Skip to content

"Vibe-Trading: Your Personal Trading Agent"

370 KB
Raw Blame

Changelog

All notable changes to Vibe-Trading are documented in this file. This project adheres to Semantic Versioning.

[Unreleased]

Fixed

Backtest runner, engines and metrics

  • A benchmark is bought at the first evaluated open, the way the strategy is. Both start from cash before the first evaluated bar, and the engine fills the strategy at a bar's open. The equal-weight basket ran instead from each bar's previous close, so with a warm-up boundary (evaluation_start_date or warmup_bars) it was credited with the move from the last warm-up close; a declared external benchmark started at the first evaluated close and missed that bar's move. Both now buy at the first evaluated open (each basket symbol at its first traded open in the window, or that bar's close when it has no usable open) and hold to the last close, so a fully invested strategy differs from buy-and-hold on the first bar only by its costs. On a 20/60 SMA run on AAPL.US over 2024-01-02..2025-12-31 the basket moves from 42.54% to 46.64%, excess return from 22.53% to 18.43% and the information ratio from 0.236 to 0.172; strategy metrics are unchanged. Runs without a warm-up boundary also move, by the first bar's open-to-close.

[0.1.19] — 2026-11-05

Rolls up 299 commits / 116 merged pull requests since 0.1.18, plus 21 pull requests whose work landed through maintainer commits, from 26 contributors.

The headline is research that outlives the chat that produced it. A finished analysis can be saved as a versioned research artifact with its evidence references and data-as-of date, searched later, and revised instead of overwritten; a follow-up turn or a later session finds the backtests already run through list_backtest_runs and reads their artifacts instead of running them again; and a research goal's objective can be edited without losing its evidence. Paper trading adds an opt-in A-share five-minute simulation that settles a completed session bar by bar on a new order lifecycle, with JoinQuant's limit-fill rules; saved accounts still settle daily. Gildata financial queries, report-period stock screening and Shannon portfolio diagnosis arrive as separate opt-ins for Gildata account holders, Gildata quotes are used only when named, and the QVeris integration is removed. Argentina (BYMA shares and CEDEARs) joins as a backtest market, web_search can use Parallel's hosted search, and Codex OAuth lists the models the signed-in account offers. Both 0.1.18 regressions are fixed: a retried provider error no longer halts live decisions for the day, and a research goal's answer that crosses its token budget is delivered. Most of the rest is correctness: formula corrections to GTJA and Qlib alphas, futures contract specifications and fees, loader paging and ticker forms, grounding of document figures and their units, background tasks that belong to the session that started them and survive a restart, and chat channels that no longer call a voice transcriber that does not exist. The Web UI gets a new welcome screen and a warm paper palette in light and dark.

Upgrading: a stored QVeris key is no longer read, and a GILDATA_TOKEN alone no longer enables any Gildata feature. Pick the features in Settings → Data Sources or VIBE_TRADING_GILDATA_FEATURES.

Added

  • Versioned research artifacts (#2079, #2159, thanks @he-yufeng; closes #1909, #2158). A finished analysis can be saved as an immutable revision with its subject, universe, data-as-of date, evidence references, verification status and optional hypothesis, goal, session or run links, through the save_research_artifact / search_research_artifacts / get_research_artifact agent tools, vibe-trading research-artifact save|search|show|revision, /api/research-artifacts and a Web page in all eleven languages. Passing an artifact_id appends a linked revision; the subject and universe cannot change. Repeating identical content and provenance returns the existing revision, and changed content or evidence defaults to unverified: saving is not verification and does not make old data fresh. The index lives in the runtime sessions.db unless VIBE_TRADING_RESEARCH_ARTIFACT_DB_PATH names another file; the idempotency lookup is indexed (#2159). The REST save route accepts at most 1,000 evidence objects (1,000,000 characters serialized) and ids of at most 200 characters.

  • list_backtest_runs finds the backtests earlier turns and sessions already ran (#2074, thanks @zeus229), so a follow-up reads their artifacts with read_run_artifact instead of re-running them. It lists runs from the req.json and run_card.json the runtime already writes, newest first and paged, with each backtest's absolute run_dir; the agent tool defaults to the current session, the MCP tool to every run. Assistant replies in the history carry [prev_run: <id>]. The grounding gate adopts a prior run read this way only when its run card and metrics.csv agree and every table it registers matches the run card manifest hash. The session filter is not an access control (runs have no owner), and a run the model forged in an earlier turn is not detected.

  • A research goal's objective can be edited in place (#2093, thanks @Shizoqua). update_research_goal (agent tool and session-isolated MCP tool) replaces the objective or display summary while keeping the same goal id, criteria and evidence, with an optional expected_goal_id stale-write guard. Replacing the goal entirely is still start_research_goal.

  • Codex OAuth model discovery (#2080, thanks @zeus229). Web Settings lists the models the signed-in ChatGPT account returns (entries marked visibility: list, including OAuth-only ones) with their display names, supported reasoning efforts and default effort, retrying once with a refreshed token on a 401. A failed catalog request keeps the configured or default model and manual entry available.

  • Opt-in five-minute A-share paper steps (#2110, #2111, #2066, thanks @shadowinlife and @Jackzigen). paper_trade_step (agent and MCP), POST /api/paper-trade/step and vibe-trading paper step accept interval="5m", plus optional session (auto / cn_a), path_model (bar_median / bar_close) and fill_model (all_or_none only). One step settles one completed cn_a session bar by bar over its 48 continuous five-minute bars from Tushare stk_mins: a market order fills at the first bar whose checks pass, zero-volume bars never fill, an order blocked by the price limit waits for a later bar and is rejected if still blocked at the close, and unfilled day orders are cancelled at session end and listed in canceled_orders. The 09:30 auction print is not matched; halts and auction-window bars are reported in events. The measured provider budget allows one symbol per step, only ordinary Shanghai/Shenzhen stocks have a five-minute leg (no ETFs, indices or Beijing listings), and strategy mode needs an explicit UNIVERSE. Such a call returns schema_version pt-v1.2 with an engine_fingerprint; a call without the new fields is byte-identical to pt-v1.1. Saved paper accounts (CLI paper advance, the REST advance route and the Web page) stay daily-only. Matching runs on a new bar-agnostic order book (#2066) with replay-stable order ids and day-order cancellation. Partial fills and volume participation remain open under #1938.

  • Five-minute limit orders follow JoinQuant's bar matching (#2127, thanks @shadowinlife). A resting buy limit fills only when a bar's low trades strictly below the limit, a sell only when its high trades strictly above it, and the fill is at the limit price; a bar that only touches the limit leaves the order resting. On the first traded bar, a reference price already better than the limit fills at that price instead. Price-limit and cash checks use the price the fill would execute at. The matching rules are versioned pt-rules-v2, so a v1.2 result's engine_fingerprint differs from one computed under pt-rules-v1 with the same bars and settings.

  • A gateway address per local IBKR or Futu connection (#1983, thanks @TerryF7). A local IBKR (TWS/Gateway) or Futu OpenD connection can store its own host and port: through Host / Port on its card in the connection center (saved, then tested), PUT /api/connections/{id}/endpoint, or vibe-trading connector setup ... --host H --port P (both required). The address is used for that connection's checks, portfolio reads and quotes and by agent read tools given its connection_id, and the connection list shows the effective endpoint. Only a hostname or IP is accepted (no scheme, path or credentials), and other transports refuse an endpoint. Existing broker configuration remains the fallback; order placement and paper/live account checks are unchanged.

  • BYMA backtests for Argentine shares and CEDEARs (#2071, thanks @zeus229). .BA codes used to resolve to their market and then fail closed in the runner; they now select a long-only, cash-only ARS engine on their own (do not set engine; leverage other than 1 and allow_short are refused, and .BA cannot share a composite pool with another market). Normal settlement is T+2, T+1 from 2024-05-27, counted in observed sessions (dates on which some loaded instrument traded); zero-volume holiday bars are dropped and never fill, and a settlement past the last bar is reported as beyond_loaded_data. Sale proceeds fund only purchases settling on the same or a later session; acquired shares may be resold before they settle. CEDEARs must be declared in ar_instruments with a ratio_history: each declared ratio change is checked against the observed price step and earlier bars are restated, an ambiguous step or an undeclared one-bar move (40% CEDEAR, 75% share) fails the run, and a position held through the change can be sold only once its earliest settlement reaches the declared credit_date. A tick rule must be declared (ar_tick_table from the BYMA Circular 3569 annex versions, or a per-instrument tick_table / tick_size); fills round against the trader. Lots default to one share and ar_brokerage_rate / ar_exchange_fee_rate / ar_fixed_fee / ar_slippage are simulation inputs, not official tariffs; all of it is labelled in the ar_data_basis metric, artifacts/byma_preparation.json and artifacts/cash_ledger.csv.

  • Optional Parallel search in web_search (#2113, thanks @RowanAldean). With VIBE_TRADING_SEARCH_PARALLEL=1, web_search queries Parallel's anonymous hosted MCP endpoint (free and rate-limited; no account, key or new dependency) after a configured Aliyun IQS and before the ddgs engines and CN fallbacks. Off by default. A Parallel failure falls through to the existing sources and its reason is prefixed to any final error; results are security-scanned and snippets cut to 300 characters.

  • Gildata combined financial queries and search candidates (#2117, thanks @TWJxi and @lishuanghong1989-netizen). The gildata_fin_query agent tool sends one natural-language question (up to 4,000 characters) covering a security's market, financial and operating indicators to Gildata FinQuery. Results keep actual periods, disclosure/trading dates and quote adjustment; amounts are converted to base currency units and percents kept as points, each with a ready figure declaration for the grounding gate. The symbol_search feature adds Gildata's A-share, HK and US candidates to search_symbol, exact matches first. Needs GILDATA_TOKEN (a commercial Gildata account) plus the fin_query / symbol_search opt-in; a FinQuery failure returns an error without trying another source.

  • Gildata stock screening with report-period indicators (#2118, thanks @TWJxi and @lishuanghong1989-netizen). gildata_stock_selection sends one question (universe, ranking or screen and its date, plus the indicators and report periods wanted, e.g. the ten heaviest CSI 300 weights with their 2023 and 2024 revenue) to Gildata's SmartStockSelection and returns the tables with the same unit normalization as gildata_fin_query. topk is a positive integer, or -1 when all results are explicitly wanted. Read-only; needs GILDATA_TOKEN and the stock_selection opt-in.

  • Shannon portfolio style and risk diagnosis (#2119, thanks @TWJxi and @lishuanghong1989-netizen). With the shannon opt-in and a token whose Gildata account carries the service, the Web Correlation page gains a Shannon tab (/api/shannon/status|catalog|diagnosis): paste up to 2,000 A-share stocks with all weights summing to 100%, or none for equal weight, and see factor exposures, per-stock characteristics and each holding's share of portfolio volatility. Indices, duplicates and partial weights are refused; the date defaults to the latest published analysis date, and a result dated after the requested date is refused. A release follow-up moved the tab's text into all eleven UI languages.

  • Factor panels combine into one composite score with walk-forward weights (#1953, thanks @BHhhh-123; #1959, #1963, thanks @rudycelekli; #2046, thanks @dajiaohuang). A new Quant Library module, combination (combine_factors, also through quantlib_call), merges several time × asset factor panels with equal, ic or icir weights, or an explicit mapping or time × factor table for any other method. IC and ICIR weights at a row use only IC observations from earlier rows and keep their sign, so a reversed factor gets a negative weight. Factors are z-scored per row by default; a factor missing on a row is left out and the others renormalised, and a row with no factor or no weight yet stays NaN. Diagnostics report per-factor and composite IC/IR, mean factor rank correlation, weight turnover and rows without weights; the IC figures are full-sample diagnostics, not the weights that were applied. Fixed before release: the insufficient="equal" fallback weights 1/n instead of 1 per factor, a single IC observation reports no ic_std instead of 0.0, and explicit weights near the float limit normalise without overflowing. It does not size positions, apply costs or neutralise exposures.

  • Alpha bench results say when each input could have been known (#2057, thanks @BHhhh-123). Bench results, plain and --strict, including those run through the alpha_bench agent tool, MCP and alpha compare, add meta.data_availability (alpha-bench.data-availability/v1): market adapter, venue, currency, timezone, calendar and observed window, and per field its source, availability mode, non-null coverage and a status of verified, assumed, unknown or failed. A declared alpha input missing from the panel is failed. SEC fundamentals stay unknown and not PIT-safe, because a filing carries a date but no publication time, even with the bench's one-bar lag. Built-in adapters cover csi300, sp500 and btc-usdt; other universes get a generic unknown context. The contract records; no IC changes.

  • Factor research runs record their mode, stage and lineage (#2067, thanks @BHhhh-123). Bench results add meta.research (factor-research.run/v1): the declared mode (explore, strict, locked), the lifecycle stage (factor_screening through monitoring), universe, period, factor ids, lineage, transformations and evaluation settings (horizon, forward-return sampling, method, and the strict runner's out-of-sample boundary). alpha bench --research-mode / --research-stage, the alpha_bench agent tool and MCP set them and refuse an unknown value; REST and alpha compare results carry the defaults. The merge exposed the mode on those surfaces and defaults it to strict for --strict runs and explore otherwise. The values are declarations: locked certifies no locked test set, and no IC, weight or backtest changes.

  • The documentation wiki can be crawled and indexed (#2165, thanks @hnq-ai). vibetrading.wiki now ships a real robots.txt (AI crawlers allowed, /api/ disallowed, a Sitemap: line) and a 42-URL sitemap.xml, both served with their own content types, plus a bilingual 404 page. The deploy prerenders every docs page to static HTML with its own canonical, the English/Chinese hreflang pair and TechArticle JSON-LD, so a crawler that runs no JavaScript gets the article body. /docs/0.1.7/* redirects permanently to /docs/latest/, and the version roots redirect to the overview. The PR wiki workflow checks these routes against local Pages responses (200, 301, 302 and 404). Whether production answers a missing path with 404 also depends on the hosting project's not-found setting.

Changed

  • The tool deny-list and bash policy apply to every provider when you set them (#2087, thanks @he-yufeng; closes #2085). VIBE_TRADING_DENY_TOOL_PREFIXES and VIBE_TRADING_BASH_POLICY (python-only or off) are enforced at the shared registry, so Agent, Swarm and MCP calls are screened even when the model names a tool it was not offered, and the advertised schemas omit denied tools. CLAUDE_CLI_* remain fallback aliases. Unconfigured, other providers keep no deny-list and off, and Claude CLI keeps its trading_,etoro_ / python-only defaults. vibe-trading provider doctor shows the active policy and where it came from. A release follow-up applies the deny-list to the MCP tools that call their code directly instead of going through the registry (backtest, paper_trade_step, run_swarm, the goal and run tools and others) and drops denied tools from the MCP tool list. These are convenience filters, not a security boundary.

  • MiniMax M3.1 Flash is the default MiniMax model (#1936, thanks @octo-patch). Direct MiniMax endpoints offer MiniMax-M3.1-Flash and MiniMax-M3, and Flash sends its reasoning back on later tool turns. On the official OpenRouter and OpenCode endpoints the configured MiniMax routes are listed first only when the gateway actually returned them; every other discovered model stays selectable.

  • Moonshot reasoning_effort reaches kimi-k3.1-pro models (#2043, thanks @Jackzigen; closes #2020). LANGCHAIN_REASONING_EFFORT is sent as a top-level field only for model names starting kimi-k3.1-pro on api.moonshot.ai; a relay MOONSHOT_BASE_URL and other Kimi models receive nothing. A level other than low, high or max fails at startup with the accepted values. Built from Moonshot's published contract; not yet verified against the live endpoint.

  • Gildata is opt-in feature by feature, and quotes only when named. In 0.1.18 a GILDATA_TOKEN alone enabled every Gildata tool and put gildata at the tail of the A-share, Hong Kong and index fallback chains. Each of nine features is now selected separately, as checkboxes in Settings → Data Sources or in VIBE_TRADING_GILDATA_FEATURES (fin_query, stock_selection, symbol_search, shannon, market_quotes, fund_nav, macro_series, announcements, broker_reports), and an existing token enables none of them on upgrade, including get_fund_nav, get_cn_macro_series, get_cn_announcements and search_broker_reports. gildata has left every fallback chain: its OHLCV needs market_quotes and an explicit source="gildata", and that request never falls back to a free source either. A saved custom order that had to list gildata keeps its other sources in order. vibe-trading data status and /data show what is selected.

  • The strategy-generation skill says the engine already delays signals one bar (#2137, thanks @davidalmeida90). The strategy-generate contract, self-check list and review categories now state that a signal set on bar t fills at the open of bar t+1, so a strategy must not shift its own signal. A model asked for a "next open" rule was adding .shift(1) itself, and every trade filled a bar late. Guidance only; the engine is unchanged.

  • The welcome screen leads with quick-action cards and featured examples. The greeting and subtitle are now left-aligned. The three quick actions (Research a stock, Test a strategy, Review my holdings) are cards that sit side by side in the 768px chat column and stack on a phone. Below them, a Featured examples row shows three library prompts, each labelled with its own category, and clicking one fills the composer. The collapsible library still holds all eight categories. The hint that examples only fill the message box now appears once, and Arabic and Persian align the copy to the start edge with the arrow mirrored. The new heading is translated in all eleven locales.

  • Light and dark themes use a warm paper palette. Background, card, popover, muted, border and chart tokens move from cool blue-grey (hue 220) to low-saturation warm neutrals. The accent changes from saturated orange (27 90% 52%) to muted umber: 24 45% 38% in light mode and 28 52% 65% in dark mode. Dark mode puts dark ink on the accent, because white on the lighter umber measures 2.26:1, below the 3:1 minimum for controls; the ink measures 7.66:1. In light mode, white measures 5.87:1 on the new umber, up from 2.77:1 on the old orange. Success, warning, info and destructive colours are unchanged.

  • Critical and high dependency vulnerabilities are patched. The minimum versions in pyproject.toml and agent/requirements.txt rise for fastmcp (≥3.4.10, was ≥2.14.0, so a 2.x install must upgrade), langgraph (≥1.2.14), pypdf (≥6.21.2), Pillow (≥13.0.0) and aiohttp (≥3.14.4). New minimums cover authlib, brotli, certifi, dnspython, langgraph-sdk, urllib3 and opentelemetry-api. The hash lock moves fastmcp 3.4.7 → 3.4.11, langgraph 1.2.12 → 1.2.18, langgraph-checkpoint 4.2.0 → 4.3.1, langgraph-sdk 0.4.5 → 0.4.10, pypdf 6.19.0 → 6.21.4, authlib 1.8.0 → 1.9.1, dnspython 2.8.0 → 2.9.0, brotli 1.2.0 → 1.2.1 and opentelemetry-api 1.45.0 → 1.45.1. A release follow-up adds a joserfc floor and pins 1.7.6 in the base and Windows locks (GHSA-9pg5-8c7x-5hfx; 1.7.5 and 1.7.4 were affected), and keeps peewee, which yfinance pulls in, at 4.5.3 in the lock, the version a plain pip install gets (the regenerated lock had moved it back to 3.19.0). The Windows desktop lock is regenerated, including Pillow 12.3.0 → 13.0.0, urllib3 2.7.0 → 2.8.0 and aiohttp 3.14.3 → 3.14.5. Electron moves 43.1.1 → 43.8.4, with overrides for undici (6.29.1, 7.30.1), brace-expansion and fast-uri. The Web frontend overrides undici (8.11.2 → 8.12.3) and postcss-selector-parser (6.1.5).

  • Routine dependency and CI updates (#2141, #2142, #2143, #2144, #2163). From the grouped Python updates, only pydantic 2.13.5 → 2.14.2 with pydantic-core 2.46.5 → 2.50.2 and ccxt 4.5.89 → 4.5.91 are applied. Dependabot now skips curl-cffi 0.17.0, which needs a newer cffi than ccxt 4.5.91 pins. The Docker image's channel lock moves pycryptodome 3.23.0 → 4.2.0, which Weixin uses for media AES. Frontend updates include react and react-dom 19.3.0 → 19.3.1, react-router 8.4.0 → 8.5.1, i18next 26.4.3 → 26.7.1, react-i18next 17.0.16 → 17.1.11, katex 0.18.10 → 0.19.1, lucide-react 1.55.0 → 1.66.0, vite 8.3.4 → 8.3.5 and vitest 5.0.3 → 5.0.5. The workflows pin actions/setup-python and actions/setup-node v7.1.0, docker/login-action v4.7.0 and actions/upload-artifact v7.0.2 by commit SHA.

  • Windows CI runs the process-isolation tests natively on x64 and ARM64. The background-lifecycle job runs on Windows Server 2025 x64 and Windows 11 ARM64 with Python 3.14. It now also runs the paper-strategy private-pipe transport test, which accepts a CRLF line ending, and the tests for owner-scoped background cancellation, cancellation during spawn, shutdown reaping and conservative restart recovery.

Removed

  • The QVeris premium data track. Retired by the maintainer in the commit that made Gildata opt-in: the qveris_search / qveris_inspect / qveris_execute agent and MCP tools, the qveris market-data source, the Settings → QVeris page and its REST routes, the /data mode and /data usage commands (/data status remains) and the QVERIS_API_KEY / QVERIS_BASE_URL settings are gone. A stored QVeris key or ~/.vibe-trading/qveris.json is no longer read, and source="qveris" is rejected as an unknown source.

Fixed

Regressions in 0.1.18

  • A retried provider error no longer halts live model decisions for the day. 0.1.18 announced every failed stream, compaction call or cancelled response that reported no usage as missing usage, which the live decision budget treats as a reason to halt; a connection reset followed by a successful retry stopped live decisions until the next UTC day. Such calls are now counted as unreported_calls in llm_usage.json; an ordinary response without usage still halts the runner as before.

  • A research goal's final answer is delivered even when it crosses the goal's token budget. 0.1.18 discarded a finished text answer whose own usage reached token_budget — the text had already streamed — and reported the run as failed. The answer is now released and the run succeeds with the budget reason; a tool call from a response past the budget is still not executed, and the exhausted goal still refuses further spending. The message says to end the goal, since a budget cannot be raised after the goal is created.

The grounding gate

  • Figures quoted from an uploaded document can be grounded (#2058, thanks @Jackzigen; closes #2050). read_document answers in prose, so its numbers never became evidence and the gate cut every figure quoted from a PDF. Up to 400 marked figures per document (percent, currency, decimal) are now recorded under its file name, never a page number or year. A document the run wrote, or one inside the run directory, is not evidence. A maintainer follow-up lets an explicitly referenced table cell be restated with the unit its header gives (1.92亿元) and used in money arithmetic, without admitting a different value, currency or a ratio. A missing file_path now names the uploads/<name> handle.

  • Percentage-point figures ground against the bare declaration (#2108, thanks @Jackzigen; closes #2107). 1.13个百分点 is conventionally declared 1.13; the gate cut such figures even when the document held the value. The borrow is one-way and exact, for whole and decimal points alike: 37% in prose still needs a percent declaration. pct / ppt / ppts / 百分点 now read as point marks, and a declared range 20%-30% becomes two endpoint declarations instead of a malformed line and a bogus -30%.

  • A percent figure can borrow a bare declaration only when its document wrote the percent (closes #2149). +17.10% grounds against a declared 17.10 | observed when the referenced document printed 17.10%, or printed 17.10 in a table column headed % / (%), including tables a PDF reader flattened to whitespace. The same number as a weight or an amount, or from another document, still fails.

  • Comma-grouped integers from a document are evidence (#2135, thanks @remote-controlled-man; refs #2126). 165,200 in a filing's plain text was read as a bare integer and left out of the ledger, so a correctly quoted table cell was refused. A visible thousands separator now marks a number as a measurement everywhere, so the same spelling in an answer needs a declaration or matching evidence too. Unmarked integers, dates and run-generated documents keep their exclusions.

  • Period valuations, dotted thousands, horizon labels and percent-point derivations (#2072, thanks @zeus229). An object naming its own ISO currency stamps the value_start / value_end and other amount fields directly beside it (a maintainer follow-up stops nested objects inheriting it). 1.618.596 is one number in any notation; 45.850 keeps its decimal reading. A 1Y / 10y table cell is a label, not a figure. A derived percent over percent-named fields the ref names grounds 41.2% for 61.56 - 20.36 and refuses 4120%.

  • Magnitude marks in a derivation are part of the operand (adapted from #2134, thanks @he-yufeng; closes #2125). ÷ 1亿, 2.5万, 3M or a currency word in a formula was evaluated as 1, 2.5 or 3, so a correct derived figure was refused. The supported marks are expanded before labels are stripped; the expanded operands still need observed or declared evidence. Magnitude marks may also follow a figure after a space (1.92 亿), from the Gildata work in #2117 (thanks @TWJxi and @lishuanghong1989-netizen).

  • A ticker's last letter is no longer read as an A-share constraint. 分析NVDA股票 matched A股, and 比较NVDA和HOOD or NVDA/HOOD matched the A/H cross-listing pattern, so the resulting market filter removed NVDA.US and the answer went ambiguous. Both patterns need a non-letter boundary now; A股, 看 A 股, 恒瑞医药 A/H and AH股 constrain as before.

Agent loop, providers and swarm

  • Stop ends a swarm run, not only the chat turn (#2059, thanks @Jackzigen; closes #2055). The swarm tool slept through its whole wait budget without checking the stop flag, so workers kept calling the model. One cancel event per turn now reaches the swarm tool, which cancels the run and returns cancelled at once; live runs are indexed per session, so a stop after the tool returned still lands and POST /swarm/runs/{id}/cancel no longer answers 404 for agent-started runs. An exhausted wait budget still does not cancel a run.

  • Background tasks belong to their owner (#2092, thanks @he-yufeng). Task reads, cancellation and completion notices are scoped to the session (or Swarm worker / isolated host) that started them, so one session can neither read nor cancel nor drain another's results; later attempts in the same session still see earlier tasks.

  • Stop and shutdown reap background processes (#2098, thanks @he-yufeng; closes #2095). A stopped or failed turn cancels its owner's running commands, including ones an earlier turn started; a successful turn leaves them running as before. Cancelling before spawn never starts the command, and the force-stop signals the process group even after the shell exited. API shutdown refuses new runs, cancels owned work off the event loop for up to five seconds and logs any task still running.

  • Background outcomes survive a restart (#2100, thanks @he-yufeng; closes #2097). Each task keeps a small redacted record under <runtime_root>/background-tasks/; after a restart a task left running reads interrupted and one left cancelling outcome_unknown, with a recovery reason, and is never re-run or signalled by its stored PID. Undelivered completion notices are restored to their owner, at least once (adapted from #2157, thanks @he-yufeng; closes #2156); corrupt or unknown-version records are reported, not dropped.

  • A silent provider stream is interrupted instead of hanging (#2099, thanks @he-yufeng; closes #2096). The wait for the next chunk is raced against the idle and total timeouts (VIBE_TRADING_LLM_TIMEOUT_SECONDS, 300 s; 0 disables) and the stop flag, so Stop reaches a stalled stream. A stopped or timed-out stream no longer falls back to a plain request that would resend an abandoned turn, and a chunk that already arrived keeps its usage. A blocking third-party adapter can still hold its socket until it returns.

  • Usage keeps cache and reasoning tokens (#2109, thanks @he-yufeng; closes #2103). Claude CLI reports cache reads and writes separately and bills every feedback retry, while context sizing uses the last prompt rather than the sum of resent prompts; Codex passes cached and reasoning tokens through. A failing Claude CLI decision carries its usage.

  • Claude CLI recovers when the model tries its own tools (#2048, #2064, thanks @he-yufeng; closes #2021, #2063). A reply with refused native tool calls, or more than two CLI turns, is sent back with a note that the host tools still work, within the existing three retries; an ordinary two-turn structured answer is accepted. The system prompt and load_tools now say to request tools through the host's tool-call protocol.

  • An answer withheld for tool-call markup gets one text-only retry (#2136, thanks @remote-controlled-man; refs #2122). When the last iteration's answer contains tool-call syntax or a tool call, the agent gets one more text-only turn beyond the iteration cap; it cannot run tools or continue a goal, still obeys Stop and goal budgets, and the neutral fallback is released if it fails again.

  • Anthropic runs survive an empty reply (adapted from #2120, thanks @he-yufeng; closes #2115). The empty-response retry, the tool-markup retry and the pending-write directive were sent as mid-run system messages, which langchain-anthropic refuses, so one empty Claude reply failed the run. They are now user messages wrapped in <system> tags.

  • Unicode line separators no longer split JSONL records (#2155, thanks @he-yufeng; closes #2154). U+2028/U+2029 inside a swarm heartbeat hid the run's liveness, so the reaper could fail a live run; the same splitlines() fault dropped records when reading traces, rebuilding the session search index and walking the audit ledger. All now split on LF only.

  • Long-session context handling. Layer-2 compaction no longer truncates the handoff summary, the only copy of older context after layer 3; a provider or model name written with underscores finds its context window; an oversized tool result carrying a lone surrogate falls back to truncation instead of failing tool delivery; and an attempt's model client is closed even when building its tool registry fails.

  • Skill and memory frontmatter (#2006, thanks @rudycelekli). A quoted scalar (description: "...", 18 bundled skills) is unquoted instead of reaching the system prompt and /skills with its quotes, and stays a string. A UTF-8 byte-order mark (Notepad, PowerShell 5) no longer hides the whole block. A memory timestamp without an offset is read as the UTC its writer stamped, not local time.

Sessions, API, CLI and tools

  • Wrapped first messages get an automatic title (#1969, thanks @rudycelekli). A session started with attachments, Swarm mode or a goal kickoff was treated as manually renamed, because the stored title was the bare prompt; the auto-title route now strips those wrappers before comparing.

  • Deleting a session clears its search index (#1970, thanks @rudycelekli). Deleted conversations stayed searchable by title and text. A maintainer follow-up purges the index first and fails the delete if that fails, so it can be retried, and refuses (HTTP 409) while a run is active. Sending a message also stamps updated_at in UTC, so session order is correct outside UTC.

  • "Long-term" and "short-term" are horizons, not orders (#1978, thanks @rudycelekli). A research objective such as "update the long-term thesis immediately" was rejected as an execution request.

  • Paper sentence splitting after numbers (#1979, thanks @rudycelekli). research_papers splits ratio is 1.5.The sample and 12%.The drawdown into two sentences, so a number cannot corroborate the next sentence.

  • A scheduled run with a channel but no target is refused (#2002, thanks @rudycelekli). POST /scheduled-runs answers 422 instead of storing a job that cannot deliver; blank channel and target strings count as unset.

  • The MCP server reads the runtime .env (#1974, thanks @alexwan02), after deciding on shell tools, so provider keys and path allowlists set there reach vibe-trading-mcp. Shell tools still need --enable-shell-tools or VIBE_TRADING_ENABLE_SHELL_TOOLS=1 in the process environment.

  • Heredocs run in the bash tool on POSIX hosts (#2077, thanks @Shizoqua). The refusal meant for cmd.exe now applies on Windows only.

  • File-tool fallbacks resolve against the run directory (adapted from #2116, thanks @Shizoqua). A relative path that escaped the run directory was checked against the allowed roots from the process's working directory, so the same call could pass or fail depending on where the server started. UNC run directories are refused before any filesystem lookup.

Tool arguments

  • Boolean tool flags sent as strings mean what they say (#2014, #1976, #1977, #2008, #2009, #2011, #2013, #2015, #2016, #2018, thanks @rudycelekli). Twelve flags were read with bool(), so the JSON string "false" counted as true: sdm_decay_scan with dry_run: "false" skipped its decay transitions, eToro's clear_stop_loss: "false" removed the stop, flatten_on_halt: "false" in a mandate proposal asked for flattening, and scaffold_signal_engine overwrote an existing signal_engine.py. true/false, yes/no, on/off, 1/0 and real booleans now read correctly, and any other value is refused before state changes; omitted flags keep their defaults and null still reads as false. Also covered: include_executions and use_rth on trading reads, eToro's take-profit and trailing-stop flags, flows_invert_sign, 13F include_changes, Taiwan active_only and the web reader's no_cache.

Market data and loaders

  • Forex detection survives AKShare 1.20 (#2083, thanks @Jackzigen). AKShare 1.20.0 deleted akshare.forex and moved the symbol map to akshare.fx.cons_forex, so the single hard-coded import made every FX pair look like a non-forex code and sent it to the A-share endpoint. Both layouts are probed now; if neither imports, the code is treated as not a forex pair.

  • Bare US ETFs and funds resolve as US listings (#2086, closes #2084, thanks @Jackzigen). search_symbol now returns QQQ.US / market us for Yahoo ETF and mutual-fund quotes on US venues instead of a suffix-less global candidate, which the identity lock then refused on the next market-data call. Non-US listings keep their native symbol. etf_holdings drops the .US suffix before matching SEC series and class tickers, so QQQ.US is no longer reported as an unknown share class.

  • A-share news survives Eastmoney's JSONP and degraded index (#2082, closes #2081, thanks @Jackzigen). The throttled JSON GET helper decodes a callback(...) envelope when the body is not plain JSON. When Eastmoney answers with its Guba user index instead of articles, get_stock_news retries once through a Chrome TLS fingerprint (curl_cffi, installed with yfinance and AKShare) on the same throttle, and raises an error naming the degraded index if that fails too, instead of reporting no news.

  • US class shares and .US symbols reach their vendors (#1997, #2010, #2012, thanks @rudycelekli). Alpha Vantage is sent AAPL instead of AAPL.US. Tiingo is asked for brk-b where BRK.B.US used to be dropped before the request, and Stooq for brk-b.us instead of brk.b.us, which is a different symbol there. Tiingo still refuses non-US suffixes.

  • AKShare US bars try NYSE American / NYSE Arca (#1998, thanks @rudycelekli). After Eastmoney market ids 105 and 106 the loader now tries 107, where most US ETFs such as SPY list, before the bare symbol.

  • CCXT history pages at the exchange's own candle cap (#1999, thanks @rudycelekli). A request for 1,000 candles from an exchange that caps lower (OKX 300, Kraken 720) returned a short first page, which pagination read as the end of history. The page size now comes from the exchange's declared fetchOHLCV limit when it is below 1,000.

  • Tushare minute volume is in board lots (#2001, thanks @rudycelekli). stk_mins reports shares while the loader declares A-share volume in lots of 100, as its daily bars are; minute volume was 100 times too large.

  • Nobitex daily bars carry the Tehran date they cover (#1989, thanks @rudycelekli). Daily bars open at Tehran midnight (20:30 UTC the day before) and were labelled with that UTC date, one day early; the first day of a window was also cut. Bars are now labelled in Asia/Tehran and the request starts early enough to include the first day.

  • screen_market returns decimal prices and percents (#1988, thanks @rudycelekli). Eastmoney's push2 list sends scaled integers unless asked for decimals (a 3.03 price as 303, 10.18% as 1018); the request now sets fltt=2.

  • vibe-trading connector history --limit prints every requested bar (#1982, thanks @rudycelekli). The table kept only the last 20 rows, so --limit 40 showed 20.

  • TTM fundamentals no longer span a missing quarter (#2042, thanks @rudycelekli). freq="ttm" summed the last four quarterly rows even when a quarter was absent from the filings, covering 15 months. A TTM value is now NaN unless its four quarters span 330 to 380 days; this reaches get_fundamentals and the fund:* panels.

  • A-share source installation is documented as shipped (#2138, closes #2121, thanks @remote-controlled-man). README, Wiki and the bundled mootdx skill now say that Tencent and AKShare ship in the base install, BaoStock needs the vibe-trading-ai[ashare] extra, and mootdx is in neither and is explicit-only, out of the automatic A-share chain.

Backtest runner, engines and metrics

  • Risk X-Ray describes a long/short book instead of refusing it (adapted from #2162, thanks @he-yufeng; closes #2147). A backtest whose average position in any symbol was net short got no risk report, and portfolio_risk_xray rejected negative weights. Short legs now keep their sign and the book its actual scale for returns and volatility; the report adds gross and net exposure with the long and short sides, and concentration (HHI, top-1, top-3) is measured on gross shares. Long-only baskets are still normalized to sum to 1.

  • STAR Market and Beijing backtests size orders on their own grid (#2105, thanks @DrDLyuan). The A-share engine rounded every order to 100-share lots. STAR Market orders (688xxx, 689xxx) now need at least 200 shares and Beijing orders at least 100, each then in 1-share steps, and a smaller size rounds to zero; main board and ChiNext keep 100-share lots.

  • Futu bars for US symbols (#2076, thanks @alexwan02). The Futu loader sent SMCI.US to OpenD unchanged and got no data back. .US codes now map to Futu's form (US.SMCI, US.BRK.B) and a bare 1-5 letter US ticker to US.<code>, while bare A-share or FX codes are left alone. Futu is still not in the US fallback chain, so this serves an explicit source="futu".

  • Global futures runs report the prices they assumed (#1965, #1972, thanks @cgycorey). A product missing from the multiplier or commission table is still priced on ES's multiplier of 50 and $2.50 a contract, but each such fallback is now warned once and listed under pricing_assumptions in the run metrics, as China futures already were; an unconfigured leverage of 10 is listed as product "*". A commission_per_contract override or a configured leverage is not.

  • Global futures contract tables corrected (#1966, #1968, thanks @cgycorey; #1992, thanks @rudycelekli). ICE orange juice (OJ) and rough rice (ZR) use multipliers of 150 and 2,000 instead of the default 50; Micro Dow (MYM) and Micro Russell (M2K) pay the micro commission of $0.62 instead of $2.50 and take the 7% equity-index price limit like MES and MNQ.

  • Digit-bearing global futures reach the global futures engine. Codes such as M2K=F, M2KZ4, 6EZ4 and 6EZ2026.CME, including local: inputs, failed the futures patterns or lost part of their product code; they are now routed as futures and keep their full product, so the table entries above apply to them.

  • Spot crypto backtests no longer pay perpetual funding by default (closes #2088). A bare spot pair such as BTC-USDT was charged funding at 0.01% per settlement, and the generic commission was ignored in favour of fixed maker/taker rates. Funding now applies to -PERP / -SWAP and CCXT settled-contract codes, strict perpetual mode, or an explicitly nonzero funding_rate; commission sets both fee sides unless maker_rate / taker_rate override one. Metrics report the signed crypto_funding_pnl and the applied fee rates.

  • Yahoo three-letter forex codes read as USD pairs (#1980, thanks @wavebyrd). JPY=X is USD/JPY on Yahoo; the forex engine read it as an unknown pair and priced it on default spread, swap and lot tables.

  • Sub-penny prices survive in trades and charts (#1987, #1990, thanks @rudycelekli). trades.csv rounded fill prices to four decimals and the reconstructed chart rows of a run rounded OHLC to six, so very low-priced instruments showed wrong or zero prices. Both now keep the full value.

  • The 30-year CFFEX treasury future pays the treasury fee (#1994, thanks @rudycelekli). TL took the generic 5 per lot instead of the fixed 3 of T, TF and TS.

  • India stamp duty follows the trade side (#1995, thanks @rudycelekli). With allow_short, stamp duty was charged on the opening short sale and the DP charge on its cover; the cover is the purchase and now pays the stamp duty, the short sale the DP charge.

  • Cross-market options runs measure their own annualisation (#2005, thanks @rudycelekli). The options engine labels its equity curve with date strings, which the span measurement could not subtract, so a run without a fixed bar count fell back to 252 bars a year. The labels are parsed first.

  • Hong Kong runs get a benchmark that loads (#2041, thanks @rudycelekli). The default was the Futu-style HK.03100, which Yahoo cannot serve, so every HK backtest lost its benchmark silently. It is now the Hang Seng China Enterprises Index ETF 2828.HK.

  • Correlation and regime analysis align prices before taking returns (#1986, thanks @rudycelekli). Returns were computed on each asset's own calendar and then joined, pairing a crypto Sunday-to-Monday return with an equity Friday-to-Monday one. Closes are now aligned on common dates first, in the Correlation page's matrix and in regime detection.

  • A missing risk x-ray says why (issue #2147). When the portfolio risk x-ray cannot be computed for a run, the run card now carries "Risk x-ray unavailable: " instead of the artifact silently not appearing.

  • Cross-market intraday runs annualise over the exact span (#1971, thanks @rudycelekli). effective_bars_per_year, used when a run's codes span more than one market, divided by whole elapsed days. 72 hourly bars spanning 2 days 23 hours were annualised as if over 2 days: 13,149 bars a year instead of 8,889, and a Sharpe 21.6% too high. It now uses the fractional span; daily runs are unchanged, and a span under one day still counts as one year.

Validation and run artifacts

  • Backtest validation refuses equity at or below zero (#2161, closes #2160, thanks @he-yufeng). Percentage returns and drawdown are undefined once equity reaches zero, yet validation still reported numbers for them. Monte Carlo (the observed trade path or any shuffled one), the bootstrap Sharpe CI and each walk-forward window now return only an error when observed equity is zero, negative or non-finite, and the bootstrap also refuses a non-positive initial capital. Monte Carlo errors no longer carry a placeholder p_value_sharpe of 1.0; the Web validation panel shows the error alone. Landed through a maintainer commit.

Quant Library and analysis tools

  • Historical VaR and CVaR use the documented number of tail losses (#1950, thanks @rudycelekli). The tail rank ceil((1 − c) × n) − 1 was computed in binary floating point, where (1 − 0.95) × 100 is 5.000000000000004, so 100 returns at 95% took the sixth-worst loss instead of the fifth and CVaR averaged six, understating both. The rank now comes from the confidence's decimal spelling, with no tolerance, so nearby probabilities keep distinct ranks. historical_var, historical_cvar and the VaR/CVaR of analyze_mc_results move at such boundaries.

  • The Corrado rank test's S(K) includes event days (#2000, thanks @rudycelekli). event_study divided corrado_rank_z by an S(K) averaged over estimation days only; Corrado (1989) averages the squared mean rank deviation over the combined estimation and event period. With a real event effect the left-out days carry the most extreme ranks, so the statistic overstated significance exactly when it rejected. Ranking and numerator are unchanged.

  • Drawdown day counts use calendar dates (#2026, thanks @rudycelekli). max_drawdown_analysis counted underwater_days and recovery_days with Timedelta.days, so on a timezone-aware series a 23-hour spring-forward day counted as 0. Both now subtract civil dates; naive midnight-stamped series are unchanged.

  • Market impact functions refuse arrays (#1960, thanks @rudycelekli). fixed_slippage, linear_impact and sqrt_impact are documented scalar-only and typed -> float, but broadcast a list, array or Series and returned an ndarray, dropping a Series' index; quantlib_call's __series__ envelope reached this. Price, volume, ADV, bps, coefficient, volatility and eta arguments now raise TypeError for a container or a bool. Scalar results are unchanged.

  • Heston prices are right for short expiries (#2140, thanks @kadyrbekovhamit-cyber). heston_price integrated its pricing formula only up to u = 200. An option a day or less from expiry has an integrand that has not decayed by then, and the truncated price can be off by more than 10%. The default is now integration_limit=None, which integrates to infinity (0–200 plus the tail); an explicit finite limit keeps the old truncation.

  • Run attribution keeps benchmark returns aligned (#2044, thanks @dajiaohuang). GET /runs/{run_id}/attribution dropped an equity.csv row with an unusable portfolio return together with that row's benchmark equity, so the next benchmark return spanned two bars and was paired with a one-bar portfolio return in the alpha/beta regression. The row now stays as a gap and only its own pair is left out.

  • The options payoff API refuses boolean quantities (#2049, thanks @dajiaohuang). POST /options/payoff accepted "qty": true as one contract, because the request model coerces booleans to integers. A boolean qty is now a validation error, as zero already was.

Alpha Zoo, factors and alpha bench

  • Twenty GTJA 191 alphas read A-share VWAP in yuan (#2039, thanks @rudycelekli). Alphas 7, 8, 12, 13, 16, 17, 26, 36, 39, 41, 44, 45, 61, 64, 73, 74, 77, 87, 90 and 92 built VWAP inline as amount / (volume × 100 + 1). On Tushare units (amount in thousand CNY, vol in lots) that is a thousandth of the traded price, and a vwap column already in the panel was ignored. They now call the shared vwap(panel, "equity_cn"), which uses vwap when present and otherwise scales amount by 1,000. Alphas that use VWAP only through ranks or correlations are insensitive to the scale; those that mix it with price levels (13's sqrt(H × L) − VWAP, the weighted prices in 74, 77, 87 and 92) change value.

  • GTJA alpha 17 keeps the sign of its exponent (#2051, thanks @rudycelekli). The report's RANK(VWAP − MAX(VWAP,15)) ^ DELTA(CLOSE,5) was computed as |base| ^ |exponent|, so after a five-day fall a base of 0.5 and a change of −2 gave 0.25 instead of 4. The base is a percentile rank in (0, 1], so the power is now taken directly with the signed change. The merge also corrected the golden fixture and the factor's English and Chinese notes. Values change wherever the five-day close change is negative (and, via #2039, wherever the VWAP scale mattered). A release follow-up keeps the value missing whenever the base or the exponent is: IEEE pow gives 1 for NaN ** 0 and 1 ** NaN, so a flat close during the 15-bar VWAP warmup, or the top-ranked name on a bar with no close change, had started to report 1.0.

  • GTJA alpha 55 has the swing index's third denominator branch (#2036, thanks @rudycelekli). The denominator chose between the high-gap form |H−C₋₁| + |L−C₋₁|/2 + |C₋₁−O₋₁|/4 and the low-gap form only. When neither of those gaps is the largest range, the report uses |H−L₋₁| + |C₋₁−O₋₁|/4, and the code now does too. Values change on those bars and on every 20-day sum that contains one.

  • GTJA alpha 170 scales by the 5-day mean of HIGH (#2037, thanks @rudycelekli). The middle term HIGH × RANK(HIGH − CLOSE) / (SUM(HIGH,5)/5) divided by the 5-day mean of CLOSE instead of HIGH, contrary to the report and the factor's own formula metadata. Every value changes.

  • GTJA alpha 140 drops an extra TSRANK (#2038, thanks @rudycelekli). The correlation leg ranked the correlation over 7 days before decaying it. The report's leg is TSRANK(DECAYLINEAR(CORR(TSRANK(CLOSE,8), TSRANK(MEAN(VOLUME,60),20), 8), 7), 3), which is what is computed now. Values change wherever that leg is the smaller of the two.

  • GTJA alpha 101 uses the report's right-hand leg (#2040, thanks @rudycelekli). It compared against RANK(CORR(RANK(HIGH), RANK(MEAN(VOLUME,10)), 11)); the report's leg is RANK(CORR(RANK(HIGH × 0.1 + VWAP × 0.9), RANK(VOLUME), 11)). Code, formula metadata and golden fixture now follow the report, and a bar is NaN unless VWAP was observed across the 11-day window. The 0 / −1 signal flips on bars where the comparison changes.

  • Qlib BETA and RESI are least-squares trend statistics (#2062, thanks @he-yufeng). Alpha158 defines BETAd = Slope(close, d) / close and RESId = Resi(close, d) / close. The five BETA factors (5, 10, 20, 30, 60) computed the endpoint slope (close − close₋d) / (d × close), and the five RESI factors the gap to the moving average (close − mean(close, d)) / close. Both now use new ts_slope / ts_resi operators: the OLS slope of close on time over a full window, and the latest close's residual from that fit. Every value changes.

  • Qlib ROC and CORD match Alpha158 (#2069, thanks @he-yufeng). The five ROC factors returned the d-day return close / close₋d − 1; Alpha158's ROCd is close₋d / close, a decreasing function of that return, so each ROC factor's rank IC changes sign and a bench may now call one reversed that it called alive, or the reverse. The five CORD factors correlated the price ratio with log((v + 1) / (v₋₁ + 1)) instead of Alpha158's log(v / v₋₁ + 1), which they now use. Formula metadata follows.

Valuation, rigor checks and report audit

  • Valuation workbooks keep text literal (#1949, thanks @rudycelekli). export_dcf_workbook, export_comps_workbook and export_three_statement_workbook let openpyxl read a string starting with = as a formula and #N/A as an Excel error, so an assumption basis such as = 2% inflation + 1% real growth read back empty. Every text cell is now written as a string; numbers and booleans keep their types.

  • financial_rigor refuses non-finite and boolean inputs (#2053, thanks @yufengsika). NaN, Infinity and true (read as 1) were accepted and answered status: "ok" with NaN or Infinity in a payload that is not valid JSON, and a finite price such as 1e308 could overflow to infinity. Every command that converts a number now returns an error envelope for such input, and a result that cannot be written as finite JSON is an error.

  • Earnings and FCF yield are omitted at a zero price (#2033, thanks @rudycelekli). verify_valuation raised DivisionByZero when the price was 0 and EPS or FCF per share was not. Those two yields are now left out, as dividend yield already was; P/E, P/FCF and the other ratios with price in the numerator are still reported.

  • Report audit skips fenced examples (#1951, thanks @rudycelekli). report_audit dropped fenced code only from its label: value pass, so a template table inside a fence was sampled as a real claim, and tilde fences or a longer fence around a shorter one were mishandled. Fenced blocks (backtick or tilde, closed by a fence at least as long, or unclosed to the end) are now blanked before both passes, keeping the report's line numbers.

Governance ledger and entity ingestion

  • The audit ledger rotates past its first segment (#1962, thanks @rudycelekli). rotate_if_needed verified only the active file before sealing it. After one rotation that file legitimately starts mid-chain, so the check reported a sequence gap on a healthy ledger and every later rotation raised LedgerCorruptionError. It now verifies the whole history, sealed segments included. A maintainer follow-up numbers a new archive after the highest existing one (past gaps and beyond four digits), orders segments numerically, and seals through a hard link so an existing archive is never overwritten. No runtime path calls the rotation automatically.

  • Run manifests detect an edited tool list (#2017, thanks @rudycelekli). RunManifest.verify_hash covered tools_hash but not the tool names it summarises, so a run_manifest.json whose tool_names gained, lost or renamed a tool still verified. It now rebuilds the snapshot from the names and fails on a mismatch, and the evaluation harness's manifest-integrity verdict inherits the check. Untampered round trips still verify.

  • Imported amounts read dot grouping, and only valid grouping (#1964, thanks @rudycelekli). In load_cashflows and load_panel, 1.234.567 failed as an unusable amount while 1,234,567 parsed. Repeated dots now group digits too. A maintainer follow-up accepts grouping, inferred or implied by a declared decimal_separator, only when every group is valid (threes, or Indian 2-2-3 grouping), so a malformed amount or a separator inside the fraction is refused instead of concatenated. 1,234 and 12.000 still ask for decimal_separator.

Paper trading

  • Strategy steps no longer hang on the stdio MCP transport (closes #2089). The strategy worker was a multiprocessing spawn child, which on Windows could be blocked by the MCP host's pending read of the server's stdin, so every strategy_path call timed out. The worker is now a fresh interpreter that talks to the server over its own private pipes, and strategy prints never reach the protocol stream. A regression test that starts the worker while the host's stdin read is blocked now also runs in CI on native Windows x64 and ARM64. Startup and per-day deadlines, sandboxing and hard teardown are unchanged.

  • ST stocks outside the main board keep their board's price limit (#1984, thanks @rudycelekli). When the exchange limit table has no row and the band is derived from the previous close, every ST stock was given a ±5% band, so a ChiNext or STAR ST stock moving 8% had its buy or sell rejected as limit_up / limit_down. Only main-board ST stocks narrow to ±5% now; ChiNext and STAR keep ±20% and Beijing ±30%.

  • A 3:2 split priced just inside the neutral band is still caught (#1985, thanks @rudycelekli). The split detector ignored any close-to-close ratio inside (2/3, 3/2), so a 3:2 split quantized to the cent (30.00 to 20.01, a ratio of 0.667) passed as an ordinary one-third loss. A ratio within 0.1% of 2:3 or 3:2 is now a candidate: with matching volume and NAV continuity the step returns PAPER_DATA_NOT_READY, otherwise that day's orders on the symbol are rejected corporate_action_unresolved. Holdings are still never rescaled from prices alone.

Live trading, brokers and portfolio

  • Robinhood portfolio positions are priced, and long accounts are read in full (#2054, closes #1428, thanks @Jackzigen). The read-only Robinhood source now declares quotes.read and prices each equity position through get_equity_quotes instead of listing it unpriced, and position and order reads follow the next cursor to the last page; a failed follow-up page fails the read rather than returning part of the account. A quote refresh never opens a browser for OAuth. A maintainer follow-up requires the quote to name the requested symbol, prices only an active instrument that has traded, uses the newer timestamped regular- or extended-hours trade and shows its as-of time; anything else stays unpriced with a price_error. The mapping follows a third-party capture of the server's schema and has not been checked against a live account.

  • Portfolio errors no longer carry credentials, and Binance HTTP 451 is explained (#1937, thanks @goatyyc). Connector read errors go through the shared redactor before they are stored, sent as refresh progress or returned: URL query strings and signature / api_key / secret values are replaced, and the latest and historical snapshot reads clean errors stored before the upgrade. A Binance HTTP 451 "restricted location" failure gets the code binance_restricted_location and a localized explanation on the Portfolio page.

  • Longbridge keeps per-currency cash and financing limits (#1940, thanks @alexwan02). Longbridge's total_cash nets every currency into one number, so a borrowed (negative) currency leg disappeared behind a positive one. Balance rows now carry each currency's available, withdrawable, frozen and settling cash and the account's maximum and remaining financing amount, through account reads, Portfolio snapshots (balance_details), the portfolio and connector-balance CLI output, the API, the agent's portfolio context and the Web account card. They are not added to totals; negative values keep their sign and unreported ones stay empty.

  • Binance spot open orders can be listed across all symbols (closes #2090). get_open_orders without a symbol always raised, because CCXT refuses a symbol-less fetchOpenOrders by default to warn about its request weight. The spot client now acknowledges that warning; CCXT's rate limiter stays on, and exchange, authentication and network failures still surface. Fill history without a symbol still reports that Binance needs one.

  • Deleting a connection keeps its vault protection (#2153, thanks @TerryF7). Deleting an OAuth or local-gateway connection, whose profile declares no credential fields, no longer needs the OS keyring; it used to fail without the keyring extra. A connection with stored credentials is still kept when the vault is unavailable or cannot confirm the secrets were removed, and the API answers HTTP 503 with the reason instead of a server error. The connection center shows the failure on that connection's card, with a localized message when the portfolio still uses the connection.

  • Zerodha and Shoonya paper orders refuse fractional quantities (#1991, thanks @rudycelekli). The local paper simulators truncated the quantity, so 0.5 shares was acknowledged as a simulated fill of 0 shares and 2.7 as 2. A quantity must now be a finite whole number of at least 1, otherwise the order is refused with "quantity must be a positive whole number of shares". Both brokers remain paper-only.

  • The broker key . can no longer act on the global kill switch (#2003, thanks @rudycelekli). A broker's live directory is <runtime_root>/live/<broker>, and . resolved to the live root itself, so that broker's HALT sentinel was the global one: halting or clearing broker . halted or released every broker. . is now rejected alongside path separators and ...

  • MT5 keeps a configured deviation_points of 0 (#2019, thanks @rudycelekli). An explicit zero slippage tolerance was read as "unset" and replaced by 20 points on every market order; only a missing or empty value defaults to 20 now.

Shadow account

  • Shadow-account backtests trade every documented entry condition (#2073, thanks @cgycorey). The generated signal_engine.py recognised only two-sided {"min", "max"} bounds, so ("<=", -0.08) pairs, bare scalars and one-sided bounds were dropped and the backtest traded a looser rule than the profile stated. Pairs and scalars (read as >=, as the live scanner reads them) now map onto inclusive bounds, each side applies on its own, and scanner and engine share one operator vocabulary. A strict > or <, ==, != or a non-string market, which the engine cannot express exactly, now raises an error instead of being widened.

  • The shadow-account result cache no longer serves stale runs (#2075, thanks @cgycorey). A cached run was keyed by window and journal hash only, so an edited rule or market selection silently reused numbers from the old definition; the key now also covers the rules, the source and preferred markets and an engine epoch. A cached report also dropped excluded_currencies and read as if every currency had been summed; the disclosure is restored. Existing caches miss once and re-run.

Trade journal and cash flows

  • Journal buys and sells need a valid quantity (#2052, thanks @MetaAviator). A buy or sell row whose quantity was missing, zero, negative or non-finite reached FIFO matching. The analysis now returns an error asking for finite positive quantities before any filter is applied. Cash dividends may still have zero quantity.

  • Cash-flow transfers are sign-checked (#2023, thanks @dajiaohuang). transfer_in and transfer_out carried no sign rule, so a transfer written from the portfolio's side entered XIRR and money-weighted returns with the wrong sign. They now follow the holder's convention used for contributions and distributions: a transfer into the portfolio is negative and a transfer out positive, and a contradicting amount is refused with a hint to flip the sign or use another kind. The cashflow_performance schema states it.

Channels and scheduled delivery

  • Feishu tool hints land on the streaming card that is already open (#1957, thanks @Shizoqua). The hint path looked up the reply's buffer by the inbound message id, but it wrote through send_delta without that metadata, so the write resolved to the chat id instead. The hint then opened a second card whose "generating" indicator never cleared. The metadata is now forwarded, so both lookups use the same stream.

  • Teams replies to a multi-paragraph message keep the quote out of the user's text (#1958, thanks @Shizoqua). For the "Reply wrapper" shape, the parser split only at the first blank line, so later paragraphs of the quoted message reached the agent as the user's reply. It now splits at every blank line and treats only the last block as the reply. Limit: if the reply itself has several paragraphs, all but its last are folded into the quote.

  • Weixin attachments with the same name no longer overwrite each other (#1967, thanks @Shizoqua). A sender-supplied filename was written as-is into the media directory, so a second report.pdf replaced the first one, which an earlier turn still pointed to. Each download is now created exclusively under a unique name that keeps the original stem (up to 32 characters) and extension. Repeated names, identical URLs, same-second retries and concurrent downloads each get their own file. The maintainer follow-up replaced the PR's timestamp-and-hash suffix, which could still collide.

  • Voice notes reach the agent as attachments on five channels (#2078, #2106, thanks @Shizoqua and @VectorPeak). WhatsApp (#2078) and Telegram, Feishu, Matrix and Weixin (#2106) called a transcribe_audio method that no channel defines. As a result, a voice note was dropped or logged as a failed download. Each adapter now forwards the downloaded audio with the same media marker it builds for other attachments. No speech-to-text runs: the agent gets the audio file, not a transcript.

  • Discord streams no longer fail on a long first message or repeat overflow (#2114, thanks @Shizoqua). The first message of a stream sent the whole buffer, so a first delta longer than Discord's 2,000-character limit failed the stream. The first message now sends only the first chunk, as later edits already did, and the remaining text goes out once when the stream finishes. The maintainer merge removed the PR's extra sends at stream start, which would have duplicated that overflow.

  • Channel path containment has boundary tests (#2094, thanks @Shizoqua). Matrix uses is_path_within to keep outbound attachments inside the workspace when that restriction is on. It is now tested against a sibling directory that shares the root's prefix, .. traversal both out of and within the root, and a symlink that points outside. Behaviour is unchanged.

Web UI and locales

  • Stopping a chat also stops its swarm run and freezes the card (#2060, thanks @Jackzigen). The stop button cancelled only the session, so a swarm that had already started kept its card spinning, and a late event could repaint it as completed. Stop now marks every pending or running swarm card in the chat as cancelled, including its running and waiting agents. It also sends POST /swarm/runs/{id}/cancel for each run and ignores any later event for a cancelled card. Limit: the card shows cancelled even when the cancel request fails, because every error is swallowed, not only the 404 from a run that already finished.

[0.1.18] — 2026-10-21

Same-day follow-up to 0.1.17: 12 commits (10 non-merge), no new external pull requests. It ships the fixes made on main after 0.1.17 — mcp>=1.30 so the HTTP MCP transport reclaims sessions, connect-time checks for custom model endpoints (including fake-IP proxy hosts), grounding of run_dir/… refs and signed amounts, Beijing codes in correlation, the macOS timeout fix and the sdist contents — together with a leaner agent runtime: tool schemas load on demand through load_tools, oversized tool results stay readable in the run directory, Swarm sizes its context by the model window, and usage accounting covers compaction, cancelled responses and failed streams. On the 0.1.17 pre-release verification prompt (DeepSeek V4 Pro, a 20/60 SMA backtest on AAPL.US) a run took 7 model calls instead of 12 and 152K input tokens instead of 706K, and reported the same figures.

Known issues

  • A retried provider error (a connection reset or rate limit) can halt live model decisions until the next UTC day, and a research goal's final answer that crosses its token budget is discarded. Both are fixed in 0.1.19.

Changed

  • Reduce agent request overhead with on-demand tool schemas while keeping the complete tool catalog and existing execution permissions. Ordinary backtests no longer request extra attribution automatically. Automatic summary titles remain available after the first chat exchange.
  • Preserve oversized tool results in readable run artifacts and stop Swarm from clearing new observations before the model sees them. Context pressure uses the model window and includes tool schemas.
  • README credits cover every cycle again: contributor lines for the 0.1.13, 0.1.15, 0.1.16 and 0.1.17 cycles in all seven READMEs, each PR attribution checked against GitHub. The data-tools list counts the 29 read-only data tools the MCP server serves (it said 22). The 0.1.13 notes below credit ModelScope (#1011) to @yrk111222 and #676 to the session lifecycle entry it fixed.

Fixed

  • Keep the custom-endpoint guard lazy-loaded for minimal installations and align the Windows desktop MCP lock with the updated package requirement.
  • Include compaction, cancelled responses and failed streams in usage accounting; stop further work when a research goal exhausts its token budget. Forward the configured output cap to both DeepSeek adapters.
  • Redact credentials before tool results reach model context or replay caches, and reject incomplete compatible-provider streams instead of releasing a truncated answer or tool call.
  • vibe-trading-mcp --transport http reclaims its sessions (CVE-2026-59951 / GHSA-84m7-p3x7-pcfv). mcp is now >=1.30.0 instead of capped below 1.30: an idle session closes after 30 minutes, at most 10,000 exist and an ended session is dropped, so a client can no longer grow server memory without bound. 0.1.15–0.1.17 held the cap because 1.30's RFC 8414 issuer check refused IBKR's OAuth metadata; IBKR's protected-resource document now names the same issuer as its metadata (checked 2026-10-21) and Robinhood's already did. The Docker lock pins mcp 1.30.0.
  • Custom model endpoints are checked again on every connection. Saving a profile resolved its host once; a name that later re-resolved to a private address (DNS rebinding), or a redirect to one, was not checked again. The provider test and the running model now resolve each direct connection once, refuse it unless every address is public and connect to the address they checked (TLS still verifies the host name). When an environment proxy carries the request the proxy resolves the name, so this check does not apply there.
  • Custom providers can be saved behind fake-IP proxies. Clash, Surge and sing-box in TUN mode answer every lookup from 198.18.0.0/15, which 0.1.17 refused, so no custom provider could be saved on such a host. A looked-up 198.18.x address is accepted now (the proxy resolves the real name); the same address written into the URL is still refused.
  • Grounding reads the run's own files under the names models use. A figures-block ref naming the active run as run_dir/… or by its directory name (20261021_…/artifacts/metrics.csv) resolves to that run's engine-written file, and a sign written before a currency mark (−$1,847.78, (−$5.00), PnL: -$3.20) is the number's sign in both prose and declarations. A live DeepSeek backtest report had its final value, trade count and losing trade's PnL cut for these two reasons. Model-written files and paths outside the run still never resolve, and a hyphen joining two figures ($10-$20) or a spaced one ($10 - $20) is not a sign.
  • Correlation analysis puts bare Beijing, Shanghai-fund and B-share codes on their exchange. 920xxx resolved to .SZ, as did Shanghai 5xxxxx funds and 9xxxxx B shares; the rule now matches the Tushare fallback's (4/8/920 → .BJ, 5/6/9 → .SH, otherwise .SZ).
  • A timed-out background command is reported as a timeout on macOS. macOS answers EPERM rather than ESRCH when a process group holds only exited, unreaped processes, so the final kill of a timed-out tree raised and the task was marked error on a loaded host.
  • Paper accounts say when nothing can be settled yet. Advancing an account before its first session closed reported a date range that ran backwards ("between 2026-10-21 and 2026-10-20"); it now names the first execution date and the settlement time in the market's time zone.
  • The sdist carries what its tests need: the test packages' __init__.py files and CSV / JSON fixtures, and no untracked test_e2e_*.py harness.

[0.1.17] — 2026-10-21

Rolls up 638 commits / 228 merged pull requests since 0.1.16, plus 41 pull requests whose work landed through maintainer commits, from 33 contributors.

The headline is daily paper trading. A paper_trade_step tool for the agent and MCP, and saved paper accounts through vibe-trading paper, /api/paper-accounts and a Web page, match orders or a strategy against completed daily bars without ever reaching a broker; saved accounts move forward only. Most of the rest of the cycle is correctness, and much of it has one shape: a value that cannot be computed now says so instead of turning into a plausible number. A profit factor with no losing trade, an undefined correlation, a factor benchmark with too few IC observations, an upside at a zero price and an unbounded ratio come back as null or a refusal rather than 0; a missing price no longer bridges a return pair, a VaR transition or a lag pair; an after-close SEC filing no longer reaches the IC a bar early; and a misspelled backtest setting is reported in the run card instead of silently ignored. Brazil (B3) joins as a market, the Alpha Zoo grows from 462 to 478 factors, Slack and Discord join the guided channel setup, scheduled reports can arrive by Email as HTML or a password-protectable PDF, token usage shows in Run Detail, chat and swarm, a claude-cli provider runs the agent on a local Claude Code login, and the Web UI adds Persian and Turkish for 11 languages.

Added

  • Daily paper trading with saved accounts (#1939, thanks @shadowinlife). A new paper_trade_step agent and MCP tool matches explicit orders, or the per-day orders of a strategy file, against completed daily bars for cn_a, hk and us; nothing in it can reach a broker order path. The same stateless step is served as vibe-trading paper step and POST /api/paper-trade/step. Saved paper accounts are built on it and reachable three ways: vibe-trading paper create|list|show|history|advance, authenticated REST under /api/paper-accounts (list, create, read, history, POST .../{id}/advance) and a Paper trading page in the Web sidebar in all eleven UI languages. An account starts from the cash and positions you enter, today or at the next session, and moves forward only: an advance settles sessions whose regular close is at least 30 minutes past and carries the revision it read, so a stale or concurrent request is refused (HTTP 409 over REST) instead of committing twice, and a failed step leaves the ledger unchanged. Ledgers live in <runtime_root>/paper-trading/accounts.db; there is no background scheduler. A strategy under <runtime_root>/strategies/ runs in a child process that sees only bars strictly before the execution day and the ledger after earlier fills; orders fill at that day's close, and any strategy failure discards the whole call. Matching is daily-close and all-or-none, with simplified fees, price limits, board lots and A-share T+1. Execution prices must be raw: adjusted or unknown price calibers and non-finite prices are rejected, A-shares need a complete ClickHouse trading calendar, and a split-like jump returns PAPER_DATA_NOT_READY rather than rescaling holdings. Corporate-action reconciliation and cash dividends are not modelled and NAV is not total return; intraday, margin, options and cross-market phases remain open under #1938.

  • Live runner model calls have their own cadence and daily budget (closes #1854). Safety checks still run every minute, but a model decision now defaults to every 15 minutes, each in a fresh conversation, with at most 26 decisions and 1,000,000 reported tokens per broker per UTC day; an optional once-daily time, timezone and operator-priced USD limit are set in Runtime → Model schedule and spending limits, the chat runner panel, vibe-trading connector runner-settings or GET/PUT /live/runner/settings. Spending history survives restarts. Limits are checked after each provider response, so one response can overshoot; estimates ignore cache discounts and auxiliary calls and are not a provider billing cap. Exhaustion or missing usage trips the existing kill switch and its cancel/flatten behaviour, and a decision left unfinished by a crash is blocked for inspection.

  • Strategy adaptation runs from preparation to evidence. A registered strategy with adequate, non-stale evidence can be adapted to a new name, universe, date window and position sizing through adapt_strategy (agent and MCP), vibe-trading strategy-evidence adapt, authenticated POST /api/strategy-adaptations and a Web welcome prompt. The SDM child keeps the parent's signal, entry and exit definitions and gets its own run directory, SignalEngine and backtest; it never inherits parent metrics. Refreshing with scope=strategies (strategy-evidence refresh --scope strategies, or POST /api/strategy-adaptations/evidence) replaces only the named strategies' evidence, and refuses an adapted run unless its config, code, metrics, equity and trades match the hashes sealed in its run card. The hash binds evidence to the persisted run; it does not certify that the generated code implements its description.

  • extraETF holdings exports feed the portfolio, with allocation proposals. vibe-trading portfolio import-extraetf, POST /api/portfolio/imports/extraetf and Portfolio → Manage accounts → Import extraETF Investments store an Investments CSV as a local source. The whole import fails on an unreadable row, a holdings count that differs from the one you confirm, a missing price, a future file time or an export older than the freshness limit you choose; transactions, dividend and X-Ray exports are refused. Portfolio ID keeps accounts apart when they hold the same ISIN. Totals are exported quantity × price and exclude cash and assets absent from the export. EUR joins USD, CNY and HKD as a display currency, valued at Frankfurter reference rates whose date is shown apart from the fetch time. An allocation proposal (portfolio rebalance --snapshot-id, POST /api/portfolio/rebalance, portfolio_summary(snapshot_id=..., target_weights=...), or the Web page) shows value changes for target percentages that total 100%; it creates no orders, quantities, tax or fee estimates, and an incomplete snapshot or stale cached rates cannot produce one. Imported holdings stay out of historical risk tools until their trading venue is resolved.

  • Portfolio research can stay on one snapshot (#1769, thanks @zeus229). portfolio_summary returns snapshot_id and read_identity and accepts a snapshot_id back, so a multi-step analysis keeps reading the same stored observation after a later refresh. vibe-trading portfolio show --snapshot-id ID and GET /api/portfolio/analysis-context?snapshot_id= read the same way, and the Web Portfolio page exports the sanitized research snapshot it displays as JSON. An unknown snapshot, one valued under an older valuation version, or one containing a source that is no longer enabled is reported as unavailable and never replaced by newer data.

  • mStock (India) read-only connector (#1675, thanks @Jackzigen). The mstock-live-rest-readonly profile sends only GET requests for the Type B holdings, net-position and fund-summary resources, with the API key and current session token kept in the connection's OS credential vault; it never logs in, refreshes a session or calls an order endpoint, and both order entry points refuse before any network call. Holdings and net positions stay separate books. Fund limits are shown raw because they are not cash or net liquidation, so the profile withholds account.read and is not a portfolio source. connector account and connector positions gain --connection-id. Tests replay the documented reply shapes; no real mStock account was available.

  • Five example options strategies and a candidate study (#1774, thanks @mdb1230). The options-strategy skill ships covered call, cash-secured put, iron condor, volatility-gated short strangle and protective put OptionsSignalEngines, and agent/scripts/options_candidates_backtest.py picks one parameter per strategy on 2016–2020 SPY/QQQ, then reports 2021 onward on SPY, QQQ, AAPL and MSFT with those parameters fixed. This is research code for choosing what to paper-trade; no engine changed and no order path exists.

  • Custom OpenAI-compatible model endpoints (#1264, thanks @Emad211). Settings gains a custom-provider manager and the CLI gains vibe-trading provider custom list|add|activate. A profile is saved only after a test chat completion with the same URL, model and key succeeded in the last five minutes; the key goes to the OS credential vault (the [keyring] extra), never to the profile JSON or argv, and activation needs an explicit confirmation. An active profile cannot be edited: save the changed endpoint under a new id. Base URLs carrying credentials, and hosts that are or resolve to any non-public address (loopback, private, link-local, reserved, or the shared 100.64.0.0/10 range that holds Alibaba Cloud's metadata service), are refused, so a model served on localhost cannot be registered this way. The host is checked when the profile is tested and saved; a name that later re-resolves to a private address is not re-checked. The test is a real request and may be billed. LANGCHAIN_MAX_TOKENS, when set, now caps output tokens on the OpenAI-compatible adapter. The PR's per-channel configuration API was not merged; channel settings stay where 0.1.16 put them.

  • Persian and Turkish join the Web UI, which now has 11 languages (#1264, #1616, thanks @Emad211 and @bygordon130-hub). Persian (fa, right-to-left) and Turkish (tr) are registered in SUPPORTED_LANGUAGES and the language switcher, and all 11 locale files carry the same keys. The Turkish file arrived with English placeholders and UTF-8 decoded as Windows-1252; both were replaced before release, and a test now rejects copied-English Turkish strings and that mojibake in tr and fa. The custom-provider strings, first merged in English in every locale, are translated in all 11. Chat messages, the composer and error lines take their direction from their own text, so a Persian or Arabic message reads right-to-left inside an English UI and the reverse. Some strings are still English in some locales.

  • Pixel Office, a shared manual research-task board (#1616, thanks @bygordon130-hub). The /pixel-office page, the vibe-trading office CLI, authenticated GET /api/pixel-office and POST /api/pixel-office/actions, and the pixel_office agent tool share one transactional SQLite board (pixel-office.sqlite3 under the runtime root, mode 0600), so separate processes see the same tasks. Tasks are created, assigned to a role, completed with a required result, and roles carry Operator / Senior / Lead ranks. Assigning a task does not start an agent or place an order, ranks grant no trading permission, and a recorded result is not an execution audit. The PR proposed a JSON file plus browser state; the merge replaced it with this shared store. The unsigned Windows review installer now builds the bundled Python backend before packaging.

  • Scheduled jobs are edited in place, with channel-aware destinations (#1649, thanks @zeus229). PATCH /scheduled-runs/{id} and the Scheduled page change a job's prompt, cadence, timezone, end date and delivery without deleting it, so its id and history stay. A job that is running, or whose delivery is being sent, refuses edits with 409. Each channel adapter describes the destination it accepts (label, placeholder, input type), and Telegram suggests numeric private-chat ids already in its allow_from; a suggestion only fills the manual target field. Changing the destination clears the old delivery receipt but keeps the session link, and a timezone-only edit does not move an interval schedule's next run.

  • Scheduled Email reports as HTML or a PDF attachment (#1680, thanks @zeus229). Email delivery takes a format: plain text (the default), sanitized HTML (Markdown rendered by mistune and cleaned by nh3), or a PDF attachment with a short notice in the body. It is chosen on the Web Scheduled page, in CLI playbooks (--delivery-format) and in agent proposals, whose confirmations show it. A PDF that fails to render fails the send instead of going out as plain text. Where WeasyPrint's native libraries are missing, the PDF comes from the bundled ReportLab renderer with embedded fonts, so desktop installs get PDFs too. A scheduled send now fails visibly when the Email channel lacks consent or an SMTP host instead of being skipped with a log line. The format selector is translated in every UI locale.

  • Password-protected PDF reports (#1709, thanks @zeus229). An operator sets a private pdf_password on the Email channel; a PDF Email job can then request protection, and the attachment is encrypted with AES-256 through pypdf. Web, CLI playbooks (--protect-pdf) and agent proposals store only the yes/no choice, a job config containing pdf_password is rejected, confirmations show the choice, and creating, editing or confirming a protected job without a configured password is refused. An encryption failure never echoes the password.

  • The agent can save a report as a real PDF. Before this, write_file wrote Markdown text into a file named .pdf. A .pdf path now renders the Markdown into a paginated PDF with ReportLab, without Pango or Cairo. Text uses a bundled Noto Sans CJK subset (VibeCJK-Regular.ttf, OFL) covering Latin, Greek, Cyrillic, CJK, kana and Hangul; Arabic lines are shaped with arabic-reshaper and python-bidi (now base dependencies) in matplotlib's DejaVu Sans and wrap across pages without clipping. A character neither font covers fails the render instead of printing empty boxes, and the write is atomic, so a failed render leaves an existing file untouched. The result carries the path, the real byte count and an authenticated /api/reports/{id} link, which the reply adds even when the model leaves it out and the Web chat downloads; IM channels that accept files receive the PDF as an attachment, taken only from this attempt's staged reports, never from a path the model wrote. Staged copies live under generated_reports/ in the runtime root; if staging fails, the written file is kept and the result reports download_error. The bundled font adds about 19.9 MB to the package.

  • Shadow Account reports render a PDF without WeasyPrint (#1683, thanks @he-yufeng). When WeasyPrint is missing or fails, the report renders the same eight sections and charts through ReportLab in a simpler layout, and only then falls back to HTML-only; the payload's engine can now read reportlab. A process no longer loses the fallback after its first render on a host without a CJK font. The PR used Helvetica when no CJK font was found; the bundled font is now tried first, and a host where no embeddable font loads gets HTML-only, because an unembedded face can print blank pages.

  • Feishu joins the guided Web UI channel setup (#1572, closes #1614, thanks @shadowinlife). The Feishu panel gets field hints and a setup guide in every UI locale. Test connection is a standalone probe that requests a tenant access token from feishu.cn or larksuite.com, whichever the config's domain names, without starting the bot. stop() now closes the lark WebSocket and stops its dedicated event loop, so a hot reload no longer leaves the old connection answering messages beside the new one.

  • Slack and Discord join the guided Web UI channel setup (#1757, #1758, #1764, #1777, closes #1763 and #1776, thanks @shadowinlife). Both get localized forms and read-only credential probes. Slack checks the bot token with auth.test, requiring a bot identity so a user token is refused, then the app token and Socket Mode with apps.connections.open; only Socket Mode is supported. Discord requires a bot identity from users/@me; privileged intents cannot be checked over REST, so the setup guide covers them. The shared token probe, also behind DingTalk, QQ and Feishu, now reports HTTP 408 and 429 as a temporary network failure rather than rejected credentials, and a failed probe leaves saved settings unchanged. Discord's allowlists, group policy and emoji settings, and Slack's allowlist and emoji, refresh without reconnecting. #1757 and #1758 landed through #1764.

  • Token usage in Run Detail, chat and swarm (#1890, closes #1889, thanks @shadowinlife). Run Detail gains a Token usage tab built from the run's existing llm_usage, chat shows a live counter from the llm_usage stream event (replay cannot double-count), and swarm cards and the dashboard show run and per-agent totals. GET /swarm/runs/{id} now returns token totals, provider and model, and swarm workers record prompt-cache tokens and whether a count is provider-reported or estimated; an estimate is labelled as one and a missing label reads "unlabeled". A reported zero stays zero beside a nonzero count, and a missing cache field is never shown as 0. These are token counts, not cost or cross-run totals.

  • get_southbound_flow serves the Stock-Connect net flow that is still published (#1486, closes #1483, thanks @shadowinlife). The new read-only agent and MCP tool returns market-wide southbound daily net buy into Hong Kong, split into the Shanghai and Shenzhen channels, in 100M HKD. Eastmoney's datacenter report is the primary source; Eastmoney's NET_DEAL_AMT arrives in million HKD and is scaled to the envelope unit, and a rejected report is an error rather than an empty history. The keyless fallback is HKEX's official daily statistics (Buy minus Sell turnover), which serves only the latest trading day. There is deliberately no Tushare path: its southbound fields now carry the cumulative stock, not a daily flow.

  • Gildata research tools and off-exchange fund NAV (#1485, thanks @sambazhu). With GILDATA_TOKEN set, three tools wrap Gildata's natural-language endpoint: get_cn_macro_series (China macro, regional and industry series), get_cn_announcements (A-share, HK and fund announcements) and search_broker_reports. Each echoes the vendor APIs its query was routed to and returns at most 200 rows (default 50); GILDATA_SRV_TOOL_URL overrides the endpoint. get_fund_nav returns unit, accumulated and adjusted NAV for funds such as 110022.OF, and only those three NAV fields count as price evidence for grounding. The Gildata loader now also serves Hong Kong equities, stamped raw, and CN indices, each resolved only when the vendor's ref_code matches the symbol exactly. It stays at the tail of each chain so paid quota is spent last. Without a token none of this registers.

  • Run cards flag backtests that may sit inside the model's training data (#1618, closes #1613, thanks @lorenzozanee). Strategy writes record the provider and model that produced config.json and code/signal_engine.py, and the run card shows them. Set model_training_cutoff (an ISO date) in config.json: a window ending before it is marked as exposed, and an unset or invalid cutoff is reported as unknown and treated as exposed. No cutoff is inferred from the model name. vibe-trading alpha bench --strict takes --training-cutoff to start the out-of-sample segment after that date, and now returns an error instead of ok when no alpha rows qualify.

  • Loader-health reports say why a source came back empty (#1643, #1698, thanks @cgycorey). The weekly public loader-health lane attaches up to three of the loader's own warnings to an unhealthy row, so an anti-bot refusal and a broken loader no longer look identical. The report is a public artifact: each message is capped at 200 characters, URLs and credential-shaped values are redacted (including a labelled Authorization: Bearer token: VALUE), and a warning that contains a filesystem path is omitted in full. The health tests now register real loaders before stubbing one, so running a single test no longer reaches the network.

  • A free BaoStock fallback for A-share val:* panels (#1799, thanks @Jackzigen). Without a Tushare token, or when Tushare returns nothing for a field, val:pe_ttm, val:pb and val:ps_ttm come from BaoStock's daily peTTM / pbMRQ / psTTM. val:pe and val:dv_ttm have no BaoStock counterpart and stay NaN rather than being filled from another field. A BaoStock outage or login failure degrades to NaN panels, not an exception. Token holders see no change.

  • Hong Kong orders are sized on each symbol's board lot (#1813, thanks @alexwan02). A backtest config may carry hk_lot_sizes: {symbol: lot}; 00700.HK, 0700.HK and 700.hk share one entry and five-digit counters such as 80700.HK stay distinct. Any HK symbol without a configured lot is sized on 100 shares and listed under lot_assumptions in the run metrics, also when no table is given. Mixed-market runs carry each sub-engine's assumptions under market_diagnostics.

  • Brazil (B3) market data and an opt-in ZimaOS template (#1783, thanks @incorpore). A .SA symbol is market br_equity, quoted in BRL and served by yahoo → yfinance → b3_cotahist → local, with BOVA11.SA as the default benchmark. The Brazil engine applies configurable B3 fees, R$0.01 ticks and a default lot of one share (br_lot_size=100 for round lots); it does not model income tax, stock-borrow cost or settlement cash locks. python -m src.br_data sync stores official B3 COTAHIST, CVM, Banco Central and Tesouro Direto data outside the checkout, and the get_br_market_data agent tool reads it. COTAHIST bars are raw, and CVM statements are latest revisions, not point-in-time inputs. zimaos/ holds a Compose file and a fork publishing example that each need explicit setup; upstream gains no scheduled publishing job.

  • A complete pip + Telegram quick start in the README (#1954). It covers installing the Telegram extra, sender authorization with allow_from, matching the API port between serve and channels start, and recovering when vibe-trading channels is missing because the package was installed into a different environment.

  • Thirteen fundamental and valuation factors (#1787, #1788, thanks @shadowinlife). Ten read point-in-time SEC XBRL filings: accruals, book-to-market, cash-to-assets, CFO yield, earnings growth, FCF yield, leverage, operating margin, revenue growth and ROA. Growth is year over year through a 365-day as-of lag, and earnings growth divides by the absolute prior value, so a loss turning into a profit reads as growth. Three read a new val:* namespace of daily A-share valuation data (fund_pb, fund_ps_ttm, fund_dividend_yield). Those are keyed by trade date and attached only on the CSI 300 universe; a nonpositive P/B or P/S is excluded rather than ranked as cheap. XOM keeps its current SEC filer, and filings by its pre-reorganisation filer that predate the change extend its history. Basic shares never stand in for missing diluted shares, and price-times-diluted-shares ratios are valuation proxies, not reported market capitalisation.

  • Industry-neutral A-share value factors (#1802, closes #1789, thanks @shadowinlife). fund_pb_ind_neutral, fund_ps_ttm_ind_neutral and fund_dividend_yield_ind_neutral z-score each name against its Shenwan (SW2021) level-1 peers on each date instead of the whole cross-section. A CSI 300 benchmark attaches a point-in-time sector tag built from Tushare index_member_all membership dates, and only when at least 90% of names carry a label. A name with no label, or a group too small for a statistic, stays NaN; there is no fallback to a whole-market z-score. Membership is point-in-time, but the SW2021 taxonomy itself is applied retroactively, and each factor's notes say so. With #1787 and #1788 the registry holds 478 factors, 20 of them fundamental (462 and 4 in 0.1.16).

  • Factor benchmarks at a longer horizon (#1801, #1802, thanks @shadowinlife). alpha bench --horizon N (1 to 252, also on the tool and MCP wrapper) measures IC against N-bar forward returns, sampled without overlap and anchored at the end of the panel; a missing close never bridges the horizon. horizon=1 is the previous next-bar formula. Result envelopes now state their horizon, sampling and universe_excluded list. SEC-fundamental factors requested on a universe other than the S&P 500 are listed as excluded instead of failing the whole run, and alpha compare shows them as skipped with the reason. The alive/dead thresholds were calibrated on daily IC and are not re-derived for N > 1.

  • Factor benchmarks feed fund:* factors their SEC inputs. Benchmarks now load each selected factor's declared fundamental fields as point-in-time SEC panels; before, those inputs never reached the benchmark. This needs the S&P 500 universe. The S&P 500 price panel goes through the shared fallback chain and reports requested, fetched and missing instruments and the price sources used.

  • Factor neutralisation and PCA in the quant library (#1817, thanks @BHhhh-123). neutralization.neutralize_factor regresses a factor on named numeric and categorical controls (OLS or WLS) separately at every time on the panel, at any bar frequency. Controls are either all required or selected greedily on coverage, rank and sample size, and every time records what was dropped and the exposure before and after. pca_factors fits cross-sectional PCA per time. Neither function fills, pools or resamples rows, and a rank-deficient time is skipped instead of being demeaned. The caller has to supply exposures known at that time. The finishing commit rescales inputs before squaring so very large or very small units neither overflow nor change the answer, keeps covariance PCA in common units, and breaks equal-loading sign ties by input order. Reachable through quantlib_call and MCP.

  • Technical indicators can be read as of a date (#1768, thanks @zeus229). technical_indicators accepts an optional end_date (strict YYYY-MM-DD; anything else is an error) as the acquisition boundary. Every response carries read_identity: the boundary mode, requested end date, bar count, serving provenance and a SHA-256 fingerprint of the exact interval, dates, closes and volumes used, so two reads can be shown to have used the same bars.

  • A claude-cli provider that runs on the local Claude Code login (#1804, closes #1800, thanks @alexwan02). Each turn runs one claude -p --restricted --strict-mcp-config --tools "" subprocess that only picks the next step (call tools or answer); Vibe-Trading executes the tools itself, so the evidence ledger, the grounding gate and swarm delivery checks run as they do for any other provider. No API key is needed. CLAUDE_CLI_DENY_TOOL_PREFIXES (default trading_,etoro_) hides tools from the model and CLAUDE_CLI_BASH_POLICY (default python-only, or off) limits bash to python scripts and read-only listing commands; refused calls are fed back at most three times. Both are convenience filters, not a security boundary: a Python script can still do anything, and live orders stay behind the mandate gate. Settings lists the models the login can run and offers only the reasoning efforts the chosen model accepts. A decision can take minutes, so the provider emits a waiting signal every second; the idle timeout no longer discards an answer still on its way, the stop button kills the child process, and progress consumers see llm_waiting.

  • The backtest tool returns a structured summary (#1646, closes #1644, thanks @shadowinlife). A successful result carries summary: the run card's scalar metrics, structured_metrics and validation, run metadata, an equity preview of at most 50 points with the first and last bar always kept, and artifact paths including each symbol's ohlcv_<CODE>.csv. The metrics no longer reach the model only as JSON inside a stdout cut to its last 2,000 characters. A missing or corrupt run card omits the summary and leaves the run successful, as does one whose artifacts resolve outside the run directory. When the envelope would exceed the 10,000-character tool-result cap, logs are dropped first, then the preview is thinned, and anything still omitted is named in the result.

  • read_run_artifact reads run artifacts as structured pages (#1647, closes #1645, thanks @shadowinlife). An agent and MCP tool with three modes: rows (offset pages of whole records with next_offset), downsample (equal-stride sample, endpoints kept) and meta (columns, row count, size and the run card's sha256 for the file). Aliases (equity, trades, metrics, positions, target_positions, ohlcv:<CODE>, run_card) resolve by a fixed map; any other path is served only if the run card's artifact manifest lists it, and paths or symlinks that leave the run directory are refused. Pages fit the same 10,000-character cap the agent and swarm apply, CSVs are scanned without loading the whole table, and zero-padded codes such as 000001 stay strings. Rows shown this way ground a report only while the file is an unchanged engine-written table; a file the model wrote never does.

  • Grounding checks are declared in one registry, and each validation names the checks that fired (#1628, #1661, #1691, thanks @he-yufeng). Each check has a stable name, its issue codes, a description, a severity and a phase. Fifteen are declared: the identity, answer and declaration checks from these PRs, plus the per-figure role checks (observed, derived, proposed, cited, count) and the price-provenance checks, which a maintainer follow-up registered. Every attempt recorded in grounding_evidence.json carries fired_checks and check_findings, taken from the predicates that actually returned a finding, so role checks that share numeric_claim_conflict do not implicate each other. Predicates and their order moved unchanged, so verdicts are the same. CONTRIBUTING.md describes how to add a check with its passing and failing answer pair.

  • The session iteration cap is configurable (#1833, thanks @sambazhu). VIBE_TRADING_AGENT_MAX_ITERATIONS sets how many ReAct iterations a session run (Web UI, API) may take before it stops with "reached max iterations". The default stays 50 and values below 1 are rejected. The no-progress budget and spending limits still apply.

Changed

  • The live runtime scheduler runs on the shared scheduler kernel (#1899, thanks @shadowinlife). It keeps its public surface, its interval and one-shot job syntax and its sleep-until-the-earliest-deadline wakeups; the mandate gate, kill switch and market-hours layers are unchanged. A second concurrent stop() now waits for the first one's teardown instead of returning at once.

  • Broker connector documentation and tests (#1770, #1793, #1809, closes #1724, thanks @lorenzozanee, @alexwan02 and @he-yufeng). The broker bring-up checklist in CONTRIBUTING.md asks for options evidence: write-side argument names and types recorded apart from read shapes, unsupported shapes kept gated, and supported ones tested against mandate caps on premium × multiplier. The readonly field on nine SDK connector configs is documented as what it is: the order methods do not read it, and order access is decided by the connector profile. IBKR's pool reset test no longer depends on test order.

  • Tencent, Eastmoney and AKShare A-share prices move to the multiplicative forward-adjusted basis when they can (#1551, closes #1541, thanks @he-yufeng). These sources serve dividend-additive qfq series, so returns on them are not total returns. The loaders now also fetch the raw series and, when qfq - raw forms valid per-action plateaus, rebuild the series with (prev_close - dividend) / prev_close factors and stamp it split_dividend. Any inconsistency (misaligned calendars, drifting offsets, implausible dividends, a failed raw fetch) keeps the additive series with its stamp and warning. A window with no corporate action keeps its static stamp, and the series is anchored at the window's end, so absolute levels depend on the requested range while returns do not.

  • mootdx is retired from the default A-share chain and the public canary (#1730, closes #1729, thanks @he-yufeng). The TDX servers stopped answering mootdx's protocol in 2026-07, and every automatic A-share fetch spent about ten seconds there before falling through. The loader stays registered for explicit source="mootdx". A saved source-order override that still lists it is no longer a permutation of the default chain, so it is ignored with a warning and the default order applies.

  • Every per-loader test suite checks the health lane's OHLCV contract (#1725, closes #1723, thanks @Jackzigen). agent/tests/loader_contract.py holds the column, index, dtype, finiteness and OHLC-ordering checks the loader-health lane applies, and each loader's canonical-frame test calls it, so a drifting source fails its own suite instead of waiting for the weekly run.

  • An MCP tool result reaches the agent once, not up to four times (#1634, thanks @zeus229). A structured result used to arrive as data, an identical structured_content, a text block restating it and the joined text. Only exact copies of data are dropped, and only from what the agent reads; MCPServerAdapter.call_tool still returns every surface to programmatic callers, and error or text-only results are unchanged.

  • The system prompt is byte-stable across runs (#1708, closes #1707, thanks @Jackzigen). The workspace state (the run directory and tool counters) changed the system prompt on every turn and tool call, which invalidated a provider's prefix cache from that point on. The state now rides in the first user message inside an <agent-state> envelope, next to <recalled-memories>; the token count is the same.

  • loop.py is split into modules, first slices (#1636, #1631, #1662, thanks @Jackzigen, @he-yufeng and @Mikesad). Tool-result helpers moved to src/agent/tool_results.py, context compaction and message repair to src/agent/context_compact.py, and LLM usage accounting to src/agent/llm_usage.py, byte-identical and re-exported from loop. Log records from the moved code now come from the src.agent.tool_results and src.agent.llm_usage loggers. A test that monkeypatches loop.X no longer reaches a function that now lives elsewhere; patch the module that holds it.

  • loop.py is split into modules, remaining slices (#1641, #1678, #1791, thanks @Jackzigen). The final-status ladder moved to run_outcome.py, the tuning getters and TOKEN_THRESHOLD to tuning.py, and written-file tracking with its pending-write directive to write_targets.py. Maintainer follow-ups moved the tool-call lifecycle (tool_lifecycle.py, still logging as src.agent.loop), per-turn streaming callbacks (loop_streaming.py) and final-answer grounding orchestration (loop_grounding.py). loop.py went from 3,761 lines at 0.1.16 to 1,628. One stated difference: the degraded flag on the end event and result now comes only from success branches, so a run cancelled or stalled between a fallback release and the status decision is no longer marked degraded.

  • Scheduled research, the Brazil data sync and memory consolidation run on one scheduler kernel (#1893, thanks @shadowinlife). agent/src/scheduling/ holds the trigger grammar, the next-due arithmetic (pinned by 133 golden vectors recorded before the move) and a tick loop with bounded retry backoff. The scheduled-research executor uses it, and its REST contract is unchanged. The FastAPI lifespan now owns the scheduler, and the unused cron_service parameter is gone. The live-trading scheduler was not part of this change. The kernel also adds protected system: jobs. They are seeded at startup, a user cannot create an id with that prefix, and REST and CLI refuse to edit or delete one. Both system jobs are opt-in and need VIBE_TRADING_ENABLE_SCHEDULER=1:

    • system:memory-consolidation runs daily at 04:00 UTC when VT_MEMORY_CONSOLIDATION is on. VT_MEMORY=on/full does not turn it on, and it does nothing unless VT_MEMORY_GC is also on.
    • system:br-data-sync needs VIBE_BR_SYNC_ENABLED=1 and replaces python -m src.br_data daemon, which now prints a migration notice and exits 0. VIBE_BR_SYNC_TIME sets the slot only when the job is first seeded, and the first sync waits for that slot. The ZimaOS sidecar container is gone. Its Compose file leaves both flags off by default (068ec3bd; the PR had turned them on).

    Since 068ec3bd, a source that fails inside the sync fails the attempt, so the scheduler's retry budget applies. Before, the sync was recorded as a success. Before downgrading to a version without system jobs, delete the system: rows from scheduled_research_jobs.json. Older readers reject those rows and quarantine the whole file.

  • A chat message can be up to 100,000 characters (#1701, thanks @mminkus). The limit used to be 5,000 characters, while scheduled-run prompts had no limit. A longer message was refused before reaching the model, the Web UI showed only "Failed to send message, please retry", and the 422 response repeated the whole prompt. Since f6e72428, the 422 returns a short message_too_long body naming the limit, without the prompt. The Web composer checks the length before sending, counting code points as the server does. Goal prompts stay limited to 5,000 characters.

  • Agent chat loads small run previews instead of full runs (#1853, closes #1852, thanks @alexwan02). GET /runs/{run_id}/preview returns only the metrics, the first 1,000 equity points and whether the run deserves a card. Agent chat requests previews for a whole session at once. Before, it fetched each full run one at a time; the PR measured about 70 MB for a session with one backtest card. GET /runs/{run_id} now builds and encodes its response off the event loop, so other requests no longer stall; its bytes are unchanged. One run type loses its chat card: a run whose only output was price data that the full response rebuilt online, with no metrics, equity, trades, positions, validation, strategy, price or OHLCV files and no run card.

  • Dependency updates (#1665, #1666, #1667, #1668, #1692). Python: 46 packages in requirements-lock.txt moved, including langchain 1.4.3, langchain-core 1.6.6, langchain-openai 1.6.7, fastapi 0.142.2, starlette 1.7.0, uvicorn 0.54.0, pyjwt 2.15.1 and akshare 1.19.1. The loader-health environment uses baostock 0.9.4. Frontend: React 19.3, Vite 8.3, react-router 8.4 and Vitest 5.0.2, with @vitest/coverage-v8 moved to 5.0.2 to match. CI: the Docker build actions, cloudflare/wrangler-action 4.1.3 and actions/upload-artifact 7.0.1. uuid-utils stays at 0.17.0 because 1.0.0 (#1670, not merged) conflicts with the uuid-utils<1.0 requirement of langchain-core and langsmith.

  • Docker image pins move past known advisories. requirements-lock.txt and requirements-channels-lock.txt, which the Docker image installs with --require-hashes, now pin aiohttp 3.14.5, certifi 2026.10.19, urllib3 2.8.0, multidict 6.9.1 and pillow 13.0.0 (ccxt 4.5.89 comes along so the newer certifi resolves). A plain pip install already resolved newer releases. mcp stays below 1.30 for the IBKR OAuth constraint, so CVE-2026-59951 (Streamable HTTP sessions are never reclaimed) remains open for vibe-trading-mcp --transport http.

  • For contributors running the suite: symlink tests skip where symlinks are unavailable (#1803, thanks @hunters1011-ctrl). Three positions/sectors API tests used to fail on Windows hosts that refuse to create symlinks (WinError 1314). They now skip there and still run everywhere else.

  • Channel edits to live-read settings apply without reconnecting (#1737, closes #1734 and #1736, thanks @shadowinlife). Each adapter declares the settings it reads per message (allowlists, emoji and similar); an edit that changes only those refreshes the running channel in place. The comparison uses validated values, so a full Web form echoing unchanged connection fields does not count as an edit. A refresh copies only the declared fields and keeps connection state resolved at login, such as WeChat's authenticated server address, and Signal refreshes its writable dm / group policies rather than its computed allowlist. The Web result now says whether the channel was refreshed, reconnected, restarted with the others, or will pick the change up when channels start, in every UI locale. When a hot swap degrades to a full reset, the reason is returned with URL userinfo and old and new secret values removed, and a best-effort notice goes to the changed channel's last chat.

  • Channel status records runtime failures (#1738, closes #1735, thanks @shadowinlife). A failure to start a channel, to send after the last retry, or to stop the old adapter during a reload is recorded on that channel's status entry as a one-line error of at most 200 characters with URL userinfo removed, plus error_context (start, send or reload_stop) and error_at. The traceback stays in the log. A successful rebuild clears it.

  • New channels are checked against the authoring contract (#1630, thanks @he-yufeng). A test walks every discovered channel and fails when one leaves a BaseChannel abstract unimplemented, has a scalar config key without a field hint, or has a credential-shaped key that would reach the browser unmasked. CONTRIBUTING.md gains an "Adding a Channel" recipe.

  • Web UI navigation is grouped and explains each page. The sidebar groups pages into everyday research, specialized analysis and automation & setup, with plainer names ("Research assistant", "Backtest reports", "Trading status"). Every page except the chat has a collapsible "How to use" panel, and the welcome screen offers three editable starter tasks (research a stock, test a strategy, review holdings). Math and code highlighting load only for completed answers. API requests time out after 120 seconds (VITE_API_TIMEOUT_MS at build time; 15 seconds for the session list) and say whether a timed-out write may still be processing; writes are never retried automatically. Session-list, rename and delete failures are shown instead of ignored.

  • Analysis pages keep their work across navigation. Alpha Zoo, Correlation and Options Lab keep their inputs for the browser session. An Alpha Zoo evaluation runs as a server job the page can reconnect to after leaving or reloading; a resubmitted request with the same id returns the existing job rather than starting another, and a job lost to a server restart says so. Before running, the page states each universe's data requirements: CSI 300 needs a Tushare token and permissions, S&P 500 uses public data, and BTC-USDT is refused as a single asset.

  • The runtime page lists connectors by status (#1845, thanks @alexwan02). Connected connectors come first, then those needing attention, then the rest; peers keep the backend's alphabetical order, so the 15-second poll never reshuffles them. Only the order changes. The page header no longer calls the page read-only, since model cadence and spending limits can be edited there while a runner is stopped.

  • Report rendering dependencies (#1680, #1683, #1709, thanks @zeus229 and @he-yufeng). The base install now requires mistune, nh3 (previously only in extras), pypdf, reportlab, arabic-reshaper and python-bidi. The Docker image adds fonts-noto-core and fonts-noto-cjk so WeasyPrint renders CJK and Arabic reports.

  • Update the Windows desktop PyJWT pin to 2.15.1. This matches the main lock and includes the fix for GHSA-ffc3-869f-jxw9. Artifact hashes are verified against PyPI; other Windows dependency pins stay unchanged.

Fixed

The grounding gate

  • Indexed and container field refs resolve, and a miss lists the refs that would work (#1633, thanks @zeus229). a.0.b and a[0].b now name the same evidence element; an unindexed field still selects none. When a call::field ref matches nothing, the correction lists exact call_id::path[i] refs, value matches first. A ref naming a container lists its numeric descendants that hold the figure's value, and a tool name shared by several calls ranks those calls by value match, most recent first. These are hints only: no hint authorizes a value, and a maintainer follow-up keeps them to the figure's own instrument and to the element an indexed ref names.

  • A ref naming no call of the session gets real refs to copy (#1638, thanks @zeus229). A scope::field ref whose scope names no call, tool or run stays rejected, now with the unknown_call_id reason and up to five exact call_id::field candidates. The system prompt tells the model to copy tool_call_id verbatim instead of inventing a short alias. A maintainer follow-up applies the same instrument and figure-kind limits a real ref has, so a candidate is never one that would be refused.

  • A correction keeps the figures that already passed (#1702, thanks @he-yufeng). The correction prompt listed only what failed, and a model facing 65 issues rewrote whole numeric sections as prose, deleting valid Risk X-Ray values and a benchmark table that had no issue. It now also lists up to 24 measured figures that checked clean and tells the model to keep them as written. A value that fails anywhere in the draft is not listed, and bare integers never are. When more than 24 issues exist, the prompt says how many are not shown and that unlisted figures are not assumed to pass.

  • Only the engine's own output grounds a backtest report (#1685, thanks @cgycorey). The active run's metrics.csv could credit an earlier run's Sharpe to the current call; that archived copy now counts only when its manifest names the full directory the call declared and the exact call id, and a model-written manifest is ignored. A metrics.csv or metrics.json the model wrote is no longer engine evidence, a date cell with an explicit year must match that year, and a resolver step that failed, was refused or came back empty no longer erases a locked or conflicting identity.

  • A derived figure's ref must name evidence that exists (#1835, #1837, thanks @zeus229). A comma- or semicolon-separated ref is resolved member by member, and an empty or unknown member fails with no_evidence instead of falling back to every value in the session. A single ref that names no call, tool, field, artifact or observed symbol is refused the same way rather than borrowing a coincidentally equal value. Unattributed multi-symbol arithmetic keeps its no_symbol diagnostic.

  • Derived notes after a label parse, and process metadata is not evidence (#1728, #1767, thanks @zeus229). A note such as scenario result: 125 / 10000 * 100 was not evaluable because only the full-width : was a separator; the ASCII : now is too. Top-level exit_code, returncode, elapsed_seconds and duration_seconds of a code-running tool's result are no longer recorded as observed values, so a backtest's runtime cannot ground a figure.

  • Percent-named fields are read as percentage points (#1823, closes #1820, thanks @he-yufeng). A-share rows carry changepercent values that are already points, so 0.52 - 1.41 refused the correct "-0.89%" and passed "-89%". When every data operand anchors only to fields named percent, pct or ending in %, the result is read in points; an operand whose value also appears under a plain field keeps the fraction reading. A maintainer follow-up converts the operands before evaluating, so ratios and compound factors of two point values come out right.

  • Every data operand of a derived figure must be observed. Only added or subtracted terms were checked, so a numerator or denominator the session never fetched could pass. A chosen multiplier or divisor (a discount factor, a weight) now needs a count declaration and its value and label on the same visible line of the answer. Literal 1 and square or cube exponents are structural, and an explicit trailing *100 on a percent claim is read as a unit conversion without exempting any other operand.

  • Instrument identity survives matrices and ticker-keyed results (#1850, #1851, thanks @zeus229). In a square Markdown matrix with an empty top-left header (a correlation table), each row's symbol is taken from its first cell, bold removed, when the row and column symbol sets are complete, unique and equal; anything else stays unscoped. A result keyed by instrument (risk_contribution_pct.ABC.US) keeps each value's instrument and exact ref, so the same number cannot ground a claim about another instrument. List-indexed scalars such as positions[0].contribution_pct keep their indexed-ref meaning.

  • A figure that names its currency must match evidence in that currency (#1905, thanks @ThreeTima). An ISO code or an unambiguous mark (€, £, ₩, ₹, 人民币, 美元, 円) fixes the figure's currency for observed, derived and proposed checks, and SEC XBRL amounts carry their declared unit (CHF, NZD, BRL, ARS and VND included). A scaled amount such as 1.2B USD keeps its code. Bare $ and ¥ stay ambiguous and match any currency.

Agent loop, providers and swarm

  • A restored read-only result stays visible until the model has seen it (#1635, thanks @zeus229). Compaction could clear a result replayed from the run cache before the next model request carried it, so the restore was wasted. Replayed results are now held back from microcompaction, text collapse and the summary layer until one successful model request has included them.

  • Writing a strategy file no longer fails with AttributeError (#1673, closes #1672, thanks @tomaszkubiak-courses). After the model-provenance writer landed, every run that wrote a strategy file crashed reading _active_model_id. The loop now sets it from the runtime snapshot and, before each batch of tool calls, from the provider-reported model, falling back to the configured one. This regression never reached a release.

  • Codex tools may omit optional arguments (#1785, thanks @zeus229). The Codex Responses endpoint treats a tool without an explicit strict as strict, so the model filled every optional property, sending "" for an optional string. Tool schemas now go out with strict: false.

  • Skills load and save under the names they advertise (#1797, closes #1795; #1858, closes #1855; thanks @Shizoqua). Two section headings that differ only by case could not be loaded, and the path the error suggested repeated the error; an exact-case match is now tried first. Saving a skill whose own frontmatter carried another name left it unreachable under its slug and let each lookup append a duplicate; the name: line is now rewritten to the slug on every save, and a fallback read never duplicates a loaded skill.

  • Research-goal evidence keeps its provenance inside the run (#1688; #1895, closes #1894; thanks @zeus229 and @Shizoqua). The goal tool and skill tell the model to copy a result's tool_call_id verbatim and never to use a ledger evidence_id (ev_...) as a grounding ref; tool_call_id stays optional for manual evidence. An absolute artifact path outside the current run directory is now rejected, as a relative escape already was, instead of being verified against another run's file.

  • Each swarm task gets its own artifact directory (#1455, thanks @Shizoqua). Artifacts were keyed by agent alone, so when one agent ran several tasks, a later task could read an earlier task's report.md back as its own output. They now live under artifacts/<agent>/<task>/, and earlier tasks' reports, summaries and message logs stay in place; sequential and parallel runs are covered through the real runtime.

  • Swarm presets take their variables from the prompt (#1674, thanks @Mikesad). Seven presets ran on fixed defaults whatever was asked: a copper request ran every commodity worker on gold, a bearish options request used a neutral view, and crypto, factor, event, fund and macro prompts were replaced the same way. These variables are now extracted from the prompt, an explicit duration ("3 months", "两周") is used as stated, and run_swarm accepts a variables object whose values override extraction. Prompts that match nothing keep the old defaults.

  • A half-written last line no longer breaks a swarm event stream (#1908, thanks @rudycelekli). read_events raised on an events.jsonl whose final line was still being appended or was cut short by a crash, which killed the SSE stream and run recovery. An unterminated last line that is not yet valid JSON, including a cut UTF-8 sequence, is skipped until the next poll; a complete invalid line still raises.

Market data and loaders

  • get_northbound_flow stops returning an empty success off a dead feed (#1484, closes #1481, thanks @shadowinlife). Eastmoney's northbound endpoint still answers but has carried no net flow since HKEX's 2024 disclosure change. When both its realtime and history sides are empty or all zero, the tool now uses the Tushare fallback or returns ok: false naming both failures. The fallback no longer multiplies every row by 100: values stay in million CNY, rows before 2024-08-19 are net buy, and later rows, which are turnover, come back under turnover_* fields with their net-flow fields null.

  • Daily fund-flow fallback covers empty responses and labels its source (#1954). An A-share daily request now tries Tushare when Eastmoney returns no usable rows, not only when the request or symbol resolution fails. Each symbol names the provider that served it, and a batch served by both reports mixed. Tushare's main is now large plus super-large net flow; net_mf_amount, which is total net flow, is returned separately as net_total. A missing or non-finite buy or sell amount stays null instead of becoming zero, and the result states its methodology because provider bucket definitions differ. Minute requests never fall back to daily data, and a missing token or permission remains a per-symbol error.

  • A blocked Stooq costs one request (#1637, #1640, closes #1648, thanks @cgycorey). Once Stooq refuses the client, the loader stops requesting further symbols instead of only silencing its warning, so the fallback chain moves on at once. The refusal is detected from the response before parsing: the proof-of-work page, a 403 or 429, or the plain-text daily-hits message. Concurrent callers serialize on the check, and the latch lasts 300 seconds instead of the life of the process. A 5xx is still a per-symbol failure and N/D is still "no data".

  • FMP and Tiingo series never mix adjusted and raw prices (#1684, #1689, thanks @cgycorey and @he-yufeng). The price basis is decided once per response from bars that can actually be emitted. A bar whose adjusted prices are missing or invalid is dropped, with a warning, rather than filled with raw prices beside adjusted neighbours. Bars with unparseable dates or non-positive prices are dropped too. Each frame carries its actual adjustment stamp, the loader cache keeps it (the cache format version was raised, so older entries are refetched), and tests pin the stamp on both paths and through a warm cache hit.

  • Local data files load every bar they should (#1651, #1825, #1819, thanks @rudycelekli and @he-yufeng). The end-date filter is half-open at the next midnight, so bars stamped in the last fractional second of the end day are no longer dropped. A Parquet file whose dates live on a DatetimeIndex loads with or without a configured date_format; before, setting a format silently dropped every bar. An explicitly mapped date column takes precedence over the index, which also removes a crash when both share a name.

  • The opt-in loader cache follows changes to local sources (#1650, #1747, #1879, thanks @rudycelekli). Local cache entries are keyed by the file, query, date format and column mapping as well as the range, so repointing a symbol no longer serves the previous source's bars. Correcting or deleting a CSV, Parquet or DuckDB file (including a DuckDB WAL change) refreshes the entry on the next fetch, reusing the same cache paths. Cache metadata that is missing, malformed or unusable is a cache miss that the loader repairs, not an exception. An edit that preserves every observed filesystem identity field is still not detected.

  • The Longbridge loader works on SDK 4.0 and later (#1805, #1847, thanks @alexwan02). SDK 4 refuses Config(app_key, app_secret, access_token), so every Longbridge fetch failed and auto mode fell through even with valid credentials. The loader and the trading connector now share one helper that uses Config.from_apikey when the SDK has it, and the initialization error names the exception type. One loader test covers both SDK shapes through windowing and caching.

  • BaoStock 0.9.4 connections keep their timeout. The bounded connection wrapper accepts the new API-key argument and routes bs- keys to the VIP host while staying compatible with 0.9.3.

  • Data tools honour their row limits (#1676, #1927, thanks @cgycorey and @rudycelekli). get_macro_series reports observations_available, truncated, the effective limit and a hint when it caps a series, and trims whole observations to fit the tool-result budget so count is what was delivered. A FRED response covering only part of the requested window is an error. get_fund_nav clamps an explicit limit=0 to one row instead of returning the 2,000-row default.

  • Symbol suffixes resolve to the right exchange (#1907, #1914, thanks @rudycelekli). get_etf_holdings treats Yahoo's 510050.SS like 510050.SH instead of sending it down the US/SEC path. Symbol search labels Beijing codes (920xxx and legacy 4xxxxx / 8xxxxx) .BJ instead of .SZ, and the Tushare fallback no longer maps 920xxx to .SH.

  • RSSHub event enrichment keeps calendar dates and share classes (#1829, #1830, thanks @rudycelekli). A timezone-aware as_of no longer raises, and aware daily bar labels keep their displayed date instead of shifting to the UTC day. Configured bare and exchange_prefix routes split only the final venue suffix, so BRK.B.US requests BRK.B rather than BRK. Intraday publication instants are not converted between markets.

  • SEC quarterly statements derive fiscal Q4 only for currency flows (#1917, thanks @rudycelekli). period="quarter" built Q4 as FY minus three quarters for every concept, which produced negative weighted-average share counts and EPS figures no filing reports. Q4 is now derived only for three-letter currency units that match across all four periods; share counts, per-share values and dimensionless ratios are left out of the derived row.

  • Correlation matches the data it compares (#1710, #1711, #1746, #1933, closes #1932, thanks @rudycelekli, @cgycorey and @Jackzigen). Daily histories from sources in different time zones are joined on each source's local calendar date in both correlation and regime analysis. Equity tickers that end like a quote asset (ABNB) are no longer routed to crypto sources. 700, 0700, 700.HK and 00700.HK are one instrument for the duplicate-asset guard, while five-digit counters such as 80700.HK stay distinct. A pair whose correlation is undefined (a zero-variance leg) is null from /correlation and /correlation/analysis, not 0.0, and the heatmap shows a muted dash with an explanation.

  • Correlation analysis reports per-symbol coverage. /correlation/analysis computes the matrix and the optional regime timeline from one price snapshot, rejects two spellings of the same asset before any fetch, and returns which symbols are missing with every source tried and why it failed. The fetch honours saved source-order overrides and tries Yahoo first for ^, =F and =X identifiers.

Backtest runner, engines and metrics

  • Profit factor is undefined when no trade lost (#1602, thanks @davidalmeida90). profit_factor and profit_loss_ratio were 0.0 for a run whose trades all won, which ranked it below every run with a loss. Both are now None (an empty metrics.csv cell, null in JSON) in the daily metrics and the options engine. A run with no closed trade, or only losing trades, keeps 0.0.

  • Sortino uses zero-target downside deviation over every period (#1693, #1727, thanks @MetaAviator). The denominator was the sample standard deviation of the losing returns, so equal losses looked riskless and a single loss left the ratio huge or undefined. It is now the root mean square of negative returns divided by the full return count, in the shared metrics and in the options engine.

  • The first bar's result counts against initial capital everywhere (#1750, #1818, #1844, closes #1749 and #1843, thanks @he-yufeng and @DrDLyuan). The options engine's drawdown peak and the drawdown column in artifacts/equity.csv start from initial cash, so a first-bar loss shows up and the column agrees with max_drawdown. Per-bar returns measure the first bar from initial cash too, so Sharpe and Sortino see it, the ret column compounds to total_return, and bootstrap validation uses the same series.

  • Calmar stays finite and zero-bar trades count in holding periods (#1901, #1876, thanks @Bortlesboat and @MetaAviator). Calmar falls back to 0.0 when an overflowed annual return makes it non-finite, like the other ratios; annual_return itself is unchanged. Completed same-bar trades now count in the average holding period. avg_holding_days remains the engine's bar-based measure, not calendar days.

  • Options legs are priced on the run's bar cadence (#1753, thanks @remote-controlled-man). Historical volatility was always annualised at 252 bars a year, so a crypto or hourly options run priced its legs on one cadence and reported metrics on another. Pricing now receives the run's factor. A cross-market run estimates each bar's cadence from the intervals observed up to that bar, so extending a run cannot change earlier option prices.

  • trades.csv reports the opening fill and keeps break-even exits (#1779, #1780, thanks @MetaAviator and @alexwan02). Each entry and exit row carries an event column. The entry row shows the price and size of the fill that opened the position, not the weighted cost after later scale-ins; those stay in fills.jsonl. Validation and run-artifact readers pick exits by the event column, so zero-PnL exits are no longer dropped. Unmarked older files keep the previous convention.

  • Benchmarks follow the symbol's exchange and the run's source (#1712, #1781, thanks @rudycelekli and @alexwan02). .SH, .SS, .SZ and .BJ symbols take the CSI 300 benchmark whatever the source; local, BaoStock and yfinance runs could fall through to SPY. A source that forbids network fallback, local included, fetches its benchmark through that same loader. When the declared benchmark cannot be resolved, the run card says that benchmark-relative metrics were graded against the equal-weight basket.

  • Crypto spellings get crypto rules (#1740, #1748, #1861, #1862, #1864, closes #1751 and #1863, thanks @he-yufeng, @rudycelekli and @Jackzigen). BTC/USD, ETH/USD and four other known slash-USD pairs are classified before the forex rule, in backtests, correlation and Agent/MCP data routing, while XAU/USD stays forex. Digit-leading bases (1INCH-USDT, 1000PEPE-USDT), separated USDC and BUSD quotes (BTC-USDC, ETH/USDC) and BASE-USDT-PERP perpetuals no longer fall through to A-share rules, where board-lot rounding rejected every order. Signal alignment uses the same classification. The same CI repair reads the alpha-readiness Tushare token through the config layer and treats a whitespace-only token as unset.

  • A-share price limits cover every board's codes (#1913, thanks @rudycelekli). ChiNext 301xxx and STAR 689xxx codes get the ±20% band, and Beijing 920xxx and legacy 4xxxxx / 8xxxxx codes get ±30%; they had fallen back to ±10% and blocked legal fills. ST status and listing-day exemptions still cannot be inferred from a code.

  • Contract specs for Yahoo forex pairs and ICE softs (#1930, #1944, thanks @wavebyrd). EURUSD=X-style symbols, and .FX in any case, are normalised before the spread and swap tables are read, so they stop paying the default 2-pip spread and swap. Coffee, sugar and cotton multipliers are dollars per cent (375, 1120, 500), like the grains; they were full contract pounds, a hundred times too large.

  • Backtests warn about config keys that nothing reads (#1935, #1947, thanks @he-yufeng and @rudycelekli). A key no reader on the backtest path recognises (initial_capital instead of initial_cash) adds a run-card warning with a did-you-mean hint. A key that only another engine reads, such as commission_rate on a crypto run that uses maker_rate and taker_rate, is named as unused by the selected engine, including the sub-engines of a mixed-market run. Runs still proceed. This resolves #1934.

Validation and run artifacts

  • Monte Carlo path metrics start from initial capital (#1664, thanks @MetaAviator). Actual and shuffled paths both begin at the starting capital, so the first trade's return and loss enter Sharpe and the drawdown high-water mark. Monte Carlo statistics and p-values change accordingly; the chart payload is unchanged.

  • Validation stops inventing returns across missing equity (#1754, thanks @MetaAviator). Bootstrap and walk-forward returns use only observed adjacent equity values, so a gap no longer becomes a flat return plus one jump. A walk-forward window without observed endpoints or any adjacent return reports an error instead of a zero score.

  • Standalone revalidation reproduces the original run's settings (#1815, #1915, #1842, thanks @MetaAviator and @rudycelekli). python -m backtest.validation <run_dir> uses the saved Monte Carlo, bootstrap and walk-forward parameters and seeds instead of defaults, still running all three tools. The runner records its annualisation factor in run_card.json and revalidation reuses it, including null for calendar annualisation; older cards reconstruct it from the config and equity timestamps. Reloaded trades recover their entry dates from marked entry/exit pairs, so walk-forward windows group them as the original run did.

Quant Library and analysis tools

  • Brinson attribution reads the book, not its liquidation (#1782, thanks @alexwan02). A run closes every position on its last bar, so the final positions.csv row is all zeros. When it is, attribution uses the row before it and the report notes that row's timestamp. It steps back only one row: a book already flat before the close is still reported as flat.

  • Attribution annualises with the run's recorded rate (#1948, thanks @rudycelekli). When the run card records annualisation.bars_per_year, attribution uses it instead of re-deriving the rate from the requested interval, which disagreed for crypto's 365-day year and any run whose served bar spacing differed from the request.

  • Attribution inputs and Carino linking (#1659, #1943, thanks @rudycelekli). brinson_fachler and carino_factor reject non-finite weights, returns and tolerances by name instead of returning NaN effects. Carino's near-coincidence branch now compares the two returns relative to the remaining wealth. Two returns close to -100% can differ by a tiny absolute amount while their wealth relatives differ severalfold, and those used to take the limit branch.

  • VaR backtests no longer join observations across a gap (#1658, thanks @rudycelekli). A pair dropped for a non-finite return or forecast still leaves the coverage counts, and now also breaks the Christoffersen independence transitions. The breaches either side of it are no longer counted as consecutive days.

  • Implied volatility meets its price tolerance (#1657, thanks @rudycelekli). The bracketing fallback passed tol, a price tolerance, to brentq as a volatility tolerance. It now solves tightly and returns NaN unless the repriced option is within tol of the market price, the same check the Newton path applies.

  • Turnover-aware optimiser and min-weight floor (#1652, #1653, #1885, thanks @rudycelekli). Turnover is measured on signed holdings: reversing a 50% long into a 50% short counts 0.5, both the close and the reopen. Allocations kept without optimisation (all cash, short history, a failed fit) now update the prior holdings the next penalty is measured against. A single-asset book is checked against its per-name and group caps and raises when the allocation exceeds them or is not finite. When a min_weight floor cannot be funded by the names above it, the date falls back to equal weights over the active names, keeping their signs and gross exposure, instead of pushing a funding name through zero.

  • Risk X-ray returns and downside deviation (#1717, #1875, thanks @rudycelekli). Returns are computed per symbol before incomplete dates are dropped, so a missing close no longer becomes a multi-bar move booked as one period. Downside deviation is the zero-target root mean square over all bars, as the backtest's Sortino computes it. The standard deviation of the losing bars alone read zero for a book that lost 1% on every losing bar.

  • Missing observations stay missing in time-series and validation helpers (#1713, #1714, #1715, thanks @rudycelekli). fit_ornstein_uhlenbeck builds lag pairs before dropping missing values, so a gap is no longer fitted as a one-step transition. purged_walk_forward_splits skips a fold whose training set purging has emptied, as its contract states. Regime edge density stays unknown on a bar where any pair's correlation is undefined (a constant or missing series) instead of counting that pair as no edge.

  • Grouped purged K-fold purges groups whose labels reach the test block (#1700, thanks @Shizoqua). group_purged_kfold_splits takes label_end_times. An earlier group whose labels resolve inside the test block is purged, as is a later group inside the test labels' horizon, and the embargo starts after the furthest test label ends. The finishing commit made the argument keyword-only, so existing positional calls keep their meaning. It also requires contiguous, chronological groups when the argument is given. Without it, only the embargo applies, as before.

  • Event-study fitting stays before the event (#1920, thanks @rudycelekli). For an event window that opens after the event, for example (1, 5), the estimation window used to end at the window's start and so took in the event day. With estimation_gap=0 it also took in the days in between. It now ends before both the event and the window, for every normal-return model.

  • Credit, stochastic-volatility and factor-model inputs (#1716, #1718, #1719, #1928, thanks @rudycelekli). cds_price keeps premium dates on the requested frequency and prices a final stub for a fractional maturity. Before, it spread the periods evenly over the tenor. portfolio_style_exposure drops an asset with any missing factor exposure and reports its weight as unmatched_weight, instead of summing NaN into a zero exposure. heston_price rejects non-finite parameters and a nonpositive integration limit before integrating. factor_risk_decomposition rejects a specific variance below -1e-8 instead of clipping it to zero idiosyncratic risk.

  • Non-finite bond and Sharpe-summary inputs are refused by name (#1824, #1821, #1822, thanks @he-yufeng and @rudycelekli). An infinite coupon frequency used to price any bullet bond near par, and an on-coupon-date shortcut in accrued_interest turned a NaN face into 0.0. ytm_solve checks its price and bracket, key_rate_duration its tenors and dv01 its par amount. probabilistic_sharpe_ratio checks all five of its inputs; a NaN benchmark used to come back as NaN "confidence". expected_maximum_sharpe refuses NaN or infinite trial counts. #1824 supersedes the narrower reports in #1821 and #1822.

  • Fund math and cash-flow ingestion refuse ambiguous inputs (#1732, #1739, #1919, thanks @he-yufeng, @Shizoqua and @rudycelekli). american_waterfall refuses a bare as_of date, which it used to ignore silently; per-deal dates are a mapping. A NaN preferred rate used to hand every profit dollar to the LP as preferred return, and the result still reconciled. Hurdle and waterfall amounts must now be finite. Cash-flow and entity-panel ingestion refuse a single dot followed by exactly three digits (12.000 is 12 in a US export and 12000 in a European one) unless decimal_separator is given. Scientific notation such as 1.2e5 is still accepted.

  • Drawdown duration and three-statement tolerance (#1784, #1786, thanks @rudycelekli and @remote-controlled-man). max_drawdown_analysis dates the peak at the latest visit to the high-water mark before the trough, so revisiting a peak starts a new underwater period. project_three_statement applies the caller's balance_tolerance inside each projected period too; a widened tolerance used to fail there at the module default.

  • Portfolio-weight and copula numerics (#1694, #1695, thanks @rudycelekli). Inverse-variance weights no longer overflow on very small positive variances. Inverse-variance weights and HRP reject a non-finite diagonal, and HRP now checks a single-asset matrix as well. A Gaussian copula fitted from a Kendall's tau that rounds to correlation ±1 is nudged back inside the open interval, and its CDF accepts the near-singular covariance.

  • quantlib_call, VPIN and options-payoff results (#1826, #1836, #1897, thanks @rudycelekli). The 5,000-value result budget is enforced while encoding. Before, a long series past the budget still emitted one <truncated> string per remaining element. VPIN completes a bucket at a relative tolerance, so scaling volume units no longer changes which buckets count as full. options_payoff returns an error envelope when extreme inputs overflow to infinity instead of raising.

  • Report audit keeps the sign of a value (#1663, #1654, thanks @he-yufeng and @rudycelekli). Accounting parentheses, ASCII or full-width, are negative: (25.3)亿 is read as -25.3 亿 with its unit, where it used to read as +25.3 with no unit. -25M, −25M, -$25M and $-25M keep their minus sign, and (-25) is still a negative. In prose a bracketed number without a unit or currency, such as (2024), is skipped as a year. Mismatched brackets, a unit both inside and outside, and non-finite or larger-than-1e15 values are rejected. Reported in #1660.

  • Financial-rigor checks stop reading undefined as agreement (#1883, #1906, #1921, thanks @rudycelekli). A zero reported market cap matches only a zero product, and against a zero cross-source median only an exact zero agrees. Either way the percent deviation is null. A negative median is compared on its magnitude. Before, these read as a 0% deviation. An explicit tolerance_pct=0 or years=0 is honoured instead of falling back to 2% and 3 years. Three-scenario upside is null at a nonpositive price, where it used to report 0% ("fairly valued"), and negative years is an error.

  • Swing-based skill signals no longer fire before the swing is known (#1873, thanks @rudycelekli). The bundled Elliott-wave and harmonic example engines stamped a signal on the swing bar itself, which is only confirmed swing_window bars later. They now replay the swings in order and stamp each signal on its confirmation bar. The optional pyharmonics path is unchanged.

  • Pattern tool extrema and arguments (#1882, #1923, thanks @rudycelekli). A flat top or bottom spanning adjacent bars is one extremum, reported at its first bar. A missing run_dir, a list of patterns, a null or non-integer window now get a JSON error or their default instead of raising out of the tool.

  • Shadow-account RSI uses Wilder's seeded smoothing (#1457, thanks @Shizoqua). The first average is the simple mean of the first 14 gains or losses, and the recursion runs from there. The extractor and the scanner share one implementation, so the RSI bounds a profile extracts are the ones the scanner tests.

  • Small-price indicators keep their digits (#1925, thanks @rudycelekli). MACD and Bollinger values were rounded to 4 and 2 decimals, so a 1.2e-5 crypto price came back as zeros. Each nonzero value now keeps at least four significant digits, including a small component next to a larger one; ordinary prices round as before.

  • Hypothesis search and backtest links (#1720, #1902, thanks @Shizoqua). A search with limit=0 returns nothing instead of one record, and the link_backtest tool forwards its optional validation results (walk-forward, Monte Carlo, bootstrap) to the registry instead of dropping them.

Alpha Zoo, factors and alpha bench

  • Fundamental factors no longer see after-close filings a bar early (#1794, thanks @Jackzigen). SEC filings carry only a date and can arrive after the close, so the benchmark IC path now uses fundamental panels one bar later and records fundamental_pit_lag_bars: 1. IC measured on same-day filings needs to be rerun. The backtest engine's separate next-bar execution is unchanged.

  • Carhart momentum no longer reads today's price (#1578, thanks @Shizoqua). academic_carhart_mom was the 12-month return minus the 1-month return, which moves with today's close, the opposite of skipping the last month. It is now the return from 252 to 21 trading days ago. A test moves every close inside the skipped month and requires the value to stay put. The finishing commit corrected the factor notes, the zoo licence text and a strategy example that taught the old formula under a factor id that does not exist.

  • decay_linear weights the newest bar heaviest (#1912, thanks @rudycelekli). Inside each window the weights ran from n on the oldest bar down to 1 on today, the reverse of the Alpha101 definition. Every factor that calls it changes value (22 Alpha101 and 19 GTJA 191 modules); windows stay causal and the missing-data masks are unchanged.

  • Rolling argmax/argmin ties no longer depend on the backend (#1898, #1696, thanks @rudycelekli and @Mikesad). Without bottleneck, a tie resolved to the oldest bar; it now resolves to the most recent, as bottleneck does. bottleneck is a core dependency, so most installs already got the most recent bar. The registry test suite also runs on Windows hosts without symlink privileges.

  • Alpha benchmark arguments and environment (#1806, #1910, #1918, thanks @xaritoku and @rudycelekli). alpha bench and alpha compare load the .env file before reading credentials, as other CLI commands do. top=0 is rejected instead of quietly becoming 20, and boolean, fractional or non-numeric top values return an error envelope. A factor with fewer than two IC observations is skipped with the reason instead of being ranked with a NaN IC standard deviation and an IR of 0.

  • 38 GTJA 191 factors show their formula. Their formula_latex read "see body", which the Alpha Library displayed as the formula. It now holds the formula already quoted in each module's docstring. Metadata only; no computation changed.

Live trading, brokers and portfolio

  • Live option orders are refused before reaching a broker (#1585, thanks @lorenzozanee). Robinhood's place_option_order is pinned as a write, so a remote read-only annotation cannot pass it through, and the profile says live order placement is equities-only. The mandate gate also refuses every option intent, on both the MCP and SDK paths and even when the mandate lists options, because share-notional limits cannot bound a contract multiplier or maximum loss. Live options execution is not implemented.

  • Direct-SDK positions are priced for the mandate exposure check (#1581, thanks @Shizoqua). Toss, Dhan, KIS, Zerodha and Shoonya report a position's price as current_price or ltp, and Futu its value as market_val; none was read, and one unpriced position fails the exposure check closed, so every later order on those brokers was refused. A zero or negative price, a boolean, or a non-finite value or total now fails closed as well.

  • The kill-switch sweep survives its own failures, and stopping a runner stops it (#1703, #1704, thanks @Romzhou). An audit-ledger write that raised used to abort the preemptive cancel/flatten sweep and latch it as fired, so it never ran again; it is now logged and the sweep continues. A resting order without an order id, or a position with a missing, malformed, non-finite, overflowing or boolean quantity or no symbol, is reported as an error instead of being cancelled blindly or closed for NaN shares. POST /live/runner/stop used to answer was_running: false while the scheduler kept firing ticks, because the driver task returned as soon as the scheduler started; it now lives as long as the scheduler, and stopping it cancels an in-flight agent decision and waits for it to end. Each broker keeps its own runtime job store, migrated from the shared one.

  • US equity live sessions close at 13:00 ET on half-days (#1706, closes #1705, thanks @he-yufeng): 27 November and 24 December 2026 and 26 November

    1. The calendar is still a static two-year table.
  • KIS reads see SOR and NXT fills and recover from a replaced token (#1726, thanks @passionbull). Order and fill reads asked only the KRX venue, while most retail orders route through SOR, so an account that had traded read as empty. They now query KRX, SOR and NXT and merge the rows, and a venue that fails or a page without a row list fails the read instead of returning a partial one. KIS issues one token per app key, so another process can silently invalidate the cached one; an explicit token-expiry rejection (HTTP error, 401/403 or an HTTP 200 business error) now drops the cache and retries once. Order placement still goes to KRX only.

  • Binance balances survive clock skew, and unpriced assets mark the snapshot incomplete (#1505, thanks @goatyyc). A balance or Simple Earn read rejected as "ahead of the server" (-1021) refreshes the clock offset and retries once. Quotes, bars and symbol search use an unauthenticated client, so a public ticker no longer needs a signed request. Binance's account total is built from priced positions, so a Binance asset without a price now marks the snapshot incomplete and names the asset, and "last complete" reads and history skip such snapshots.

  • Futu reads a saved non-HK market and reports both cost bases (#1812, #1814, closes #1807, thanks @Frank-zhu2000 and @alexwan02). All four Futu profiles pinned filter_trdmarket to HK, hiding a US account saved in futu.json; HK is now only the default when nothing is saved. A multi-currency account's balances, which Futu labels N/A, carry the HKD they were requested in. Positions add average_cost, pl_ratio_avg_cost, unrealized_pl and realized_pl (None where Futu reports N/A), and the portfolio derives Futu unrealized P&L from the average cost, because Futu's cost_price is diluted by realized gains and can turn negative.

  • Futu HK.00700-style codes reach the risk x-ray and holdings merge (#1922, thanks @rudycelekli). Market-prefixed codes are rewritten to the suffix form the loaders use (00700.HK, BRK.B.US) before risk inputs are built and holdings are merged across brokers. Snapshots valued before this change are hidden from latest and history views until the next refresh rebuilds them.

  • The Portfolio page's holdings and history follow the display currency (#1810, thanks @rwxyx518-prog). The holdings table's value and P&L columns and the history chart used USD whatever currency was selected. Converted values use the snapshot's own rate and show a dash when that rate is missing, rather than an assumed one.

  • Weekly and monthly risk x-rays annualise per period (#1656, thanks @rudycelekli). portfolio_risk_xray annualised 1W and 1M volatility with 252 bars a year; it uses 52 and 12, and 1M is not read as a minute. A monthly request without a start date fetches 31 months, enough for the minimum sample.

Trade journal

  • Mixed-currency journals are reported per currency (#1811, closes #1808, thanks @Jackzigen). total_pnl, max_drawdown and dividends summed HKD and USD round-trips without conversion. With more than one settlement currency, total_pnl, total_dividends, max_drawdown and profit_loss_ratio become currency-keyed objects with a warning, the overtrading diagnosis is given per currency, and top_symbols ranks by trade count with each amount's currency; no FX conversion is applied. Single-currency journals keep scalar values. Drawdown now starts from zero realized P&L, so a losing first close counts.

  • Amounts and symbols parse the way the broker wrote them (#1766, #1871, #1872, #1881, #1916, thanks @remote-controlled-man and @rudycelekli). A generic Date,Symbol,Side journal is claimed as a Futu export only when Futu's Quantity or Amount column exists, so its quantities and prices are no longer read as zero. Accounting negatives such as (1,234.50) and $(1,234.50) keep their sign instead of becoming 0. 1.234,56 reads as 1234.56 (a comma-only 1,234 is still thousands). Fiat pairs such as EUR/USD are no longer classified as crypto, .US symbols, share classes like BF-B.US included, are US equities, and .SS is Shanghai.

  • Dates are combined, ordered and validated before analysis (#1900, #1941, #1942, thanks @MetaAviator). Separate generic date and time columns are combined, including Excel serial dates and day fractions, and a time alone never stands in for a missing date. Rows with equal timestamps keep their file order through FIFO matching. A journal with a missing or unreadable date returns an error instead of being analysed or filtered without those rows.

  • FIFO P&L counts reversal fees once, and unbounded ratios are valid JSON (#1891, #1884, thanks @MetaAviator and @rudycelekli). A fill that closed one side and opened the other charged its whole fee to both; the fee is now split by quantity. A journal with no losing round-trip produced an Infinity profit/loss ratio that strict JSON parsers reject; it is null. When winners and losers both have zero holding time, the disposition check reports an undefined ratio instead of a bias.

  • Unknown analysis types and filters are rejected (#1911, #1924, thanks @rudycelekli). An unrecognised analysis_type returned an empty ok result, and an unrecognised filter returned the whole journal as if it had been applied; both now return an error naming the accepted values (full, profile, behavior, strategy; a date range, symbol= or market=).

Strategy store

  • Strategy records keep their trading rules, and declared backtests feed decay monitoring (#1848, #1849, thanks @alexwan02). sdm_register accepts entry_rules, exit_rules and position_sizing (a list is stored one rule per line) instead of dropping them. Nothing used to write bench_history, so sdm_status(decay_check) and sdm_decay_scan always reported insufficient data: a backtest whose config.json names an artifact_id now records its Sharpe, annual return, max drawdown and Calmar and reports a bench_record (recorded, skipped or failed, with a reason). Decay monitoring reads only initial and periodic rows, so an undeclared rerun (manual) cannot move the baseline. Failed runs, runs without trades or metrics, and a measurement already stored, even from a copied directory, are skipped, atomically across connections; a different evaluation window is a new measurement. Bookkeeping errors never fail a successful backtest.

  • SDM decay compares disjoint IC windows (#1699, thanks @Shizoqua). The baseline and recent windows are the oldest and newest five IC observations; with fewer than ten they overlap (and with five or fewer they are the same window), which pinned ic_ratio near 1.0 however far the IC moved. ic_ratio is now null until ten observations exist.

Tools and documents

  • A refused file write names the setting that allows the path (#1679, thanks @he-yufeng). When write_file or edit_file refused a path outside the allowed roots, the error said only that the path escaped. Now it lists the allowed roots and names VIBE_TRADING_ALLOWED_WRITE_ROOTS; read-side refusals name VIBE_TRADING_ALLOWED_FILE_ROOTS. The variable is also documented in .env.example and the README. The boundary itself is unchanged. Since aac03fa6, on a case-insensitive filesystem a path spelled differently from an allowed root (dropbox for Dropbox) is accepted when the filesystem confirms it is the same directory. Path names are never case-folded.

  • Skill file removal follows the same rules as skill file writes (#1639, thanks @Shizoqua). remove could delete any file in a skill directory. It is now limited to references/, templates/, examples/ and assets/, like write. Since 8ec3b80f, both actions also refuse .. segments and symlinks that resolve outside those directories, and remove deletes only regular files, never a directory.

  • read_file reports an unreadable directory as a structured error (#1946, thanks @rudycelekli). Listing a directory inside the allowed roots ran outside the tool's error handler. A permission error therefore escaped instead of returning the usual JSON status: error.

  • Document reader keeps more of PDF, Word, PowerPoint and Excel files (#1838, #1839, #1840, #1841, thanks @rudycelekli).

    • PDF: with no OCR engine, a page whose native text is shorter than min_text_per_page (50) used to be skipped. Its text is now kept, and the result is marked no_ocr_engine. If every selected page has only such short text, the read is still refused with the OCR install hint.
    • Word: paragraphs and tables come out in document order. Before, every table was appended after all paragraphs.
    • PowerPoint: text inside grouped shapes and tables is read; merged table cells appear once.
    • Excel: cells holding NA, N/A, null or NaN keep that literal text. Only empty cells count as missing.
  • README tables match the code (#1671, #1733, thanks @he-yufeng and @Jackzigen). The Indonesian README lists get_southbound_flow and the current tool counts. The Quant Library contract table in README.md had duplicate rows for credit, timeseries, factormodel and multipletesting; each now has one row with the newer description.

Sessions, memory and search

  • Runtime files stay out of a checkout used as the runtime root. When VIBE_TRADING_HOME points at the repository, memory_index.db (moved to the runtime root in this release, and holding memory contents), the paper trading strategies/ folder and shadow_reports/ were not git-ignored. They are now, and the guard test also finds names joined onto the runtime root inline or through *_DIRNAME constants, not only *_FILENAME ones.

  • Session search finds Chinese, Japanese and Korean text (#1686, thanks @Shizoqua). SQLite's default tokenizer treated a run of East Asian characters as one token, so a query rarely matched anything shorter than the entire run. The ported fix indexes those characters split apart. aac03fa6 reworked it: ideographs, kana and Hangul are indexed one character per token in a separate search_content column; queries become phrase matches; and the stored message keeps its original text for snippets. An existing sessions.db is upgraded and reindexed once, on first open.

  • Session search snippets, limits and concurrency (#1834, #1827, #1828, thanks @rudycelekli). A snippet now centres on the match SQLite actually found and highlights that span in the original text. Before, it highlighted the query wherever the text contained it, so alpha lit up inside alphabet while café (matched by cafe) was not highlighted. The result limit now counts distinct conversations. Before, the search read only the first five matching messages per requested conversation, and one chatty conversation could fill them all and hide the others. Index writes and searches on the shared SQLite connection are serialized, so concurrent callers no longer interleave on one cursor.

  • Renamed conversations show their new name in search (#1832, thanks @rudycelekli). The REST rename and auto-title routes now update the search index's title as well. Auto-titling reloads the session after the model answers, so updates saved meanwhile are kept, and a manual rename made meanwhile is kept with status kept.

  • Reconnecting to a session's event stream no longer repeats an event (#1798, closes #1796, thanks @Shizoqua). An event published between the replay snapshot and the new subscription reached the client twice, once in the replay and once live. Both steps now happen under one lock.

  • Memory search snippets and titles read back as written (#1682, thanks @Shizoqua). A CJK snippet could show its span twice ("使用用移移动…") because the index stores both single characters and pairs. A title mixing Chinese with a ticker, number or punctuation came back with extra spaces (比特币A策略 became 比特币 A 策略). Since 8ec3b80f, highlights and ... truncation markers also survive in snippets built only from character pairs.

  • Memory removal, links and recall (#1697, #1771, #1772, thanks @Shizoqua). remove(), which the remember tool's forget action uses, accepts a filename stem as find() already did. Before, it reported "not found" for an entry that find() and vibe-trading memory forget could see. A blank name is refused (f6e72428). Semantic links match two-character identifiers such as BP or GE; the minimum used to be three. FTS recall skips index rows whose memory file was deleted outside the process. The PR fetched four times the requested count; 54e225de replaced that with a window that doubles until enough current memories are found or the index runs out.

  • Verifying an archived audit ledger reports a corrupt line instead of crashing (#1721, thanks @Shizoqua). verify_chain_with_archives parsed every segment with a bare json.loads, so a damaged line raised JSONDecodeError. It now uses the same parser as single-file verification and returns a malformed_json break at that line's position across all segments.

Credential redaction

  • Environment-style credential names and URL credentials are redacted (#1765, thanks @cgycorey). TUSHARE_TOKEN=…, AWS_SECRET_ACCESS_KEY=… and aws_secret_access_key lines in ~/.aws/credentials are now caught. Lowercase identifiers such as pad_token and max_token are still left alone. scheme://user:password@host becomes scheme://[redacted]@host, and a password containing @ is handled. A bare username before @ is redacted only when it is a 20-plus-character token, so [email protected] stays readable.

  • Authorization: Basic and other non-bearer schemes are redacted (#1846, thanks @cgycorey). The key pattern used to redact the scheme word and leave the base64 user:password visible. Basic, Digest, Token, Negotiate and NTLM values are now redacted, but only right after an authorization label, so prose like "basic information" is untouched.

  • More secret formats are redacted (#1874, #1866, #1867, #1868, #1869, #1870, thanks @rudycelekli). #1874 combined the five smaller PRs. These are now redacted:

    • unlabelled GitHub fine-grained (github_pat_), GitLab (glpat-), npm, Hugging Face, Stripe (sk_/rk_ live and test), Google API (AIza) and AWS ASIA keys;
    • PEM and PGP private-key blocks, including a block cut off before its END line, which is redacted to the end of the text;
    • Cookie: and Set-Cookie: header values;
    • fields named like cookie or private_key in structured payloads.

    Public keys and certificates are left as they are, and running redaction twice changes nothing. Since 47de62de, a quoted header such as curl -H "Cookie: a=\"x\"; b=y" is redacted up to its closing quote.

  • Telegram bot tokens, Slack and Discord webhook URLs, Groq and NVIDIA keys are redacted (#1880, thanks @rudycelekli). The Telegram token is caught inside https://api.telegram.org/bot<token>/…, which HTTP error messages repeat in full. For a webhook URL, only the secret last path segment is replaced. The PR accepted only 8–10 digit Telegram bot ids and lowercase hooks.slack.com. 2c00cd20 widened that to any length of numeric id, to GovSlack (hooks.slack-gov.com) and to webhook hostnames in any letter case.

CLI, API and scheduling

  • vibe-trading update leaves a Git install alone (#1655, thanks @rudycelekli). An install made from a Git URL was treated as a PyPI wheel, so the update replaced the user's chosen revision with a released wheel. The command now detects vcs_info in the install metadata, reports the newer PyPI version and tells the user to update from the original source.

  • The terminal dashboard matches parallel tool results to the right call (#1859, closes #1856, thanks @Shizoqua). When several calls to the same tool ran at once, progress and results were matched by tool name and recency, so results could be swapped between calls. They are now matched by call id. Since 47de62de, a result whose id matches no step gets its own step and no longer completes a different call that is still running.

  • CLI limits and number formatting (#1886, #1926, thanks @rudycelekli). vibe-trading list --limit and memory search --limit reject zero and negative values instead of passing them on. Token counts, abbreviated numbers and durations switch unit when rounding reaches the next one: 999,950 tokens shows as 1.0M tokens instead of 1000.0k tokens, 999.6 ms as 1.0s, and 59.96 s as 1m 00s instead of 60.0s.

  • A cancelled upload leaves no partial file (#1945, thanks @rudycelekli). If the upload request was cancelled while a chunk was being read, the half-written file stayed in the uploads directory with no reference to it. The file is now deleted and the cancellation is re-raised.

  • More runtime files follow VIBE_TRADING_HOME (#1896, #1904, #1887, thanks @rudycelekli). These paths were fixed to ~/.vibe-trading and now use the runtime root:

    • persistent memory and its search index (memory/, memory_index.db);
    • autopilot run directories;
    • the alpha-bench panel cache and reports;
    • the Gildata code cache.

    The Shadow Account report route also reads from the runtime root, which is where the report writer saves. Report links therefore work again when VIBE_TRADING_HOME is set. Existing files under ~/.vibe-trading are not moved, and the memory paths are read when the process starts.

Channels and scheduled delivery

  • IM channels show agent progress again, without repeated lines (#1878, thanks @alexwan02). Channel replies now forward the attempt's status lines (working, waiting for the model, using a tool, checking the figures), tool hints and opt-in reasoning while the agent runs, so Feishu's streaming card and other streaming channels update again (reported in #1792). Only those labels go out: answer drafts and tool results stay private, and the final answer still passes grounding first. #1878 applies the channel's own visibility settings before deduplicating, so hidden tool hints no longer break the dedupe and a chat no longer stacks one "Waiting for the model" per tool round.

  • Streamed replies are no longer lost or split (#1860, #1903, #1929, closes #1857, thanks @Shizoqua). Telegram (#1860) and Matrix (#1903) dropped the whole answer when the live preview message could never be created; they now send the accumulated reply as a new message. CI installs the matrix extra so the Matrix test runs instead of skipping. WeCom (#1929) generated a new stream id for every chunk, so progress and the final answer arrived as unrelated messages; one id now covers each reply.

  • Email sender authentication cannot be satisfied by forged headers (#1508, thanks @Shizoqua). SPF and DKIM checks read only the single Authentication-Results header bearing the configured trusted_authserv_id; with none configured, or with two matching headers, verification fails closed. A pass counts only when its domain equals the From domain or an aligned dmarc=pass is in the same header, and comments or quoted text cannot pose as results. The receiving provider must strip forged copies bearing its id; the id is not a signature. The guided Email setup exposes the setting, and a blank one rejects inbound mail while SPF or DKIM checks are on. The PR described trusting the topmost header; what shipped selects by configured id instead.

  • URL credentials in channel settings are neither leaked nor erased (#1742, #1744, thanks @cgycorey). Status and config redaction strips user:password@ even when the port is malformed or out of range or the URL is followed by punctuation, where it used to return the URL untouched (#1741). Saving or testing a form that echoes the redacted URL keeps the stored credential instead of overwriting it (#1743); a different URL or a clear flag still applies. Connection-test and reset messages, including raised SDK exceptions, are scrubbed too.

  • Unchanged empty fields no longer block channel saves (#1681, thanks @zeus229). A nullable setting such as Email's post_action was submitted by the form as an empty string and rejected with a 422 on unrelated edits. An empty value now maps back to null where the adapter's default is null, for both Save and Test.

Web UI and locales

  • Run trades list every trade, newest first (#1865, #1877, #1831, thanks @alexwan02 and @he-yufeng). The Trades tab and its CSV download used the first 500 rows of trade_log, which dropped the newest trades of a long run; they now use the full trades.csv rows (#1865). Rows sort newest first with a header toggle, keeping same-time fills in engine order, and the symbol filter lists codes in order (#1877). #1831 removes a race in the Settings effort-level test that intermittently failed frontend CI; the component is unchanged.

[0.1.16] — 2026-09-29

Rolls up 492 commits / 116 merged pull requests since 0.1.15, from 16 contributors.

The theme of this cycle is a number that can show where it came from. The final-answer grounding gate stopped guessing what a number is from the words around it: the model declares each figure's role — observed, derived, proposed, cited or count — the gate checks that declaration against the session's tool evidence, and a figure that fails is cut from the answer instead of the whole answer being refused. A backtest's own output (its Sortino, turnover, weights and Monte Carlo p-values, not only its Sharpe) now grounds the report written about it, run cards cite a metric only where the saved CSV agrees, and the generated broker matrix states each profile's permissions as declarations rather than as runtime verification. Research runs stopped losing what they fetched: compaction follows the model's real context window instead of a fixed 40K estimate that began clearing tool results from the third step. The missing-data sweep reached 59 more alphas: 15 qlib158 count factors, 20 that filled a missing input with a constant and 24 comparison gates. Argentina (BYMA), weekly and monthly bars, Bahasa Indonesia, Gildata, OpenCode, Email and WebSocket channels and four broker connectors (KIS, Upbit, Toss Securities, Scalable Capital) land alongside.

Added

  • Stock profiles carry the listing's own identity (#1577, thanks @Shizoqua). get_stock_profile returns a listing block beside the issuer's fundamentals: Yahoo's symbol, exchange, quote type and quote currency, plus the financial currency and underlying symbol when Yahoo declares them. A secondary listing no longer reads as the primary one. GOOGL.BA is an ARS line in Buenos Aires carrying Alphabet's USD financials, and GGALD.BA is a USD line whose issuer reports in ARS. Nothing is inferred from the ticker.

  • Volume statistics in technical_indicators (#1571, thanks @zeus229). The tool returns the latest volume, its 20-bar mean and their ratio, computed from the bars it already fetched. The mean needs a complete 20-bar window. volume.unit is the unit the serving source declares (board lots on the A-share sources, shares on the Yahoo family, null when undeclared), because the two are 100x apart and nothing in the numbers tells them apart.

  • Argentina (BYMA) market data (#1543, thanks @zeus229). A .BA symbol — a BYMA listing or a locally traded CEDEAR — is its own market ar_equity, quoted in ARS, served by yahoo → yfinance → local, and reported as market ar by get_stock_profile. Execution is deliberately not modelled: an Argentine backtest raises instead of borrowing US or crypto commissions, lot sizes, settlement and short-selling rules. The Web UI carries the market too — Settings source priority, the positions asset-class grouping and the sector union — and with it uk_equity, vietnam_equity and index, which had been missing from one or more of those three since they landed. EnvConfig.data's per-market MARKET_DATA_ORDER_* fields are now guarded against FALLBACK_CHAINS instead of a hand-written list, which is why uk_equity never had one.

  • Weekly and monthly bars (#1479). get_market_data and backtests take 1W and 1M (and 1w / 1wk / 1mo). Loaders are still asked for daily bars; loaders.base.resample_bars builds the period bar at the three fetch boundaries (runner, market data, benchmark), so a weekly bar is the same whichever source served the days, at that source's daily caliber. A week runs Monday to Sunday, a month is a calendar month, and each bar is dated on the last trading day inside it: first open, highest high, lowest low, last close, summed volume and amount, volume-weighted vwap, mean funding rate. Annualisation is 52 / 12; a weekly perpetual bar settles 21 funding periods. 1M is matched before any case fold, which read it as one minute in strict-100x validation and in attribution (98,280 bars a year). The technical-indicator tool canonicalises its documented 1wk / 1mo and fetches whole periods instead of two days per bar.

  • IM channels are configured from the Web UI (#1520, #1529; #1519, #1530, thanks @shadowinlife). The Settings panel renders each built-in channel's fields from backend metadata, masks secrets (**** plus the last four characters, never sent back), tests the unsaved form against the provider (ok / invalid_credentials / network / unsupported), and hot-swaps only the saved channel with no restart. DingTalk and QQ have guided setups; a shared token probe backs both. Saves write channels.<name> in ~/.vibe-trading/agent.json atomically at mode 0600; a YAML config is shown read-only. Reloads of one channel are serialized in ChannelManager, and stop_all cancels a reload's start that is still connecting.

  • Email and WebSocket join the guided Web UI channel setup (#1544, after #1519, thanks @shadowinlife). Both channels get hand-written field metadata — localized labels, masked secrets — plus connection tests and per-channel hot apply. Email's test probes the real IMAP login, mailbox select and SMTP login without ever sending a message; WebSocket is a server channel with no remote credentials, so its test validates the TLS cert/key material and binds every resolved address locally — an address held by the already-running server reads as the expected state, not a failure. Saving a WebSocket config hot-swaps the server: connected clients (including the Web UI chat) briefly disconnect and reconnect. A hand-written field hint's secret flag is now authoritative for the keys it covers — the regex fail-safe still masks every unhinted key — so token-shaped non-secrets such as websocket_requires_token stay visible and editable instead of being masked, and a form save can no longer silently flip the stored token requirement. The web config routes now build the WebSocket adapter with the gateway services it requires; previously every WebSocket save or test failed validation. Email's polling loop also skips stale-config delete/move post-actions once a stop or hot swap begins, while the already-fetched batch is still delivered. Both the probe and the polling loop now send an RFC 2971 IMAP ID after login: NetEase mailboxes (163/126/yeah.net) accept the login but reject the first SELECT with Unsafe Login until the client identifies itself, so without this the Email channel could not read the most common Chinese mailboxes. The identification is static and carries no user data or secrets, and servers that ignore ID are unaffected; verified end-to-end against a live 163.com account.

  • Bahasa Indonesia UI locale and README_id.md (#1482, thanks @Yoruxyv). The README count tests cover it, and MANIFEST.in now ships README_es.md and README_id.md.

  • Portfolio valuation accepts any ISO-4217 currency (#1510, thanks @zeus229). A source whose currency has no FX rate fails with the currency named and is excluded from the totals; the healthy sources stay in.

  • Gildata (恒生聚源) joins the A-share fallback chain as a token-gated source (#1474, thanks @sambazhu). The new gildata loader talks to the vendor's raw-api MCP endpoint (one JSON-RPC POST per call, token in the Authorization header) and serves A-share daily OHLCV through the StockDailyQuote tool with restorationStatus=1 — forward, split-AND-dividend adjusted bars, stamped split_dividend in the price-caliber table after being measured against a live payload. Volume arrives in 万股 and is converted to shares (volume_units: shares); avgprice/prevcloseprice are ignored because the vendor keeps them on a different adjustment basis than the adjusted OHLC. An unresolvable symbol answers rows: [] (never an error), so the chain keeps walking. Auth is GILDATA_TOKEN (Settings page field or env; GILDATA_BASE_URL overrides the endpoint, VIBE_TRADING_GILDATA_MIN_INTERVAL the 0.3s default spacing). Without a token the loader reports unavailable and the chain skips it — no behavior change for existing users beyond one reorderable entry at the chain's tail. One migration note: a MARKET_DATA_ORDER_A_SHARE value saved before this change is a permutation of the old 7-source chain, so it stops validating once gildata joins — the Settings card flags it and the default order applies until the saved order is re-saved (one click) with the new source included.

  • A Robinhood account can be a read-only portfolio source (#1428). The new robinhood-live-mcp-readonly profile uses the same MCP server and OAuth grant as the trading profile. Its connection reads exactly one account, picked from Robinhood's own get_accounts list in the connection center or with vibe-trading connector select-account <id>. Nothing is preselected, and a source with no account errors before any broker call. Holdings come from get_portfolio and get_equity_positions and are mapped only from the shapes the reporter posted. A second page of positions, a null list or item, or an unknown field is an error, not a shorter account. Positions show quantity and cost but no price until the quote reply is mapped. An account that also holds options, crypto, futures, event contracts, mutual funds or fixed income fails the source instead of showing an equity-only view.

  • A weekly health check of every keyless data source (#1627, closes #1623, thanks @he-yufeng) — python -m backtest.loader_health fetches one liquid symbol from each of the 15 keyless network loaders, each in its own process with a temporary home, no inherited keys and the loader cache off, and requires normalised OHLCV whose last bar is at most 14 days old. The Public loader health workflow runs it every Monday and on demand. An unreachable endpoint, a missing dependency, malformed bars and a stale series all fail the lane; as first proposed, an unreachable endpoint counted as a skip. An offline test fails when a new keyless loader has no canary symbol.

  • OpenCode is a built-in provider, and each request carries the session id OpenCode Go requires (#1416, closes #1415, thanks @tonydo) — LANGCHAIN_PROVIDER=opencode with OPENCODE_API_KEY / OPENCODE_BASE_URL, in Web Settings and both CLI onboarding flows. OpenCode Go refused every call without x-opencode-session (400 MissingSessionID); the adapter now sends the run's session id to any opencode* provider or opencode.ai host, and the compaction summary call, made on its own thread, carries the same id (fd6c7c7c) instead of reaching OpenCode as a second conversation.

  • KIS (Korea Investment & Securities) is a connector for KOSPI and KOSDAQ equities (#1407, thanks @as950118) — paper (모의투자) and live sit on separate hosts with separate OAuth tokens, so the paper-trade profile sends orders to KIS's own paper server rather than a local simulator. Live is read-only: no kis-live-trade profile exists, and place_order / cancel_order refuse a non-paper config before anything else. A finishing commit (970c235f) pins the paper sell and cancel TR_IDs, which no test covered, and makes a balance or order walk that hits its 20-page cap raise instead of returning a short position list that reads as complete.

  • Upbit is a KRW crypto connector, capped at paper plus read-only (#1408, thanks @as950118) — one key pair reads the same account whatever profile is declared, so there is no runtime paper/live discriminator and no live order path. Paper orders are simulated against a live Upbit quote: no quote means no fill, a limit order fills only if immediately marketable, and cancel acknowledges only ids the simulator issued. Needs pip install "vibe-trading-ai[upbit]".

  • Toss Securities is a read-only connector for Korean and US equities (#1409, thanks @as950118) — Toss documents no sandbox and no paper/live discriminator, so, like Trading 212, it refuses every order and cancel, paper included. Reads follow the published schema (/api/v1/holdings, /api/v1/orders); closed-order history is paged to completion, and a walk past its cap, or a hasNext page with no cursor (b5ff36ba), is an error rather than a silently truncated list.

  • Scalable Capital is a read-only MCP connector that declares quotes only (#1417, thanks @cgycorey) — Scalable has no paper account and publishes tool names without argument schemas. scalable-live-mcp-readonly withholds account.read / positions.read because nothing maps the holdings reply yet: declaring them made it a portfolio source whose refresh stored a complete snapshot with zero positions. The 29 published tools are pinned literally as 18 read and 11 write (unknown names resolve to write), and a hand-enabled order tool is wrapped by the live order guard and refused.

  • The README's broker capability matrix is generated from the profile registry (#1629, thanks @he-yufeng) — one row per built-in profile (55 across 18 connectors): environment, transport, read-only or write-enabled, declared reads, and whether placement requires a mandate. A test fails CI when the block and the registry disagree; regenerate with PYTHONPATH=agent python -m src.trading.capability_matrix. As merged it folded each connector into one row; dec2c0e6 renders each profile on its own, so a paper order capability never reads as a live permission, and says the rows are declarations, not runtime verification.

  • Run cards record the backtest's execution and cite metrics against the saved CSV (#1612, closes #1610, thanks @lorenzozanee) — run_card.json (schema 1.0) carries a trace with UTC start and end, status, and SHA-256 hashes of the config and the metrics, never the payloads. As merged, every scalar metric was cited whenever artifacts/metrics.csv existed; after 0244ecea a metric is cited only when its value matches the file's single data row and the file matches its manifest checksum. The Markdown card and the Run Detail page show both sections.

  • A read-only extraETF export reader (#1406, thanks @cgycorey) — the Investments export becomes positions and the Transaktionen export typed movements, so a Dividende row is never read as a holding and a transactions file cannot pass as a portfolio. It fails closed on anything it cannot read without guessing (unknown header or label, US-ordered numbers, more than 8 decimals, one instrument twice under one Portfolio ID). Nothing calls it yet; the portfolio container side is still open under #1170.

  • Strategy-store adaptation substrate (#1391, thanks @modelpath-dev) — a child strategy points at its parent through derived_from (SQLite schema 3, an additive nullable column), and AdaptationPatch can change only universe, name and position sizing, never signal_definition. The child starts with no evidence and UNVALIDATED; a7bd2e30 also clears the parent's validator, approver and versions, and makes register_adaptation require the parent's evidence, refusing a stale or insufficient parent. Nothing outside tests calls it yet (#1149 stays open).

Changed

  • source="qveris" refuses markets with splits and dividends (#1494). It picks a capability by search rank, which ignores adjustment: for 600519.SH the top pick was FMP's non-split-adjusted EOD, and a tool named adjusted returned raw closes. Stocks and ETFs are refused before anything is searched or billed; crypto, forex, futures, macro and index bars are served as before. Both QVeris budget gates reserve only a flat per-call quote now: 1 credits/result had reserved 1 credit for a call that billed 9.66, and qveris_execute returns quote_not_bounded for any other shape.

  • A price index is stamped raw (#1541). Eastmoney's fqt=1 equals fqt=0 on every bar of the A-share indices measured, Tencent serves only day for the SSE/SZSE ones, and Yahoo's adjclose equals close on ^GSPC / ^NDX / ^HSI, but the per-source table stamped an A-share index additive and a Yahoo index dividend-adjusted. price_caliber takes the symbol; A-share index codes are matched per exchange, since 000001.SZ is Ping An Bank.

  • Portfolio valuation version 3 (#1510, thanks @zeus229): snapshots stored under version 2 are no longer read, so portfolio history starts again after upgrading.

  • Tencent, Eastmoney and AKShare A-share prices are labelled split_dividend_additive (#1497, #1493, thanks @cgycorey). Their forward adjustment subtracts cash dividends from the level instead of scaling by a ratio (on 600519.SH, qfq minus raw takes five values over 500 bars against 367 distinct ratios), so a basket mixing them with multiplicative sources warns, and a run served only by them warns that its returns are not total returns. Tencent's HK series is labelled raw: it serves no adjusted HK bars. The prices themselves are unchanged.

  • connector account renders Binance spot balances as Asset / Free / Locked / Total, Futu's per-currency assets, and Trading 212's cash and metadata (#1539); before, Binance printed blank rows and the other two printed "No account summary returned."

  • The final-answer grounding gate no longer guesses what a number is from the words around it. Only a number's SHAPE is inferred — a decimal point, a percent sign, a currency mark or a table cell makes it a measurement; dates (including year-less "09-14"), years, security codes, line-leading ordinals and code in a language-tagged fence are structure; a plain integer is checked only as a price of an instrument quoted in the thousands — so a sentence gets the same verdict in every language. A measurement that equals a price or volume this session's tools returned needs nothing more. Any other one is declared in a fenced figures block at the end of the answer, one line per figure — value | role | note | ref — and checked by role: observed must appear in the referenced call or tool (ref may name either), filtered to the figure's own symbol, with a currency-marked figure compared only against money-denominated values and a percent only against non-price ones; derived needs an evaluable formula whose added or subtracted operands are all observed, matching the figure at the precision and sign it was written with; proposed is a price inside its own symbol's observed range or a derivation; cited must name its source on the figure's own line, because the block is stripped before anyone reads the answer; count covers counts, weights, thresholds and probabilities, while a currency-marked figure, or one inside its instrument's price range that no derivation uses as a factor, is checked as observed when declared count. No role other than observed may sit in a price column. The block is stripped from the released answer and never streamed; the artifact keeps it. The price-word, level-word, derivation-phrase and metric-subject catalogues are deleted with it (65 compiled regexes → 12, all shape), and src/agent/grounding.py is now the src/agent/grounding/ package.

  • A rejected answer costs one correction round, not four, and is then released with its failing figures cut rather than refused. The correction prompt lists every failing figure as written, with its declared role and what the evidence says (the observed range, the nearest observed values, or what its own formula evaluates to, in the figure's units), and names the three ways out: declare, rewrite to an observed value, or remove. If the second draft still fails, each failing figure is replaced by (略※) / (omitted※) at its own span, bare-integer restatements of a cut figure are swept too, a footnote says how many were omitted and why, and the whole document — footnote included — must pass the same gate before release. The canned fallback is left for an identity finding, a run that never observed a price, or a cut that still fails. Drafts that triggered bounded search_symbol / get_market_data recovery do not spend the correction budget. The run stays degraded with a reason naming the redaction, the release path's rechecks are not counted as rejected drafts, and the shipped document is recorded under released in artifacts/grounding_evidence.json.

  • A draft missing only a source or currency word gets a data note appended, not another model round. The canonical symbol is deliberately not repaired this way: it is the figure's subject, and a note naming it under an answer about another instrument would footnote a misattribution.

  • The chat shows that the answer is being checked. The agent loop emits grounding_status ({stage: "revising", round, issues} after a rejection, {stage: "released_redacted", removed} before a cut release), and the chat page shows "Checking the figures in this answer (round N)…" until answer text arrives, in all eight locales.

  • MetaTrader 5 no longer picks a broker alias when several match (47a036d6, with #1597 and #1598). Quotes, history, orders, symbol search and the backtest loader took the shortest suffixed name (EURUSDm over EURUSDz); they now refuse and ask for the exact native symbol. The connector still tries a configured symbol_suffix first; the backtest loader does not read it.

  • run_card.json schema_version is 1.0 (was 0.1), with the optional tool_traces and citations keys from #1612.

  • Broker connectors: 18, up from 14 (KIS, Upbit, Toss Securities, Scalable Capital). test_readme_counts.EXPECTED_BROKERS is checked against the profile registry, not incremented by hand.

  • The opt-in loader cache is at version 7 (#1411, thanks @shadowinlife). Entries written under version 6, which include Tencent series the forward walk truncated, are never matched again and are fetched afresh.

  • VIBE_TRADING_BAOSTOCK_TIMEOUT (default 30 seconds) bounds each BaoStock lock wait, connection and complete message (#1615, thanks @he-yufeng). It is not a deadline for a whole multi-symbol fetch.

  • The a_share fallback chain is tencent, mootdx, eastmoney, baostock, akshare, tushare, gildata, local (#1474, thanks @sambazhu).

  • Factor values change for 23 alphas, so an IC, bench or comparison stored before this release is not comparable with one computed after it: qlib158_cntp* / cntn* / cntd* (15 factors; the first value lands one bar later and windows touching a gap are NaN, #1459), qlib158_wvma5 … wvma60 (5; every value, #1594), and gtja191_103 / 133 / 177 (3; direction and scale, #1604). Patell z for market_adjusted and mean_adjusted event studies moves slightly as well (#1449, thanks @Shizoqua).

  • A grounding correction turn is offered no tools (#1600, thanks @zeus229) — after a rejected draft the model was handed every tool again, so revising a figure could restart research. Unless the gate asks for recovery (an unresolved symbol or a missing price), the correction turn gets none. A call made anyway is not run and, since 47a036d6, spends one of the two correction rounds; at the cap the last real draft goes through the existing cut-and-recheck release.

Removed

  • The Requesty built-in provider. Its capabilities were OpenRouter's, with no adapter of its own. Same request shape: LANGCHAIN_PROVIDER=openrouter, OPENROUTER_BASE_URL=https://router.requesty.ai/v1, and the Requesty key as OPENROUTER_API_KEY.

Fixed

Long research runs stop losing what they fetched

A user's two runs on the Codex provider (GPT-6 Sol) failed as "0 steps · no_progress": a risk-parity backtest after 1m03s and a comparison of A-share and Hong Kong game companies after 4m50s. They were two different failures.

  • Context compaction follows the model's real context window — compaction fired at a fixed estimate of 40,000 tokens (characters / 4 of the messages, not counting tool definitions), while the system prompt alone is about 12,500 of them. From the third step it cleared every tool result but the last three, so a comparison across a dozen companies re-fetched the same filings until the no-progress stop: reproduced on v0.1.15 with DeepSeek V4 Pro, seven A-share issuers' filings fetched four to five times each, stop at iteration 16; with a larger budget the re-fetches went from 54 to 0. The window now comes from VIBE_TRADING_CONTEXT_WINDOW, else a limit learned from the provider's own context-length error during the run, else src/providers/context_windows.json (272K for the Codex GPT-6 / GPT-5.6 models, 1M for DeepSeek V4, each entry naming its source), else 128K, and VIBE_TRADING_CONTEXT_MAX_TOKENS (200K) caps what a prompt may cost. Usage is read from the input tokens the provider reports. The three layers fire at 80 / 90 / 100% of what the system prompt and tool schemas leave, and clear the oldest results first, only down to the line. A context-length error is compacted and retried instead of failing the run. TOKEN_THRESHOLD is deprecated and ignored with a warning; remove it from .env.

  • read_file says where the agent may read, and lists a directory — the 1m03s failure: on the same prompt, a live DeepSeek V4 run went looking for the backtest engine's source (../../agent/src/backtest, the checkout's absolute path, skills/../backtest/engine.py) and every attempt got the same "File not found or path escapes workspace". Each spelling was a new call and none an observation, so eight rounds exhausted the no-progress budget in 93 seconds, before any backtest ran. The refusal now names the case (outside_readable_roots or not_found), the readable roots (the run directory and skills/) and where the engine's contracts are documented (load_skill); a directory inside those roots returns its entries. The boundary itself is unchanged.

  • A failed or cancelled run keeps its steps — only a completed run saved its tool trail, so the Web history showed "0 steps" for exactly the runs someone would want to inspect.

  • A no-progress stop says which calls produced nothing, and why — each failing or repeated call is listed with its reason and count ("read_file failed (… not found) ×3"), instead of a message about previous summaries and artifact paths written for a different failure.

  • get_market_data fails when no requested symbol returned data — an envelope holding only _unresolved was a success, so a run whose sources were all down counted it as progress. It is now error_code: no_market_data; a partial result is still a success, and grounding still records every symbol as unavailable.

  • Codex (ChatGPT sign-in) keeps the system prompt, its reasoning and its call ids — a system message the loop adds mid-run (a retry hint, a write-the-file reminder) replaced the whole 49,000-character system prompt for every later request; it is now sent as a developer message, as the Codex CLI does. Reasoning items are carried from turn to turn the way the Codex CLI replays them, and resent without them if the backend refuses; tool calls keep the item ids the backend issued instead of a fabricated fc_0; the prompt-cache key is stable for the session instead of changing on every request. The default Codex model is gpt-6-sol: gpt-5.4 is refused for ChatGPT accounts.

A backtest report is checked against the backtest

A risk-parity versus equal-weight report ended in the canned refusal in 12 of 15 replayed runs. Replaying all 26 stored drafts through the gate (880 of 882 recorded rejections reproduced) showed that four in five refused figures were values the backtest itself had written.

  • A backtest's own output grounds the report about it — only metrics whose names the gate knew (Sharpe, return, drawdown, volatility, win rate) were evidence. Every scalar of the run card, the metrics row, the risk X-ray, the Monte Carlo validation and the rebalance summary now is (lists such as the 1,000 simulated Sharpes are not descended, and a file other than the metrics must match the run card's SHA-256 manifest). A per-bar table counts only for the rows a read_file showed, while it is still byte for byte what the engine wrote; a file the model wrote is never evidence. A figure may name the run it came from — rp, rp/artifacts/metrics.csv, risk_parity target_positions.csv, rp::sortino — and a value from one backtest does not ground a figure declared from another. Replayed: value_mismatch 196 → 1, not_in_referenced_call 167 → 7. On the same runs a real value declared with its run passes 97% of the time and the same value moved 2–10% passes 0.7%; the undeclared path was deliberately not widened (the whole metrics row there would have passed 16% of invented values).

  • A comparison's arithmetic is read as it is written — a formula glued to its labels ("等权 Sortino 1.2115 − 风险平价 1.1329") or followed by a bracketed aside was "not arithmetic"; an operand written as a percent was read as a whole number, so "13.28% − 12.98%" anchored on nothing and "0.666 × (1 − 3%)" computed the wrong value; a sum of squared weights (HHI) could not be written at all; a formula's negative operand (a drawdown) could never anchor; and a field two runs both hold anchored neither. All now evaluate (only a literal square or cube is accepted as a power), and the ones that still fail are the model's own arithmetic. A declared value may be a fraction (1/3) or a multiple (5.5x).

  • A report that still fails after a completed analysis is released with the failing figures cut — naming a symbol made any request a market answer, and a market answer with no fetched price was never released cut. A completed analysis now is; a price question with neither prices nor an analysis still gets the refusal, and the refusal after an analysis no longer says the draft cited prices.

  • A draft ending in DeepSeek's closing tool-call tags is caught — </||DSML||invoke> slipped past the pattern for opening tags and reached the gate as unreadable declaration lines.

  • A correction says what went wrong with a figure the model did declare — 0.3897 and 38.97% stay two assertions, and so do −1.51pp and a formula that computes +1.51, but the model is now told which value it declared (also when it has the opposite sign) and that its formula runs the other way round, instead of "not declared" or "its own note evaluates to 1.51%". In a live run after the evidence fix, eight differences were still cut because the second draft rewrote the numbers rather than the formulas. The correction also asks for the corrected answer alone, in the user's language: that draft had opened with "The rejection was because…", in English to a user writing Chinese, and the sentence was released.

The grounding gate

  • A tail-risk figure names its own field once a session holds more than one (#1425, after #1444, thanks @he-yufeng). A call- or tool-scoped ref (ref x1) pooled every tail-risk field that call returned, and an undeclared tail-risk percent pooled every tail-risk value in the session, so an ES 95% could quote the VaR 95% value and pass. The rule is read off the evidence, never off prose: tail_risk_identity() takes (measure, confidence) from the field name a tool returned, so var_95 and es_95 are two identities while cvar_99 and es_99 are one. When a scope holds two or more and the figure's value matches one of them, it is refused with tail_risk_needs_field_ref, whose correction names every identity the session holds; a figure matching no tail-risk value keeps its old reason, so a fabricated number is still a mismatch rather than a missing ref. Accepted cost, stated in the issue: an undeclared VaR 95% beside a VaR 99% is correct today and now costs one correction round. The system prompt states the rule so a risk report declares the ref on its first attempt.

  • Tail-risk figures keep their identity under a field ref (#1444, thanks @he-yufeng). A declared ref may name an evidence field by its full path (data.tail_risk.var_95) or trailing part (var_95), or one call's field (q1::historical_var); a VaR 99% quoting the 95% value, or ES quoting VaR, is rejected under such a ref. A field two calls returned with different values is ambiguous and the correction names the refs to choose from. A claim scoped to a whole call or not declared is still matched against every tail-risk value it can see (#1425 stays open).

  • The grounding gate reads decimal commas (#1517, thanks @zeus229): 17,93 % and 1.410,00 CNY ground against tool results, two unspaced numbers such as 1400,1777 stay two numbers, and a written figure must match its evidence within half a unit of its last written digit (a truncated 30.20% for 30.2052% is now sent back for correction). One shape regex added (14 → 15).

  • A redacted answer keeps its row numbers and its labels (#1471). When the grounding gate released a draft with its unverified figures cut, a ranked table came back with (omitted※) in place of 1, 2, 3, of the 12m in its headers and of "3 names" in the prose. Every integer in a table cell was a measurement, so the rank column was cut; and each cut figure's digits were then swept out of the rest of the page, single digits included. Three shape rules, no word list: a column whose numbered cells read 1, 2, 3 … in row order is the table's index; a plain integer in a cell that also holds a word (12m mean return, Mean beta (9 reported)) is read as it would be in a sentence, while a cell that holds only a number, any decimal, percent or currency figure, and every cell under an OHLC column stay measurements; and the sweep never keys on a single digit. An integer price of an instrument quoted in the thousands is still checked inside a worded cell.

  • A real VaR result grounds the VaR it returned (#1464). quantlib_call records a scalar result under the function name. For historical_var and parametric_var that name reduces to var, which the grounding gate accepts only as a whole field name, so the result recorded no evidence and a correct answer quoting it was rejected. Both names are now tail-risk aliases. Variances such as residual_var stay unmapped.

  • Plain integers in prose are not price claims. A list number ("输出原则 4"), a window length or a count was checked like a quoted price and could be cut from a released answer.

  • A drawdown against an observed high passes wherever the endpoints are written. "较 5 月高点 1.053 元已回撤约 37%" was rejected when its endpoints sat on another line; a derived declaration does not depend on layout.

  • Price-denominated indicator values are evidence, by registration. Moving averages, Bollinger bands and similar levels returned by technical_indicators are registered per tool as observed price evidence; RSI, MACD and other non-price leaves are not, and nothing is decided from a field name's words.

  • A year-less date is a date, and "24.6M" is 24.6 million. Two real runs that followed the contract were still released redacted: "2026-09-11 / 09-14" in a table cell shipped as "2026-09-11 /(略※)-(略※)", and "24.6M → 22.6M lots" as "(omitted※)M → (omitted※)M". A zero-padded MM-DD, or one opened by a full date in the same cell or line, is structure (an unpadded "| 11-12 |" is still checked, since it may be a price range). K / M / B / 千 / 万 / 亿 glued to a figure scale its comparison with the evidence, never its shape, and a cut removes the mark with the figure. Replayed on both runs' tool results, each second draft now passes as written.

  • Metadata counts are not metrics, and a tool's tail-risk result can ground a VaR (#1420, #1426). return_observations, n_returns, return_window, vol_lookback, max_drawdown_duration, aligned_days and their *_window / *_lookback / *_obs / n_* / *_count / *_days / *_duration families no longer ground a percentage ("年化收益 81%" from 81 observations); var, var_95, var_99, cvar, es and expected_shortfall leaves are tail-risk evidence, so a correct VaR is no longer refused. A field's metric comes from the head of its name, so sharpe_sample_size and drawdown_threshold are no longer a Sharpe ratio or a drawdown.

  • A correction names the prints of the figure's own instrument (#1433). The "nearest observed" values come from that symbol (and its table column or ref), not from every field of every symbol in the run.

  • A number is read the way the chat renders it (#1418). Zero-width characters, a full-width or Arabic decimal point, a no-break space before "%", the U+2212 minus sign, a backslash escape ("0\.888"), an HTML entity and an ISO code glued to the digits ("CNY0.888") no longer split a price into unchecked integers; a decimal comma ("0,666 CNY", "12,5 %") is a decimal while a valid grouping ("-1,250.00") stays grouped, and an answer that writes a marked decimal comma reads "2,237" and "−5,132%" beside it as decimals too. Fences follow CommonMark, and an unterminated code fence no longer exempts the rest of the answer.

  • A year is not a price shield. "$2050", "1999 元" and a 2031 under a close column are checked; a price placed in a date or code column is checked; a compact date, a dotted date and a time are structure. 円 and 원 are currency marks, and "52pp" / "5200bp" are measurements.

  • A declaration cannot relocate or launder a figure. A figure the sentence writes about one instrument cannot be declared another's close; a currency-marked result cannot be derived from an RSI or a volume; count, derived and proposed cannot sit in a price column.

  • A general answer with no tool evidence is cut, not refused. A question about no instrument ("印花税怎么收") whose figures the gate cannot check is released with those figures omitted and a footnote, instead of a refusal about prices it never mentioned; a market answer that observed no price still falls back. A malformed figures block no longer forces the fallback either: the block is dropped and the draft is checked as written.

  • The tool-call-syntax fallback reaches the chat once. When a model answered the forced-text iteration with tool-call markup in a run whose output is buffered, the replacement message was streamed by its own branch and again as the released answer, so the chat and the CLI showed it twice.

  • No unchecked number reaches the chat before the gate runs. An unbuffered answer (no instrument asked about, no tool evidence yet) streamed its first draft live, numbers included, and a rejection then replaced it; the stream now stops at the first measurement-shaped number, and a rejected draft's shown prefix is cleared with stream_reset.

  • A trading plan in a bare code fence is checked. Only a fence tagged with a language holds code; an untagged fence is read as prose.

  • An integer price of an instrument quoted in the thousands is checked ("600519.SH 最新收盘 1520"), within half to twice its observed range; a window such as "200 日均线" stays unchecked.

  • A price declared count is checked. A number inside its instrument's observed price range may be a count only when a declared derivation uses it as a factor.

  • A redaction footnote the model wrote itself is removed before the real cuts, so a released answer never carries two contradictory footnotes or a marker for a cut that did not happen.

  • A figure about a second instrument is checked against that instrument's evidence (#1584, thanks @zeus229) — symbol resolution saw only price-comparable records, so a line naming an instrument the session held only fundamentals for took the answer's primary symbol and failed as not_in_referenced_call. Any observed evidence with an explicit symbol now resolves it; 47a036d6 scopes analysis metrics likewise, so a metric from a call about another symbol, or several, cannot ground it.

  • A field ref may contain |, and a correction for an ambiguous field lists the exact refs to use (#1586, thanks @zeus229) — the figures parser split on every |, cutting such a ref short; a declaration is four columns now, the last taking the rest of the line. When several calls returned a field, the correction lists each valid call_id::field (only the figure's own symbol's since 47a036d6), and a tool_name::field ref is refused as field_ref_needs_call_id with that list.

  • A decimal-comma answer's dot-grouped integers are read as thousands (#1588, thanks @zeus229) — AR$5.165; volumen 502.408 was read as two decimals and failed as figure_undeclared. Once a document writes a decimal comma, now including a 0, with two or more digits, 5.165 is 5165 and is still checked; without that evidence it stays 5.165. The PR's wider 1% rounding band was not kept (47a036d6): at 0.5%, 0.82467 has to be written 0,825, not 0,82.

  • Figures from technical_indicators keep the date the tool reported (#1590, thanks @zeus229) — the tool dates its snapshot as latest_date, a key evidence ingestion did not read, so latest_close, the indicators and the volume metrics landed in grounding_evidence.json with timestamp: null. It is inherited now, below a row's own date fields and above as_of; an undated payload stays undated.

Market data

  • get_market_data asks for AAPL.US instead of returning empty frames for a bare US ticker — only codes the market detector reads as US equities are refused; a Chinese futures code (RB0, IF2412) or a joined crypto pair (BTCUSDT) is served as written. The public-source health canary probes Yahoo with AAPL.US for the same reason.

  • A mixed-market backtest falls back per market — a symbol missing from its primary source was retried along the fallback chain of the first symbol's market, so a Hong Kong gap in an A-share-led universe was never retried on a Hong Kong source. Each missing symbol now follows its own market's chain, and the incomplete-data error lists which symbols were served.

  • A dollar-quoted BYMA or TSX line, or a Hong Kong RMB or USD counter, no longer enters a single-currency backtest (#1576, after #1543, thanks @Shizoqua, @zeus229). Both venues list USD lines beside their home-currency ones. On 2026-09-24 Yahoo declared USD for GGALD.BA (4.20), SPYD.BA and AAPLD.BA, and for DLR-U.TO and XUS-U.TO. ar_equity and ca_equity are each one static-currency pool, so such a line was priced as pesos or Canadian dollars. As on the LSE, the loaders now read the declared currency. They admit a .BA line only in ARS and a .TO / .V line only in CAD. A trailing D does not decide it: YPFD.BA is a peso line. Every other declared quote currency is recorded in market-data provenance, and the grounding gate asks an answer to name that currency rather than the one the suffix implies. Hong Kong is decided by the code, because the Hong Kong chain starts with sources that declare no currency: HKEX's Stock Code Allocation Plan assigns 80000-89999 to products traded in renminbi and several 09xxx ranges (plus 10900-10999 and 41500-41599) to products traded in USD. So an RMB counter (80700.HK, 375.40 CNY on 2026-09-24) or a USD ETF (9834.HK) is refused in a Hong Kong or cross-market backtest, is counted in its own currency by the mixed-currency guard, and is quoted in that currency by the grounding gate, whichever source served it. The table matched the currency Yahoo declared for all 24 codes probed across the ranges.

  • Grounding identity and symbol search know every market the data layer routes (#1565, #1575, thanks @Shizoqua). .BA, .L and .VN symbols had no canonical-symbol scan, venue or currency in grounding. Symbol search labelled Indian, Korean, British, Vietnamese and Argentine results global. A parity test now takes the market list from the backtest's own currency table. An answer may write £, ₫ or AR$ for the currency it has to name.

  • A local: code in a backtest is served from your own dataset or not at all (#1467). The market-data tool and the README already treated local:AAPL.US this way, but the backtest runner did not. With source="local" it counted the served AAPL.US as missing and sent it down the A-share network fallback chain. With source="auto" it ignored the prefix and fetched the symbol from network loaders. The prefix now picks the local loader only, a symbol the dataset lacks fails with incomplete data for source=local, and the rest of the run sees the bare symbol: market rules, artifacts (ohlcv_AAPL.US.csv, no colon in the file name) and the run card. A local: code requested from a network source, or one symbol requested both with and without the prefix, is refused up front.

  • Block trades, margin trading and financial statements no longer read Eastmoney rejecting a stale query (code 9501) as an empty result; all six datacenter callers share eastmoney_client.datacenter_rejection.

  • Lockup expiry: free_shares is the unlocking quantity (CURRENT_FREE_SHARES), not the float; units are stated (万股 / 万元) (#1513, #1501, thanks @lorenzozanee). The report is RPT_LIFT_STAGE; the old one rejected six columns.

  • Dragon-tiger seats come from the buy/sell detail reports, surface a provider rejection, and rank within each listing reason (#1512, #1502, thanks @lorenzozanee).

  • Shareholder-count history returns real periods from RPT_HOLDERNUM_DET, states each row's previous period, and warns when the average-holding join fails (#1518, #1503, thanks @he-yufeng).

  • QVeris resolves adjusted fields as a complete set, one family per response, and no longer gives an unadjusted bar the adjusted volume (#1527, #1494, thanks @cgycorey).

  • Tencent history walks backward from the end date (#1411, closes #1410, thanks @shadowinlife) — fqkline serves the last 500 bars of a window, not the first: sh000300 over 2018-01-01..2026-06-30 comes back as 500 bars starting 2024-06-06 (verified live). The forward walk that shipped in 0.1.15 (#1154) therefore returned only the last 500 bars of any longer window, with exit code 0. The cursor now moves back from end_date, and an error-shaped or mid-walk empty reply is retried instead of being read as the start of history.

  • A BaoStock server that stops answering no longer hangs the fetch (#1615, closes #1492, thanks @he-yufeng) — baostock 0.9.3 reads its socket with no timeout, so a silent server blocked forever and a closed connection spun at 100% CPU. The connection, each complete message and the wait for the session are now bounded by VIBE_TRADING_BAOSTOCK_TIMEOUT (30 seconds by default), and a failed call is a per-symbol miss the chain moves past. After review, BaoStock sessions also run one at a time per process, because the SDK keeps one global socket, and a socket that failed mid-reply is closed so the next query cannot read its late answer.

  • Concurrent first lookups no longer find an empty loader registry (#1405, closes #1403, thanks @alanwilhelm) — _ensure_registered marked itself done before importing the loader modules, so a thread arriving during a cold start returned with nothing registered and an available symbol came back _unresolved. The flag is published after the imports now, under a lock, and concurrent cold callers wait for them.

  • An explicit source that must not fall back no longer has its gaps filled (#1441, thanks @chiww) — local, fmp, qveris, tickerall, nobitex and wallex were kept from degrading only when the loader was unavailable as a whole. A symbol the named source lacked was still fetched down the market's chain, so one run could mix a local snapshot with Tencent bars behind two log lines. The run now fails naming the missing symbols, pinned by a test for every source in that set.

  • The Stooq challenge warning names a variable that exists (#1402, thanks @chiww, thanks @chiww) — it told operators to set VIBE_TRADING_MARKET_DATA_ORDER_*, which nothing reads, or to drop stooq from the chain, which override validation refuses. It names MARKET_DATA_ORDER_* now (for example MARKET_DATA_ORDER_US_EQUITY), suggests moving stooq to the end, and says an override must be a permutation of the default chain.

  • QVeris no longer sends the end date as indicators (#1496, thanks @lorenzozanee, thanks @lorenzozanee) — end-date parameters were recognised by substring, so any capability parameter containing to, such as indicators, was given the end date. Only exact names are mapped now (end, end_date, enddate, to, to_date, until, until_date, with hyphens folded). The rest of #1494 was fixed separately.

  • get_shareholder_count answers again (#1490, closes #1489, thanks @cgycorey, thanks @cgycorey) — it asked RPT_HOLDERNUMLATEST for AVG_HOLD_AMT, a column the report no longer has, so Eastmoney rejected the whole request and the tool reported "no shareholder-count disclosure found" for every A-share. It requests AVG_MARKET_CAP, the same per-account value (the output key avg_hold_amount is unchanged), and a rejected request now returns the datacenter's own message instead of an empty result.

  • get_market_data no longer answers 1M with one-minute bars (#1487, closes #1480, thanks @shadowinlife) — intervals were matched through an upper-case fold under which 1m (a minute) and 1M (a month) share a key, so a monthly request was served minute bars without an error. An exact-case match now wins before any fold. The PR refused 1M and 1W; since #1479, in this same release, they return monthly and weekly bars built from daily ones.

Backtest runner and engines

  • Backtests align calendars at any timestamp resolution (#1560, thanks @Shizoqua). The engine merged symbol calendars and placed prices through raw nanosecond integers, but a local duckdb source serves microseconds. Such a run was dated 1970-01-21, and beside a nanosecond source one symbol's closes came back empty. The rebalance mask had the same unit mismatch.

  • Mean-variance style optimizers size a short by its own expected return (#1548, thanks @Shizoqua). The optimizers scored every position as if it were held long, so the strongest shorts (most negative drift) looked worst and got the least capital. BaseOptimizer.optimize now negates a short's column before building the context, so mu is the position's expected return AND cov is the position covariance D Σ D: a long and a short of two +0.92 correlated names hedge instead of reading as correlated (the old objective put the whole book on the long, the new one splits it 0.514 / -0.486). Risk-parity and max-diversification contexts get the position covariance with it; volatility-only contexts are unchanged, since std(-r) == std(r).

  • Monte-Carlo validation annualises at the venue's bars per year (#1546, thanks @Shizoqua). It hardcoded 252 while bootstrap_sharpe_ci and walk_forward_analysis in the same report used the resolved value, so a crypto (365) or forex (260) run published two disagreeing Sharpes.

  • A futures order on a negative-price bar is no longer rejected outright (#1547, thanks @Shizoqua). FuturesBaseEngine dropped the abs(price) its base class relies on under allow_nonpositive_prices, so size came out negative and the caller's size <= 0 guard refused every order.

  • A backtest aborted when a funding debit left cash below zero (#1542). CompositeEngine and CryptoEngine subtract crypto funding from capital with no floor, and the next open then fitted at no scale, not even an empty plan: the open-basket search kept the full-scale plan and the run died with "planned order … exceeds available capital", and a rebalance raised "insufficient capital for position rebalance". Such an open is now skipped and reported once as insufficient_capital, the bar's reductions still run, and a close whose loss exceeds its margin still aborts the bar (#1274).

  • One cash-starved open is one rejection, not twenty-six (#1470). In position_adjustment="hold", the search that scales a basket down to the cash available re-planned every open on each step, and every step whose size rounded to zero booked a zero_size rejection for the same symbol on the same bar — one contract the cash could not hold read as 26 lot-rounding failures. The trial plans are silent now; a sleeve that was a real order at full scale and left the basket is reported once, after the search, as insufficient_capital, which the rebalance path's dropped sleeves now use too. zero_size means the lot rule and nothing else.

  • A-shares from a source without its own engine branch run under A-share rules (#1431, thanks @chiww) — for such sources the runner followed the instrument's market only for US, HK, Canadian and UK equities, so 600519.SH served by local, tencent, eastmoney, baostock, mootdx or sina ran on CryptoEngine: no stamp tax, T+1, price limits or 100-share lots, and a perpetual funding fee charged against the position. They route to ChinaAEngine now.

  • The declared interval is checked against the bars that arrived (#1432, thanks @chiww) — a single-market run annualised at the interval it asked for, but a loader can serve coarser bars (the local loader cannot turn a daily file into 1H), and CAGR, Sharpe and volatility then scale by the ratio. The runner compares the median bar spacing of the longest price frame with the declared interval; past a 1.5x gap either way it annualises as the interval the spacing matches, or from the calendar when bars four or more days apart match none, and the report is written to the run card beside the caliber warning. Median spacing replaced a bars-per-calendar-year count, which measures five daily bars starting on a Monday at 456 a year.

  • Options annual return is annualised over the elapsed bars (#1458, thanks @Shizoqua, thanks @Shizoqua) — the options engine raised growth to bars_per_year / (n - 1) where the shared calc_metrics uses bars_per_year / n, which exaggerated annual return in both directions, most on short runs. Both use n now.

  • An India short is covered against the buy-side circuit band (#1608, thanks @Shizoqua) — the India engine never passed the held position's direction to the shared band check, so covering a short, which is a buy, was tested against the lower band a long's sale uses, and a cover at a locked upper circuit filled. The direction is passed now, in single-market and composite runs alike.

The backtest tool and what a run leaves behind

  • The backtest tool honours VIBE_TRADING_TOOL_TIMEOUT_SECONDS and keeps a timed-out run's output — the engine subprocess had a hardcoded 300-second limit regardless of the setting (default 1,800; 0 disables it), and a timeout discarded everything the engine had printed. The partial stdout and stderr are now written to the run's logs/ and the last 2,000 characters returned with the error.

  • Runs from the backtest tool record state.json (#1413, closes #1412, thanks @shadowinlife) — only the agent runtime wrote it and refresh_strategy_evidence refuses a run without one, so every run from the backtest tool, or the MCP tool that delegates to it, was rejected as run status unknown despite complete artifacts. The tool now records success, failure with the exit code, or a timeout, through the same RunStateStore writer; a config that fails validation records nothing, since the engine never ran.

  • The backtest tool names the directory it looked in (#1429, thanks @cgycorey, thanks @cgycorey) — a missing config.json or code/signal_engine.py was reported with no path, so a caller whose run_dir resolved somewhere unexpected could not tell that from an empty directory. Both errors now name the resolved directory and say where the file belongs.

  • Dict- and list-valued metrics reach the run card (#1424, thanks @cgycorey, thanks @cgycorey) — the card kept only scalar metrics, so unfilled_plan_rejections_by_symbol, which names the rejected sleeve and the reason, was dropped while the anonymous count beside it stayed. Non-scalar metrics now go to a structured_metrics block, shown in the Run Card tab and the Markdown card, with metrics unchanged; a value over 4 KiB of card JSON, or one that cannot be serialised, is left out and named under _omitted.

  • A text column in metrics.csv no longer drops a run's metrics from the reply (#1579, thanks @Shizoqua) — the session service converted the whole row at once and swallowed the error, so a run with an explicit benchmark, whose benchmark_ticker is text, lost every metric without a log line. Each field is converted on its own now, and only the unconvertible one is skipped.

Alpha Zoo

  • 24 alphas emitted a value computed from a missing bar (#1463, #1523, #1534, #1452, thanks @Shizoqua). A comparison over an operand that is NaN because its window holds the gap is False, and the #1463 sweep, which nudged the bar by a tick, could not see that. A second oracle compares the gapped run with the gap-free one. Comparison gates and np.fmin/np.fmax now mask on their inputs' reach (factors.base.observed_over) rather than on their operands' NaN, so a correlation undefined on complete data keeps its old verdict; the six fmax/fmin alphas converted on 2026-09-18 had lost up to 87% of their post-warmup values in a 5-symbol universe that way and are restored. On gap-free data no finite value changed.

  • Twenty alphas no longer fill a missing input with a constant (#1463). A halt left a comparison reading 0 or False, a .where(cond, 0) reading a flat day, or an np.fmax returning the other side, and the value then fed every rolling window that reached back to the gap. Those cells are NaN now. On gap-free data every finite value is bit-identical to before. A perturbation sweep over all 462 alphas counts 35 still carrying a missing bar forward, down from 55, and every one of them is accounted for: 28 are recursive statistics (the GTJA SMA(A, n, m) smoothers and one running product), which by the policy set on #1463 skip a missing observation and continue from their last state — the operator layer's header now states that exception, and a test pins it; 3 declare their own partial window; 4 are the sweep's own rank-tie artifact. Nothing rewarms a smoother after a gap, and the registry's look-back mask was measured and rejected: it hid 4,270 legitimate cells to remove 41% of the fabricated ones.

  • Asset growth is year over year on the daily panel (#1564, thanks @Shizoqua). Derived fundamentals run on a panel densified to the price calendar, where a one-row lag compared yesterday with today: 0% on ordinary days and a one-day spike at each filing. The lag is now 365 days, as-of.

  • A factor's IC ratio is unset for a negative baseline (#1549, thanks @Shizoqua). Dividing two negative IC means gives a positive ratio, so a rolling IC that got further below zero reported as improvement. The signal itself took the worst of the available metrics and was not fooled; the published number was.

  • qlib158 CNTP / CNTN / CNTD stop counting a missing close as a flat day (#1459, thanks @cgycorey) — the 15 factors (three families × 5 / 10 / 20 / 30 / 60 windows) turn close > prev close into 0/1 and average it, and a NaN comparison reads False. The registry mask from #1377 blanks the gap bar itself, but every later window that still held the gap counted the halt as a day that neither rose nor fell, so cntd read closer to zero than the observed days say. The comparison is NaN now wherever either close is missing: a gap blanks window + 1 rows and the value returns, equal to the gap-free run, once it scrolls out. The same rule moves the first value from row window - 1 to row window, matching each factor's min_warmup_bars, because the first bar has no prior close. On sparse panels this can trip the registry's >95% NaN guard; each factor's notes states from what missing-bar fraction.

  • qlib158 WVMA takes the absolute return in its numerator, as upstream does (#1594, thanks @Shizoqua) — wvma5 … wvma60 computed ts_std(ret·v) / ts_mean(|ret|·v), a signed numerator over an absolute denominator, where qlib's formula takes Abs on both. Every value of the five factors changes. The finishing commit (a5d47c20) points the attribution header at the upstream file that actually defines WVMA (qlib/contrib/data/loader.py, MIT) instead of handler.py, and pins the formula against an independent scalar oracle, a gap and all-zero volume.

  • GTJA alphas 103, 133 and 177 score a fresh extreme high, not low (#1604, thanks @Shizoqua) — HIGHDAY / LOWDAY count bars since the extreme, but the port put the 0-based window index, counted from the oldest bar, in their place. Alpha 103 read a low set today as 5 and a low 19 bars old as 100: the recency signal was inverted. The PR restored the direction one bar off (95 and 0). The finishing commit (0244ecea) counts days since the extreme with today as 0, so alphas 103 and 177 run from 5 to 100 over a 20-bar window, and takes the most recent bar on a tie. It also corrects gtja191/LICENSE.md to cite the Guotai Junan report of 2017-06-15 by Li Chen and Liu Fubing, where the zoo had named a 2014 report.

  • A custom zoo root no longer replaces the bundled alpha of the same name (#1468, fixes #1465, thanks @cgycorey) — Registry(zoo_root=...) loaded custom alpha files into sys.modules under the bundled module's own name. Once a custom root had computed its own qlib158/cntp5.py, a default Registry() in the same process computed the custom code too. The import cache is keyed by module path and zoo root now; module names and relative imports inside a custom root are unchanged. No shipped caller passes zoo_root; it is the path tests and plugins use.

Quant Library and analysis tools

  • A strict-bench OOS split needs two IC observations on each side (#1559, thanks @Shizoqua). A split outside the loaded prices is refused. An alpha whose IC series leaves fewer than two observations on either side is skipped, with the counts in the reason. The IC series ends one forward-return horizon before the prices, so a split one bar from the end used to publish verdicts measured on nothing.

  • Quant inputs are checked before they are used (#1555-#1558, thanks @Shizoqua). HRP aligns a supplied correlation matrix, and the covariance's own rows, to the covariance's column labels; swapping two rows used to move the weights with no error. Purged cross-validation refuses an unordered or duplicated timestamp index, and a label that ends before it starts. Impact models and every fixed-income entry point refuse NaN and infinities instead of returning NaN prices, durations and curves.

  • Shadow-account overtrading uses the whole trading window (#1563, thanks @Shizoqua). A long hold that closed before later short trades no longer shrinks the span.

  • Event-study z-statistics no longer over-reject (#1466). A CAR's standard error summed each day's variance, ignoring that every day is forecast with the same estimated parameters, so the days' abnormal returns are correlated. On null data, the standardised CAR's variance was 1.11 with a 120-day estimation window and 1.41–1.46 with a 30-day one. It is now 1.02 and 1.07, which is only the t correction for an estimated residual variance. Patell and BMP inherit the fix; market_adjusted was already right.

  • Shadow account: an other-market rule no longer matches every symbol (#1538, thanks @Shizoqua), and overtrading PnL no longer double-counts trades already explained (#1475).

  • Quantlib: comps refuses negative bridge magnitudes like the DCF bridge (#1535, thanks @Shizoqua); accrued interest is exact on coupon dates under 30/360 and rejects an unknown day count there too (#1509); Kyle's lambda fits an intercept (#1477).

  • ML strategy: a single-class training window is skipped and the previous model keeps predicting (#1522, thanks @Shizoqua). Memory: two entries with one title and different types keep separate index rows, and forget names the type when ambiguous (#1525). Swarm run and task files use unique temp names (#1536).

  • Patell z uses each normal-return model's own degrees of freedom (#1449, thanks @Shizoqua) — the df / (df - 2) correction in event_study assumed the market model's two estimated parameters for every event, so market_adjusted (none estimated) and mean_adjusted (one) over-stated the variance and reported a slightly smaller |z|, and a larger p-value, than their own fits support. The market model is unchanged. The finishing commit (844b13b9) has estimate_market_model read its residual degrees of freedom from the same table, so the two cannot drift apart.

  • Microstructure estimators refuse NaN and infinite inputs (#1451, thanks @Shizoqua) — a non-finite price came back from roll_effective_spread as a 0.0 spread, kyles_lambda returned nan, and vpin produced a series from NaN or negative volumes. All three raise now, and vpin also refuses a non-finite bucket_size. amihud_illiquidity drops an infinite row the way it already dropped a NaN one; before, an infinite return made the ratio infinite and an infinite volume counted as a zero-illiquidity day. The accept side is pinned too (c6010208): constant prices are a zero spread, all-zero order flow a zero lambda, and zero-volume bars leave VPIN unchanged.

  • Copula pseudo-observations give ties one rank and leave NaN missing (#1453, thanks @Shizoqua) — the double argsort ranked tied values by position ([1, 1, 3] became 0.25 / 0.5 / 0.75) and sorted a NaN last, handing it the column's top rank and counting it in the normaliser. pseudo_observations uses average ranks now, keeps a NaN as NaN with each column normalised by its own count of observations, and refuses infinite values.

  • spread_term_structure keeps tenor 5 and 5.0 in one column (#1507, thanks @Shizoqua) — the column name was built from the raw key, so issuers quoting the same tenor as an int and as a float were split across 5Y_spread_bp and 5.0Y_spread_bp, each NaN for the other group. The reference-curve lookup already treated the two as one tenor.

  • Purged cross-validation refuses a timestamp label index it cannot map (#1556, thanks @Shizoqua) — label end times passed as a Series are turned into positions with searchsorted, which assumes a sorted index. An unordered or duplicated index mapped label ends to the wrong rows with no error, so the purge protected the wrong rows; it raises now. The finishing commit (250ef746) closes the neighbouring gap: a timestamp label ending before its own start was clipped up to it (ends at positions [1, 0, 3, 3] came back as [1, 1, 3, 3]) and now raises, as the positional form already did. A NaT end stays allowed and purges through the last observation.

  • Market-impact models refuse infinite inputs (#1557, thanks @Shizoqua) — NaN was already refused, but positive infinity passed every positivity check: an infinite ADV priced a linear_impact or sqrt_impact order at zero impact, and an infinite price, traded volume or bps returned an infinite buy fill. Price, volume, ADV, bps, impact coefficient, volatility and eta must all be finite now.

  • The correlation-analysis cointegration examples call the Quant Library (#1404, thanks @Shizoqua) — the skill's Engle-Granger example carried its own statsmodels fit and raised KeyError on an unnamed Series, because it looked the hedge ratio up by the series' name. The PR named the regressor; the finishing commit (bd29a187) replaced the local fit with find_hedge_ratio, cointegration_test and adf_test from src.quantlib.timeseries, and the half-life example with compute_half_life. The example now inherits their refusal to pair two legs on different indices positionally, keeps its result keys and adds half_life.

  • portfolio_risk_xray works on windows longer than 250 bars (#1462, closes #1461, thanks @he-yufeng) — the tool fetched prices under the shared 250-row cap. Above it fetch_market_data returns a sampled envelope rather than the bar list the tool parses, so a window of 251 bars or more failed with "no close prices returned for any requested symbol", which is about anything over a year of daily bars. It requests every row now (max_rows=0, as technical_indicators does); volatility, VaR/ES and drawdown need consecutive bars regardless.

  • benford_check reads the leading digit from the number as written (#1591, thanks @Shizoqua) — financial_rigor's Benford check took the digit from 10 ** frac(log10(v)), which lands just under an integer for some exact round values, so 5000 counted as a leading 4. Among the integers 1 to 100,000, twelve were misread (8, 30, 40, 50, 80, 500, 600, 900, 5000, 6000, 9000, 50000). The PR rounded before truncating, which would have read 4999.9999999 as a 5; the finishing commit (d0ef98a6) takes the first significant digit of a Decimal built from the supplied value instead, with no float round-trip. Zero, NaN and infinite values are still skipped.

  • The sentiment scorer reads "low" as bearish only (#1448, thanks @Shizoqua, thanks @Shizoqua) — "low" was in both the bullish and the bearish word lists, so it cancelled itself and shares tumbled to a new low today scored neutral. It was the only word in both; a test now keeps the lists disjoint.

  • Head-and-shoulders detection checks shoulder symmetry on negative series (#1456, thanks @Shizoqua) — the pattern tool divided the shoulder gap by the signed shoulder average. On a series below zero the ratio went negative and always passed the 5% test, so clearly asymmetric shoulders were flagged as a pattern. It divides by the absolute average now, as double_top_bottom already did.

Live trading, brokers and the strategy store

  • Robinhood live trading reads and trades the account the mandate names (#1442). The runner, the pre-trade gate and the commit-time ceiling fetch called Robinhood with no account_number, and read replies one level too shallow (data.positions where Robinhood nests data.data.positions). So every reconciliation aborted, and the gate refused every order because it could not read positions. The tests stayed green only because their fake replies had a shape Robinhood never sends. A Robinhood mandate is now bound at commit to one account from get_accounts: listed, not deactivated, and open to agentic trading. The Web confirm dialog and the CLI both ask for it, and never take it from the agent-written proposal. Every runner and gate call sends that account, and an order naming another account is refused. Held positions, which Robinhood reports without a price, are priced the way a quantity order already was, and one that cannot be priced fails the exposure check closed. Existing Robinhood mandates must be committed again with an account.

  • SDM decay metrics treat NaN and infinite bench results as missing (#1447, thanks @0xouzm) — three stored sharpe=inf rows read as healthy, so a non-dry sdm_decay_scan moved a strategy in monitoring back to active, and -inf read as critical. SQLite keeps infinities as written, so the persistent store was affected too. compute_decay_metrics now drops non-finite values, independently for IC and Sharpe, before the three-observation minimum.

  • A mandate narrowed to zero exposure is stored as zero (#1593, thanks @Shizoqua, thanks @Shizoqua) — narrowing max_total_exposure_usd or account_funding_usd to 0 at commit is legal, but the cap mapping used x or default, so the saved mandate got the order cap back as its exposure limit: the commit widened what the user had just narrowed. An explicit zero now persists, and an order checked against that mandate breaches max_total_exposure_usd.

  • A malformed Robinhood buying_power is a shape error, not an unknown (#1526, thanks @lorenzozanee) — a non-null buying_power that was not an object (a bare number, string or list) was read as unknown and passed; it now raises, so the portfolio source errors instead of carrying the reply. An omitted or null value stays unknown and zero stays valid. As merged the PR also refused an omitted buying_power; 46892136 restored that case to unknown, which is what the PR description specified.

  • Interval and event triggers carry market=None again (#1599, thanks @woshi77777stars, thanks @woshi77777stars) — Trigger.market was both a dataclass field and a classmethod factory, so every non-market trigger carried the bound method as its market, and dataclasses.asdict() + json.dumps() raised on it. due_now() was unaffected. The factory is bound after the class body (d0ef98a6 keeps subclass binding intact), and a repo-wide static test fails on any dataclass or NamedTuple field shadowed this way.

  • With an MT5 profile selected, search_symbol resolves the broker's own symbols (#1597, closes #1589, thanks @lorenzozanee) — it returned no MT5 candidate for an explicit symbol such as XAUUSD, so the identity gate blocked trading_quote, trading_history and get_market_data. Search now reads the configured terminal's catalog by terminal_path with no login credentials. After 47a036d6 every query is answered from that terminal while MT5 is selected, a web-search hit is never substituted, and an ambiguous alias is refused.

  • Backtests attach to the running MT5 terminal without seeing its credentials (#1598, thanks @he-yufeng) — the sandbox's ephemeral HOME had no mt5.json, so the loader attached bare and a forex backtest could fall through to akshare while get_market_data used MT5 in-process. The sandbox now gets only terminal_path and timeout (validated by 47a036d6), never login, password or server, and a bare attach warns that the SDK may pick another installation logged into another account.

  • Dhan paper orders reject fractional share quantities (#1595, thanks @Shizoqua, thanks @Shizoqua) — 0.5 passed the positive check and truncated to 0, returning a successful simulated fill for zero shares, and 1.5 silently became 1. Only whole numbers of at least one share are accepted; a5d47c20 parses with Decimal, so 1.00000000000000001 is refused too, and NaN, infinity or text return an error instead of raising.

Swarm workers

  • A swarm worker can load only the skills its spec lists (#1414, thanks @shadowinlife, thanks @shadowinlife) — SwarmAgentSpec.skills was read only to filter the prompt, so load_skill loaded any skill a worker named. The worker's load_skill now refuses a skill outside the list and names the ones allowed. An empty list stays unrestricted, as the prompt filter treats it; the main agent is unchanged.

  • A swarm tool that declares run_dir gets the directory the model named, confined to the agent's workspace (#1430, thanks @cgycorey) — the worker overwrote every call's run_dir with the workspace, so backtest, exposed to 19 preset agents, never saw the directory it was given. A relative path now resolves under the workspace, an absolute one inside it is kept, and one outside is refused with an error naming the workspace. Tools that do not declare run_dir still get the workspace.

  • Every swarm worker that writes its summary also writes its message log (#1439, thanks @cgycorey) — messages.json is the only record of the arguments the model asked a tool for, and the completed, token-limit, content-filter and incomplete exits wrote summary.md without it, so the most common outcome left no log. Both files now go through one writer. A failed LLM call still writes neither.

Agent loop, providers and goals

  • An explicit provider header survives an ambient twin (#1573, #1568, thanks @zeus229). When OPENAI_CUSTOM_HEADERS names a header the provider also sets (a user-agent beside the Kimi / NVIDIA / OpenCode User-Agent), openai 3.19.2's case-insensitive merge dropped the provider's value. That turned CI red for every PR on 2026-09-24. The first fix proposed would have sent both headers under openai 2.53, the version the lock file and the Docker image install. Every ambient spelling is now omitted and the explicit value re-set after the omits, which holds under both.

  • Read-only results lost to context compaction are restored, not refetched (#1488, thanks @zeus229). A successful read-only call whose payload compaction removed is replayed from the run's own cache, at most six times a run; past the cap it runs again, as it did before, instead of being refused with "use the previous result" for a payload the model cannot see. The cap is checked per restore, and any successful write empties the cache, so a regenerated file is read afresh. read_url opts in; no_cache=true always fetches.

  • OpenAI's gpt-5.6 models run with their reasoning on the agent's default configuration (#1473). /v1/chat/completions refuses function tools for gpt-5.6-terra and gpt-5.6-luna unless reasoning_effort is none, including when no effort is configured and the model applies its own default, so a plain LANGCHAIN_PROVIDER=openai setup failed on its first tool call with Function tools with reasoning_effort are not supported. The adapter now treats that refusal the way it treats a rejected temperature or stream_options: the request is retried on /v1/responses, where the effort travels as reasoning.effort, and the model is remembered for the rest of the process so later calls go there first. Nothing changes for a model whose chat endpoint accepts tools, and LANGCHAIN_USE_RESPONSES_API=true still selects the route up front and skips the one failed request.

  • A research goal that is an order is still refused (#1562, thanks @Shizoqua). The execution filter no longer rejects research that mentions shares or coins ("should the fund sell its GOOGL shares"). It refuses an objective that opens with buy or sell, puts a quantity right after the verb, or writes 买入 / 卖出 with a quantity and a unit. The first version of the change had made "Buy 100 shares of NVDA" an accepted goal.

  • Research goals refuse an order that puts "now" or "immediately" before the verb (#1605, thanks @Shizoqua) — the execution filter matched only the verb-then-cue order, so Immediately buy TSLA, Now sell GOOGL and Please immediately short TSLA were accepted as research objectives. The mirrored rule requires the cue to sit directly before the verb, with only separators and an optional please between them. The first version allowed up to 40 characters there and also refused Now research whether to buy AAPL; the finishing commit (0244ecea) keeps that and similar research phrasings accepted. The check runs on a goal's objective, on each of its criteria and on an objective update.

  • The tail a full compaction keeps counts a thinking model's reasoning (#1582, thanks @Shizoqua) — the layer-3 tail is sized to a 20K-token budget, but the sizing read content and tool-call arguments and skipped reasoning_content, so on reasoning-heavy turns the kept tail could run far past its budget and compaction freed less than intended. The reasoning text is counted now, as the loop's token estimate already counted it.

  • technical_indicators results are restored after compaction instead of fetched again (#1601, thanks @zeus229) — the tool is read-only but repeatable, and a repeatable tool is replayed from the run's memory only if it opts in; it did not, so an identical request after compaction re-ran the market-data fetch. It opts in now, like web_reader, within the loop's existing replay budget.

  • Token counts and Polymarket CLOB token ids are no longer redacted as credentials (#1606, thanks @Shizoqua) — the token credential marker is a substring match, so token_budget, token_usage, total_input_tokens, total_output_tokens and clob_token_id read [redacted] in traces, event previews and the live audit ledger. Since 0244ecea each is released only in its expected shape (a non-negative integer, a map of counts, a decimal uint256 id); any other value under those names stays redacted.

  • A remote MCP tool with a malformed properties schema still gets a valid one (#1607, thanks @Shizoqua) — normalize_mcp_tool_schema dropped a non-object properties such as a string or a list and then skipped the {} fallback, so the tool reached the model as an object schema with no properties key at all. It gets {} now, unless the top level composes anyOf / oneOf / allOf.

  • The Copilot credential fallback re-reads the token after 60 seconds (#1619, closes #1617, thanks @lorenzozanee) — when neither COPILOT_GITHUB_TOKEN nor OPENAI_API_KEY is set, get_llm_credentials falls back to GH_TOKEN / GITHUB_TOKEN or gh auth token, and it cached that answer for the life of the process. The cache now expires after 60 seconds. Its consumer is provider diagnostics: Copilot requests go through the SDK client, which resolves credentials each time it is created.

Channels, memory and configuration

  • The Email channel verifies TLS certificates by default (#1544, thanks @shadowinlife). The implicit-SSL paths (IMAP4_SSL / SMTP_SSL) in both the connection probe and the polling/send adapter used Python's default unverified context (CERT_NONE, no hostname check), so the mailbox password travelled to a server whose certificate was never checked — an active network attacker between the operator and their mail provider could harvest it. All four call sites now build their context through a shared email_tls_context() that verifies the certificate and hostname against the system CA bundle, matching the STARTTLS path, which already did. A new verify_tls field (default true, surfaced in the guided Email setup in all nine locales) is the documented opt-out for self-signed or internal-CA servers. Upgrade note: a mail server with such a certificate connected before this change and now fails the TLS handshake until verify_tls is set to false for it.

  • Plain IMAP no longer sends the mailbox password in clear text (after #1544). With imap_use_ssl off, the Email channel and its connection test logged in over an unencrypted connection, and the guided setup put that switch one click away. A new imap_use_tls (on by default, the IMAP twin of smtp_use_tls) upgrades the connection with STARTTLS before LOGIN; a server that offers no STARTTLS is reported as a network failure before the password is sent. verify_tls now governs STARTTLS on IMAP and SMTP as well as implicit SSL, so a self-signed server can opt out on any path. Upgrade note: an imap_use_ssl: false setup against a server without STARTTLS stops connecting until imap_use_tls is set to false for it, which sends the password in plain text again.

  • Settings-write routes reject cross-site browser requests (#1544, thanks @shadowinlife). require_settings_write_auth did not apply the cross-site guard that its siblings require_auth and require_event_stream_auth enforce on unsafe methods, so a malicious web page could attempt a CSRF write against any settings route — including the new POST /channels/email/test, which merges stored credentials with a caller-supplied patch and could be steered to send the stored mailbox password to an attacker-chosen host. The guard now runs first on every settings-write route; same-origin Web UI calls and non-browser clients (CLI/curl, no Origin) are unaffected.

  • Channel adapters log through stdlib again: {} placeholders, napcat's 17 sites and loguru-only .opt() calls lost their messages (signal's _safe_handle raised instead of swallowing) (#1533, #1531, thanks @shadowinlife). Slack DMs from unapproved senders get a pairing code (#1524, thanks @Shizoqua); Discord no longer pings @everyone or roles (#1537); duplicate-reply suppression covers integer message ids and is bounded (#1476); the WebSocket channel starts again (from #1521).

  • Two memories sharing a title keep separate index rows and links (#1545, thanks @Shizoqua). Under VT_MEMORY_HIERARCHY an entry lives at {memory_type}/{slug}.md, so two entries with one title had the same path.name: the second add() overwrote the first's MEMORY.md row, and the semantic-link block beside it excluded the other entry as "self" and wrote an ambiguous target. Both key on the path relative to the memory dir now; a sidecar written earlier, with a bare filename or an absolute path, is still read.

  • /agent.json, /agent.yaml and /agent.yml are gitignored at the runtime root; the Feishu QR login already wrote channel credentials there.

  • Signal channel: stripping the bot's mention no longer removes the wrong character after an emoji (#1478, thanks @Shizoqua) — Signal reports a mention's start and length in UTF-16 code units, and the channel sliced the Python string with them directly. Each character outside the Basic Multilingual Plane (most emoji) before the mention shifted the slice by one, so the mention placeholder stayed in the text handed to the agent and the neighbouring character was cut instead. Offsets are converted to string indices before a mention is stripped or tested as a leading mention.

  • A file received over DingTalk is saved inside its download directory whatever it is named (#1506, thanks @Shizoqua) — the adapter joined the inbound fileName to the per-sender download path unsanitized, so a name with ../ segments, or an absolute path (which replaces the whole path when joined), could write outside it. The name goes through safe_filename now, as the other adapters already sanitize a remote file name.

  • A channel that fails to load logs which one and why (#1580, thanks @Shizoqua, thanks @Shizoqua) — two channel-registry log calls had two %s placeholders and one argument, so an external plugin that failed to import, or a built-in channel skipped for a missing SDK (at debug level), printed --- Logging error --- instead of its name and the exception.

  • The README says which process owns Telegram polling (#1583, closes #1553, thanks @lorenzozanee) — Telegram allows one getUpdates poller per bot token. vibe-trading serve owns the channel runtime, and channels start asks the running server (VIBE_TRADING_API_URL, default http://127.0.0.1:8000) to start it rather than polling itself, so the conflict to avoid is a second server process with the same token enabled.

  • Memory garbage collection no longer compresses an entry it just archived (#1450, thanks @Shizoqua) — the compression pass reused the entry list read before archiving, so an entry just moved to archive/ was compressed from a path that no longer existed and logged a failure. 451c554a skips by whether the file still exists, not by id: ids are six hex characters two files can share, and an archive step that failed leaves its file to be compressed.

  • An unparseable optional numeric setting falls back to its default instead of failing config load (#1592, thanks @Shizoqua) — the coercion that drops a bad int / float env value skipped int | None fields, so ANTHROPIC_MAX_TOKENS=not_a_number raised while building EnvConfig. Optional fields get the same fallback now; a parseable value outside the field's bounds (ANTHROPIC_MAX_TOKENS=0) is still rejected.

Web UI and locales

  • Leaving a chat cancels its pending history scrolls — a scroll scheduled while a conversation's history loaded could still fire after leaving it, and a history load still in flight could schedule another.

  • A welcome-screen quick action stays highlighted after you click it (#1500, thanks @shadowinlife) — the first chip was styled as selected by its position, whatever had been clicked, so the chip you chose never lit up. The highlight now follows the click and moves with the next one, and each chip reports aria-pressed.

  • Broker SDK diagnostics no longer name Longbridge for every broker (#1436, thanks @tingkk) — the shared sdk_missing and network_unreachable messages on the Runtime page and in the chat connector status said Longbridge whichever broker raised them, so an unreachable Futu OpenD gateway sent the user to check Longbridge. They now name the affected broker's SDK, service or local gateway, in every locale.

[0.1.15] — 2026-09-09

Rolls up 551 commits / 162 merged pull requests since 0.1.14, from 35 contributors.

The theme of this cycle is data that says what it is. A price frame now carries the adjustment caliber it was served under. A factor propagates the gaps in its inputs instead of filling them. A loader that cannot serve a market no longer claims it, and a source that fails to refresh is an error rather than a smaller portfolio. Three of those were the same bug wearing different clothes: a default that silently substitutes a plausible value for a missing one, which is indistinguishable downstream from a real observation. Three new markets, a fourteenth broker and fifteen Quant Library additions land alongside.

Added

  • UK equity market (#1206, thanks @cgycorey) — LSE .L and .IL symbols end to end: market data, Yahoo-backed financial statements and indicators, trade-journal inference into uk_equity, and its own engine path. SDRT is modelled as what it actually is — a 0.5% duty on the purchase side only, not a symmetric round-trip cost. Charging both sides overstated the cost of every round trip by half, which is exactly the size of edge a mean-reversion strategy lives on.
  • Zerodha Kite Connect (#1193, thanks @ashutoshsinghpr7) — a fourteenth broker connector for Indian equities. Kite exposes no runtime paper/live discriminator — no account-id format, host separation, demo flag or trade environment — so under the red line it is capped at paper plus read-only: place_order and cancel_order hard-refuse any non-paper config at the first line, and no *-live-trade profile exists to select. That now covers Longbridge, Dhan, Shoonya and Zerodha; Trading212 goes further and refuses all order placement including paper.
  • Read-only multi-broker portfolio (#1072, thanks @goatyyc; onboarding contracts in #1250) — one aggregated snapshot across every enabled connection instance, reachable four ways: the Web /portfolio page, REST under portfolio_routes.py, the portfolio_summary agent tool, and vibe-trading portfolio show | refresh | sources. Three design points that are not incidental: a source that fails to refresh is an error excluded from the totals (status error, last_success_at, complete=false), never a carried-forward cache, so a partial snapshot cannot read as a smaller portfolio; every remote_mcp read passes interactive_oauth=False, so aggregation can never pop an auth prompt; and analysis_context() carries risk_xray_args so portfolio_risk_xray is fed rather than reimplemented. Eligibility is one rule — is_portfolio_connection_profile requires readonly and account.read and positions.read — under which the IBKR official-MCP profile, which advertises only discovery, is correctly not a portfolio source.
  • Quant Library, fifteen additions (thanks @santhreal) — Heston (1993) stochastic volatility option pricing (#1195); Hierarchical Risk Parity allocation (#1196); Archimedean and Gaussian copula analytics (#1197); market-microstructure analytics — VPIN, Roll spread, Amihud illiquidity, Kyle's lambda (#1198); analytical single-barrier options with cash-rebate handling (#1163) plus finite-difference sensitivity Greeks for them (#1203); an ISDA standard-model single-name CDS valuation engine with hazard-rate conversions (#1164); Key Rate Duration decomposition across benchmark tenors (#1165); the Vasicek credit portfolio loss model and spread DV01 (#1167); cross-sectional factor Information Coefficient time series (#1166); Ornstein-Uhlenbeck exact calibration with half-life (#1161); drawdown distribution analytics with the Ulcer and Pain indices (#1162); Corrado rank and Cowan generalized sign non-parametric event-study tests (#1160); multi-factor risk decomposition with Euler marginal contributions, reporting unmatched_weight rather than normalising it away (#1159); and group-purged k-fold cross-validation for panel data (#1168). The layer's rule holds throughout: a formula has exactly one implementation, and skills import it rather than carrying the source in a SKILL.md code block.
  • Brazilian Portuguese locale (#1327, thanks @nandofmike) — the eighth shipped language, landing with full key parity. The parity suite globs locales/*.json instead of listing languages, so a new locale is covered the day its file lands; German had previously shipped with perfect parity and zero coverage because that list was hand-written.
  • Nobitex and Wallex (#1263, thanks @Emad211) — read-only Iranian exchange sources on keyless public UDF endpoints, explicit-source only. Both sit in _NO_NETWORK_FALLBACK_SOURCES and are barred from the crypto fallback chain on purpose: they are the only Toman-quoted sources and declare markets={"crypto"} purely to be reachable, so degrading an unavailable BTCIRT request into that chain would hand back a USDT-quoted series as if it were Toman.
  • Binance USD-M evidence path (#1229, #1230, thanks @honginp; #1248, thanks @lorenzozanee) — read-only USD-M account snapshots, drift evidence artifacts comparing simulated fills against recorded ones, and deterministic tolerance calibration derived from those recordings rather than assumed. The point is that the tolerance has a provenance: it is measured against real fills, not chosen.
  • Offline evals harness (#1271, thanks @AirHua-byte) — verifies a run's persisted artifacts (prompt preservation, tool calls and results, identity, evidence, provider usage, iteration budget, RunManifest integrity) against a versioned case/verdict schema, with no LLM, network, broker or MCP call: it reads files only. Absent instrumentation emits NOT_EVALUABLE rather than passing, which is the whole point — a missing field must not read as good behaviour.
  • Anthropic prompt caching (#1366, thanks @averatec0773) — an explicit cache breakpoint on the system block so the static tools+system prefix is read from cache instead of resent at full price, with cache usage recorded in the usage envelope. Native adapter only, behind VIBE_TRADING_ANTHROPIC_PROMPT_CACHE; set it to 0 if a compatible proxy rejects the cache_control request parameter.
  • Per-market data-source priority (#1231, thanks @sambazhu) — a Settings card and MARKET_DATA_ORDER_* env overrides that reorder a market's fallback chain. Validated as a multiset permutation of the default, so reordering passes while adding, dropping or duplicating a source is refused and the default stands. The override is applied as a setitem on the same dict object, because runner.py and market_data.py import FALLBACK_CHAINS by name and a rebind would leave them on the old order.
  • Calendar-triggered partial rebalancing (#1277, thanks @thisisjun786) and a data window separated from the evaluation window, so warm-up bars stop counting as evaluation and a strategy is not graded on the bars it needed to become defined.
  • Agent-confirmed scheduled research (#1187, thanks @AirHua-byte) and each monitor's latest verdict rendered on the job list (#1156, thanks @he-yufeng), so a recurring job's state is legible without opening it.
  • Swarm replay and retry (#1158, thanks @cgycorey; vibe-trading swarm retry --resume in #1194, thanks @SiMinus) — resume a failed or cancelled DAG run while keeping the artifacts of tasks that already completed, rather than paying for the whole graph again.
  • Order-plan rejection reporting (#1245, thanks @lorenzozanee) — an _on_plan_rejected hook plus a run-level report of the trades the engine wanted but could not take. A rejected plan previously vanished, so a strategy that was being silently throttled by buying power looked identical to one that simply had no signal.
  • Read-only Binance-connector crypto identity (#1242, thanks @pengpengyi92) and public venue-catalog pair resolution, so a crypto pair resolves without requiring a broker connection at all.
  • Multi-file drag-and-drop and paste uploads in the Web chat (#1179, thanks @AirHua-byte).

Fixed

Missing data stops being filled with plausible values

  • The Alpha Zoo NaN contract, enforced at the registry (#1377, closes #1376, thanks @cgycorey) — factors/base.py states the contract in its own header: NaN is preserved through warm-up and missing data, no silent fillna(0). An alpha that branches with np.where on a comparison violates it without looking like it does, because a NaN comparison is False and the ternary falls through to its constant. Blanking every declared input on one bar across the whole zoo, 84 of 462 alphas returned a number anyway — and since dropna() is the mechanism that keeps a gap out of an IC, each fabricated ±1 was consumed as a real signal. Registry.compute now masks the output to NaN wherever a declared dependency is missing on that bar, and the same sweep afterwards finds none. Individually fixed on the way there: gtja191_004 (#1371), gtja191_003 and gtja191_059 (#1372), gtja191_069 (#1373), gtja191_019 and gtja191_086 (#1379), alpha101_007 and alpha101_051 (#1380), alpha101_009 and alpha101_046 (#1381), alpha101_021 and alpha101_024 (#1382), alpha101_010/023/027 (#1383), and alpha101_049's warm-up mask (#1374) — thanks @Shizoqua. Stated rather than quietly carried: the warm-up half is still open. Measured the same way, 52 alphas still emit a value inside their own declared min_warmup_bars (gtja191_154 declares 198 and emits at row 0), because on those bars the inputs are present and only the rolling window is undefined. A registry-level warm-up mask would also blank rows for alphas whose declaration merely over-declares by two or three bars, so that half needs a direction call rather than 52 pull requests.
  • pct_change() forward-fills by default (#1397, #1398, #1399, #1400, thanks @Shizoqua; #1172) — pandas still defaults pct_change() to forward-filling, so a missing close becomes a 0.0% return that never happened, and the next bar reports a two-day move as one day. Four PRs each fixed one call site; sweeping the skills exhaustively found 24 sites across 10 files, two of them inside the very files those PRs were editing. The class is closed repo-wide now: the SDM strategy template (#1397), cross-market volatility weights (#1398), multi-factor TopN selection which now excludes assets with no factor observations rather than ranking them at zero (#1399), and sentiment-factor orthogonalization which no longer regresses on zeros it filled in itself (#1400). Alpha Zoo returns stopped forward-filling gaps in #1172.
  • Survivorship-bias disclosure reaches the report (#1289, thanks @bonyohana) — the flag was computed and then not rendered in the HTML bench report, which is the same failure mode one level up: a caveat that exists in the data and not in what the reader sees.

Prices say which caliber they are

  • Adjustment caliber stamped on served frames (#1317, closes #1301, thanks @he-yufeng) — a frame now carries what its prices actually mean. A caliber is recorded only where it was measured against a live payload or pinned by the loader's own endpoint choice; anything unmeasured resolves to unknown on purpose, because origin-side adjustment is invisible from loader code — Yahoo serves split-adjusted quotes with zero adjustment logic in this repo.
  • The loaders that still booked a dividend as a loss (#1287, #1288, thanks @he-yufeng; #1320, thanks @lorenzozanee) — Yahoo, FMP and Tiingo paths serve adjusted prices, and an unresolved symbol now retries down the fallback chain instead of stopping at the first source that lacks it.
  • Explicit sources answer as themselves (#1276, #1342, #1185, #1316, thanks @lorenzozanee, @he-yufeng, @cgycorey) — FMP moved to the Stable endpoint and no longer silently falls back when named explicitly; stooq's anti-bot challenge is flagged rather than parsed as data, which is the difference between "no data" and "a JavaScript page interpreted as prices"; codes, date range and source are validated before the fetch rather than after it; and symbol search is aligned with fetch for FX pairs and indexes, so the thing you searched is the thing you get.
  • Tencent history pagination (#1154, thanks @BigFishEmily) — long ranges were silently truncated to the first page.

Futures reach the right engine, and now have a source at all

  • Classification (#1369, #1389, thanks @Shizoqua; #1396, thanks @he-yufeng; and 349bf202 on the global side) — digit-prefixed and month-code-colliding product codes, Tushare's own exchange spellings (SHF/CZC/CFX/GFE), case folding at extraction rather than per lookup, and dated global contracts carrying their venue (CL2412.NYMEX, ESZ4.CME) each fell through every pattern to the a_share default. A US crude contract was therefore run under T+1, with no shorting, settled in CNY.
  • Coverage (#1395) — the loader chain was the last link, and it was empty. It named tushare and akshare, and neither implemented a futures endpoint: tushare._fetch_daily_frame branches on ETF / index / HK and sends everything else to daily(), the A-share equity endpoint, while akshare._fetch_one ends with # Default: try A-share. AKShare now serves Chinese contracts off the token-free Sina daily endpoints, dated and main-continuous, verified live across SHFE (RB2601, RB0), CFFEX (IF2512, T2512), GFEX (SI2601) and ZCE (MA2605). A global contract returns empty and reaches local rather than being priced as an equity. Trimming tushare's markets set alone would not have worked: resolve_loader walks the chain and never consults markets, so tushare would have stayed first in line. The main continuous contract (RB0) is routed too — a dated contract lives about one year, so any backtest longer than a contract cycle has to name the rolled series.
  • China-futures margin provenance (306aca2c, closes #1393) — a run now states which products were priced on a generic default instead of a table entry. rb's real margin rate is 0.10, the same number as the default, and "looked up" must not read like "assumed".

Live trading fails closed

  • Broker reads that are not answers (#1212, #1232, thanks @he-yufeng; #1244, #1254, thanks @cgycorey) — a position read returning an API error, and error envelopes arriving mid halt-sweep, were both treated as successful empty results. The sweep now latches only the episodes it actually swept, under per-episode latch files, and only after attempted side effects.
  • State that has to survive a restart (#1233, thanks @he-yufeng; #1213, #1221, #1222, thanks @Elfsa-Miranda) — the flatten latch persists across runner restarts, and unresolved Alpaca submissions are owned before broker writes and recovered by exact client ID, including fills, so a restart mid-submit cannot produce a duplicate or an orphan.
  • Order sizing and timing (#1312, #1361, #1253, thanks @he-yufeng and @cgycorey; #1209) — buy-limit orders are sized at the worse of quote and limit on both transports, market-triggered ticks are skipped while the market is closed, and Alpaca position quantity is signed by side before it reaches the mandate gate.
  • The mandate gate narrows in both directions (#1285, thanks @Jackzigen) — the change enforcing narrowing semantics for non-numeric and list adjustments also rejected "none" as an invalid type, which made the most conservative mandate unsubmittable. Nine CI checks were green over it: the eight new tests were all raises assertions, and a one-way suite cannot see that something which should have been allowed was refused.
  • Four connector gaps from the #1207 minor batch (#1388, thanks @he-yufeng) — an MT5 order above the symbol's volume cap was silently clamped and sent while the below-minimum side errored (it fails closed naming the cap now); an eToro limit order with no limit_price went out as an MIT with no TriggerRate and rested untriggered while reading as accepted; futu K-lines inherited an undeclared adjustment caliber; and bare BTCUSDT was routed through A-share rules — T+1 and no shorting, on a perpetual.
  • Report audit fails closed when nothing was verified (#1362, thanks @cgycorey) — an audit that checked zero claims returned PASS.

Backtest execution semantics

  • Options (#1299, #1305, thanks @cgycorey; #1306, thanks @he-yufeng; #1177) — signals fill on the next bar rather than on the date they were computed from, which had let a strategy trade at the price that generated its own signal; HV warm-up bars use the configured default IV instead of an undefined one; short option legs hold margin and gate opens on buying power, so a naked short could no longer sell premium it could never cover; and an explicit expiration is validated against the dates actually available rather than silently resolved to a neighbour.
  • Shorts, rebalancing and accounting (#1298, #1344, #1281, thanks @cgycorey) — 1x shorts liquidate on adverse moves in non-strict mode; an over-committed rebalance basket scales to fit instead of aborting on a commission overdraft; and requested target changes are separated from executed rebalance fills (closes #1275), so the rebalance count shown to a reader is the number of fills rather than the number of intentions.
  • Market rules through composite state (#1309, #1332, #1370, thanks @he-yufeng and @Shizoqua) — A-share and India rules are enforced through composite state rather than per-engine, a halted position marks at the last traded close instead of entry cost, and the sub-engine's active symbol is synced before every composite dispatch rather than once per bar.
  • Perpetuals and forex (#1307, thanks @he-yufeng; #1226, thanks @P1Piyush) — funding settles by bar span on 8h+ intervals rather than once per bar, and forex metals are treated as metals for pip size, lot size and spreads. A 50-ounce gold order had been rounded down to zero.
  • Symbol routing into the engines (#1351, thanks @cgycorey) — dotted US class shares (BRK.B.US, BF.B.US) fetched an empty chart.
  • Cross-market annualization (#1239, thanks @cgycorey) — risk x-ray honours bars_per_year=None instead of assuming a trading calendar that does not apply.
  • ML walk-forward label leakage (#1392, thanks @Shizoqua) — future labels were not purged from the training window in the example pipeline.
  • Intraday fundamentals lookahead (closes #1387) — an announcement date carries no time of day, so on an intraday frame a filing became visible from the first bar of its own announcement day: a full session of lookahead. Sub-daily frames are refused now, with fundamental_subdaily="next_day" as the explicit opt-in.

Shadow account

  • (#1217, #1310, #1311, #1314, #1356, thanks @he-yufeng) — PnL is derived from runner metrics and fails closed when unknown rather than reporting a number it cannot support; attribution is scoped to the pool currency with mutually exclusive buckets; short lots are modelled in FIFO pairing with legs restated across corporate actions; the result cache is keyed by window and journal hash rather than by shadow alone, so editing a journal no longer returns the previous answer; and cash-dividend journal rows are parsed and booked into real PnL.

Grounding gates

  • Both directions (#1326, #1338, #1375, #1378, thanks @cgycorey) — full-width brackets no longer split a clause; analysis metrics are gated on a completed analysis result (#1336); a metric claim formatted as a generic | Metric | Value | table no longer escapes the gate that rejects the same claim in prose; and the mirror-image error is fixed too — a price word used as a formula variable is not a price claim, so close/SMA50 > 1 is no longer read as an asserted price (#1354).
  • Vocabularies completed on both sides (#1346, #1357, thanks @saju01) — percentage-point deltas (3.6pp, 3.6 个百分点, 250 基点) were read as quoted prices, and the adjacent-CJK boundary is now exercised directly. The English gate had been the leaking side: \bclose\b does not match "closed", so a fabricated The stock closed at 412.35. passed while the identical Chinese claim was correctly rejected. Both vocabularies are complete now and a language-parity suite asserts the two reach the same verdict per case instead of checking each in isolation.
  • Identity (#1265, #1280, #1282, thanks @aminak58; #1384, thanks @laiyierjiangsu; #1269, thanks @AirHua-byte; #1333, thanks @Shizoqua) — bare joined crypto pairs (BTCUSDT) went unrecognised while spot platinum resolved to a crypto pair that does not exist; FX, metals and futures symbols route through the right identity path (GC=F and XAUUSD had been classified as China A-shares); search_symbol("XAUUSD") returned exactly one candidate — a Swedish Bitcoin ETP — which then locked the run's instrument identity, so a metal or FX query is an exact instrument assertion now; a pasted broker code (HK.00700, US.AAPL, SH.600519) scanned as no symbol at all; market context stated earlier in a conversation can narrow resolution behind VIBE_CONTEXTUAL_IDENTITY_CONSTRAINTS; and generic evidence timestamps are preserved rather than dropped.

The agent loop keeps what it still needs

  • (#1341, #1358, thanks @saju01; #1349, thanks @he-yufeng; #1352, thanks @cgycorey) — the loop discarded tool results the model was still using, and microcompaction cleared pairs it then could not reconcile. A session that had successfully fetched fundamentals, fund flow, margin and research data had those results cleared to free context, and the de-duplication ledger went on refusing to re-fetch them: 44 blocked retries in one run, ending in "fundamental data not retrieved" for data the model had already received. Clearing a result now re-opens that tool, the ledger keys on arguments rather than tool name, results are reconciled by exact call identity, and the token budget counts tool-call arguments instead of scoring a 100 KB payload as ten tokens.
  • The no-progress budget counts observations, not activity (#1363, thanks @be-student) — a run that keeps issuing tool calls without a new successful observation stops after NO_PROGRESS_LIMIT iterations with a visible recovery answer and a failed terminal state. An identical call identity is refused from its second failure, not its first: only a successful mutating call clears the ledger and a research run may have none, so refusing on the first failure made a transient rate limit or timeout permanent for the whole run.
  • Run-stall watchdog and bash tree-kill timeout (#1169, thanks @wiliao) — plus a compaction verification ledger, so a stalled run is detected rather than waited on.

Providers and transports

  • (#1225, thanks @he-yufeng; #1246, #1247, thanks @lorenzozanee; #1330, thanks @averatec0773; #1334, #1348, thanks @Shizoqua; #1182, thanks @AirHua-byte) — real token usage is requested on streamed calls instead of being estimated; a temperature rejection self-heals on the OpenAI-compatible path, and the Anthropic self-heal strips the relocated extra_body.temperature rather than only the original position; stream retry delays escalate and honour Retry-After; Codex token usage and response model metadata are propagated rather than discarded; and shared HTTP clients survive cleanup instead of being closed out from under an in-flight request.
  • Dependency caps that were load-bearing (#1313, thanks @he-yufeng; 4abd6cc2) — fastmcp is capped below 4.0.0 until the constructor/snapshot adaptation lands, and mcp below 1.30, which broke IBKR OAuth discovery. The second was diagnosed as a red CI on tests nothing in the diff had touched; adapting the fixtures would have hidden a genuinely broken broker path.

Web UI, CLI and surfaces

  • (#1257, #1258, #1259, thanks @iagop03; #1319, thanks @guestccc) — a stale streamingSessionId is cleared on session reopen, the main region scrolls and long run prompts collapse, a non-backtest run points at Studio instead of a blank dashboard, and the Vite dev server proxies /api so a fresh checkout works without a manual proxy step.
  • Session recovery across restart (#1180, thanks @AirHua-byte; #1340, thanks @averatec0773) — streamed assistant text is checkpointed to partial_response.json at ≥0.25s intervals with an atomic os.replace, and every pending or running attempt on disk is reconciled at service construction into an explicit interrupted transcript entry. The reply is committed before attempt.json so recovery can finish either side of that write. The attempt start persists too, so the elapsed clock survives navigation, reload and history.
  • Windows and desktop (#1261, thanks @Emad211; #1284, thanks @lorenzozanee; #1359, thanks @birdxs) — cross-platform locks and path handling, EPERM handled in the desktop updater's shutdown checks, and Docker actions moved to Node.js 24 compatible versions.
  • Swarm and MCP plumbing (#1175, #1331, #1176, #1335, thanks @Shizoqua; #1210, thanks @pengpengyi92; #1173; #1345, thanks @cgycorey) — cancel_run() is threaded into the in-flight worker rather than only marking the DAG; goal sessions are isolated per MCP connection rather than per process; goal_id/expected_goal_id default to the current goal; beginTime/endTime are forwarded through get_research_reports; worker retries use bounded backoff; the MCP stdio registry path has coverage; and investment-committee researchers get the fundamental data panel their prompts already assumed (#1343).
  • Portfolio OAuth lifecycle (#1211, #1251, thanks @goatyyc) — the reconnect lifecycle is isolated, and reconnect failures are classified rather than collapsed into one opaque error.
  • IBKR (#1186, thanks @sykuang; #1190, thanks @goatyyc; #1178, thanks @c020627) — public MCP OAuth flow support, verified read tools enabled, and the official read-only seed pointed at /mcp-public.
  • Futu HKD valuations (#1228, thanks @JaxonHu1024) and futu connection_state handling.
  • Feishu QR login credentials persist (#1188, thanks @AirHua-byte).
  • Persistent memory cleans up after itself (#1174, thanks @Shizoqua) — garbage collection and compression left the FTS index and the semantic links pointing at entries that no longer existed, so a cross-session recall could surface a row whose body had been collected.
  • Strategy store and discovery (#1347, thanks @ethanstoner; #1368, thanks @Shizoqua) — created_at ties are broken so history is really newest-first, and position size is resolved per regime rather than once for the whole run.
  • Valuation and credit refuse non-finite inputs (#1184, #1215, #1386, thanks @Robin1987China) — comps, three-statement inputs and every credit-risk function reject non-finite values rather than propagating them. The three Archimedean copula CDFs returned wrong extreme values under strong dependence — Clayton 0.0, Gumbel 1.0, Frank inf — and are computed in log space now, verified against a 2500-digit reference across θ ∈ [-5000, 5000] (closes #1385). Note that the first fix covered only θ > 0; θ < 0 overflowed to nan from |θ| ≈ 355 and the rewrite dipped below the Fréchet bound as θ → 0, both caught before merge.
  • Cross-validation edges (#1204, #1220, thanks @santhreal and @he-yufeng) — multi-segment test sets are supported, empty training partitions are guarded, and rows between combinatorial test blocks are pinned as still trainable.
  • Skill and documentation links (#1252, #1328, thanks @ethanstoner; #1189, thanks @youngjincho02-arch) — a skill-relative reference link resolves in read_file, the skill-name prefix is dropped so links resolve on GitHub, and the historical_var quantile formula in the risk-analysis skill is corrected. Verification note for the first two: the agent reads through ReadFileTool rooted at src/skills/, not through GitHub's renderer, so the two conventions had to be checked against the real consumer.
  • Dependencies (#1321, #1322, #1323, #1324) — dockerhub-description 4.0.2 → 5.0.0, docker/login-action 3.7.0 → 4.6.0, 15 npm minor/patch updates and 43 pip minor/patch updates, via Dependabot.

Changed

  • The futures fallback chain is now ["akshare", "local"]. resolve_loader walks FALLBACK_CHAINS and never consults a loader's markets set — the only .markets read is the explicit-source fallback — so trimming that set alone would have left tushare first in line, constructed and available on any TUSHARE_TOKEN, returning empty frames before akshare was ever asked.
  • tushare no longer declares the futures market. Serving it needs pro.fut_daily, which sits behind a points tier nothing here implements.
  • China futures main continuous contracts (RB0, IF0) now classify as futures rather than falling to the a_share default. The rule is generated from the product whitelist rather than from a width heuristic, so an ordinary ticker ending in 0 keeps its own market.
  • test_release_version_consistency.py did the job it was written for. The bump missed the app.version mock in frontend/src/components/layout/__tests__/Layout.test.tsx, and the guard named it — that site has been in its enumeration since it shipped with 0.1.14, which is the point of enumerating declaration sites rather than spot-checking the ones somebody remembers. The locale check globs its directory, so pt-BR.json was covered the day it landed without anyone extending the test. The version named in a test_cli_update.py comment is de-versioned here — a version string in a comment has no guard at all and goes stale at the next bump.
  • Reader-facing counts were re-measured against the code rather than incremented by hand: 74 MCP tools, 107 registry tools, 10 backtest engines, 90 bundled skills, 462 alphas, 27 data sources, 30 swarm presets, 14 broker connectors. None had drifted this cycle.

[0.1.14] — 2026-08-20

Rolls up 272 commits / 74 merged pull requests since 0.1.13.

Added

  • Options Lab (#1096, closes #1095, thanks @shadowinlife) — a Web UI page with four surfaces: expiry payoff diagram, spot×IV scenario P&L matrix, portfolio Greeks cards, and a live US options chain. The math is not new: two read-only HTTP endpoints wrap the existing options_payoff / get_options_chain tools, so the page and the MCP surface compute from the same test-pinned src/quantlib/options.py. Research only — it places no orders.
  • Factor Research tab in Run Detail (#1099, thanks @shadowinlife) — IC statistics cards, the daily IC series with its mean line, quantile-group equity curves, and an IC correlation heatmap, all rendered from the factor_analysis artifacts a run already writes. The new read-only GET /runs/{run_id}/factor endpoint scans the run's artifacts/ tree and computes the pairwise IC time-series correlation matrix, which existed nowhere before; a has_factor_artifacts flag gates the tab so runs without factor output do not show an empty panel.
  • Positions tab in Run Detail (#1097, closes #1098, thanks @shadowinlife) — portfolio-book structure from the existing positions.csv: a per-symbol weight pie/treemap with a date slider, sector and asset-class net-exposure bars, and a weight-evolution stacked area. The pie is gross composition (absolute weights including shorts and cash) while the bars are net exposure per industry, so a long/short pair in one sector nets to zero on the bars while both legs stay visible on the pie. The two charts answer different questions on purpose.
  • Tearsheet tab in Run Detail (#1091, closes #1090, thanks @shadowinlife) — monthly-returns heatmap (years × months with a full-year column), annual returns bar chart, top-5 drawdown table, and an equity curve annotated with the ranked drawdown zones. Everything is derived client-side from the equity.csv rows the run response already carries: no backend change, no new dependency, and month axis labels come from Intl.DateTimeFormat rather than per-locale month keys.
  • Interactive backtest research dashboard (#1084, thanks @AndyLongest) — an optional report-style view for a completed backtest alongside the existing candlestick and trade-detail views: headline return/risk/trading KPIs, normalized equity versus benchmark, drawdown with rolling Sharpe, realized trade P&L, a trade ledger, and the complete metric table. CLI backtests get a quiet local report-server bootstrap so Full Report routes to the dashboard without a manual server step.
  • Strategy Discovery (#978 Phase 1 and #1007 Phase 2, refs #969, thanks @shadowinlife) — a read-only facade answering "what strategies exist and what state are they in?" across the Alpha Zoo and the SDM strategy store, with per-(strategy, regime) evidence rows rather than scenario tags. Ships list_strategies / query_strategies / get_strategy_evidence as agent and MCP tools behind an evidence gate (minimum-trade, coverage and cost-breakeven thresholds, each with an explicit warning), plus a startup guard that drops the routing text when any advertised tool is not actually registered. Phase 2 adds the population path — refresh_strategy_evidence as an agent tool, an MCP wrapper and vibe-trading strategy-evidence refresh --manifest — with the backtest-diagnose Hard-Gate Checklist as the ingestion gate (stable hard-gate:* skip tokens) and atomic rebuilds. Freshness is computed at read time from the evidence-window age (fresh/aging/stale); stale rows fail closed out of default recommendations behind an include_stale opt-out, and sdm:* rows mirror the SDM lifecycle.
  • Scheduled research now delivers itself. A finished briefing leaves the run through an outbox rather than waiting to be read: the outbox row is claimed under a lease before sending so a crash mid-delivery cannot double post, the send is wired to both the session and the channel runtimes, and the push happens on the terminal event. Delivery is configured and watched from Market Watch.
  • Each monitor's latest verdict is persisted on its job (#1152, refs #943, thanks @he-yufeng) — the run's terminal briefing is parsed server-side into a structured last_verdict record and the /scheduled-runs list carries it inline, so the list renders a verdict without re-parsing free text per row or going N+1 on sessions. The five research playbooks grow a strictly additive ## Verdict tail — one - SYMBOL: STATE - reason line per tracked symbol, each playbook declaring its own state vocabulary. Parsing is deliberately strict: an ad-hoc prompt lands permanently at no_verdict_section and renders as nothing rather than a warning, a run that tracked nothing is a real "no calls" answer rather than an absence, and a malformed section degrades to contract_violation instead of showing a wrong verdict. The prior record is embedded one level deep at write time and no deeper.
  • Futu connector — seven extended read-only endpoints (#1135, thanks @549236606-oss) closing the gap between what the SDK exposes and what the connector consumed: get_rehab (dividend/split/rights adjustment factors, so backtests stop reading dividend gaps as price moves), get_capital_flow (historical super/big/mid/small inflow buckets), get_capital_distribution (today's in-flow versus out-flow snapshot), get_history_deals (fill records for true cost-basis reconstruction, capped at the SDK's 360-day window), get_acc_cash_flow, get_financials (income/balance/cash-flow statements), and get_earnings_calendar with EPS and revenue consensus. Each routes through the same _quote_ctx / _trade_ctx / _assert_gateway envelope as the original five, so missing data, missing privileges and a missing OpenD all degrade to a clean fail-closed payload instead of an SDK stack trace. The futu-api SDK stays lazily loaded.
  • Vietnam equities (HOSE) as a backtest market (#1033, thanks @ngoanpv). .VN matched no entry in _MARKET_PATTERNS, so _detect_market fell back to a_share and executed Vietnamese bars under China A-share rules — and asking for Yahoo explicitly did not help, because both Yahoo loaders gated on suffix and rejected .VN before building the request. .VN now resolves through the Yahoo loaders with a ["yahoo", "yfinance", "local"] fallback chain and runs on a dedicated VietnamEquityEngine.
  • Offline USD-M account reconciliation (#1106, relates #1030, thanks @honginp) — immutable Binance USD-M account and position snapshot contracts that compare an exchange observation against the existing AccountState and RiskSnapshot without mutating either, reporting numeric, structural, missing-symbol and unexpected-symbol drift deterministically. Liquidation-engine validation is explicitly left unassessed rather than silently assumed. This is the Shadow 1 slice: contracts and fixture-only offline reconciliation.
  • Novita AI as a built-in OpenAI-compatible provider (#1059, thanks @jax-novita) — registered in the provider registry with NOVITA_API_KEY / NOVITA_BASE_URL and wired through capabilities, CLI onboarding, .env.example and the README provider list, so it is selectable from the built-in list instead of hand-configured as a custom endpoint.
  • GitHub Copilot as a provider through the official github-copilot-sdk (#990, supersedes #899, thanks @sykuang), with SDK-managed authentication from COPILOT_GITHUB_TOKEN, GH_TOKEN, GITHUB_TOKEN, stored Copilot CLI credentials or gh credentials, plus LangChain-compatible invoke, streaming, reasoning and tool-call handling and Settings/preflight integration. This implements the boundary #899 was closed on: no borrowed OAuth client ID and no editor-impersonation headers.
  • tickerall hosted MetaTrader 5 data source (#968, closes #897, thanks @miguelangelo78) — a broker's MetaTrader 5 candle feed over the hosted TickerAll HTTP API, so forex and metals backtests run on any OS with no local, logged-in MetaTrader 5 terminal. Purely opt-in: is_available() is False unless TICKERALL_API_KEY and TICKERALL_ACCOUNT_ID are set, and the loader never joins an automatic chain.
  • A drift tolerance band for rebalance mode — a rebalance that would move the book by less than the band is skipped rather than executed for a rounding difference.
  • Spanish and German locales (#1087 thanks @daviddaco1, #1117 thanks @1psconstructor). Spanish ships a full key-for-key es.json plus README_es.md, making Spanish the sixth README; German ships de.json with the full UI key set. Both register in SUPPORTED_LANGUAGES and localeLoaders on the existing lazy-loading path, and the locale-parity and interpolation-variable tests cover them.
  • Desktop update safety boundary (#1101, refs #1016, thanks @QCYTSN) — a strict PID-scoped backend/watchdog shutdown result for a future update handoff, dormant Windows candidate verification, interrupted-attempt recovery primitives, and a documented, tested rejection matrix for tampered, unsigned, invalid-signature, wrong-publisher and downgraded candidates. The signed 0.3.0 → 0.3.1 run itself remains blocked on a signing identity; this is the part that could be made executable before a certificate exists, including the proof that desktop cleanup never kills unrelated Python processes.
  • Docker images carry the Feishu and Telegram channel dependencies (#1088, thanks @birdxs), with a manually triggered workflow that builds and pushes to both GHCR and Docker Hub and syncs the project description to the Docker Hub page.
  • Strict alpha t-stats in the bench JSON and HTML report (#1085, thanks @jay79-boop), so the strict-mode number is readable from the artifact instead of only the console.
  • Offset paging on the MCP surface for SEC filings and financial statements (#1138), and load_skill routed through the registry so an oversized skill pages instead of being truncated at the transport limit (#1137).

Changed

  • The MCP surface grows to 74 tools (from 70 in 0.1.13).
  • smartmoneyconcepts is now an opt-in [smc] extra, not a base dependency — and the stale <3.14 cap is gone from both the project metadata and the distribution manifest. The package pulled in numba → llvmlite, and llvmlite ships no macOS x86_64 wheel from 0.46 onward, so as a base dependency it turned every Intel-Mac install into a source build that needs CMake (discussion #1035). It backs exactly one skill example, whose SKILL.md already tells the reader to install it. Intel-Mac users on Python 3.11–3.13 can now pip install vibe-trading-ai without a toolchain.
  • CI runs the test job on both the floor and the ceiling of requires-python, so a version-specific break at either end is caught before release rather than by a user.
  • Vite moves from 6.4.3 to 8.2.0 in the frontend dev toolchain (#1022).
  • Market-data provenance now declares per-market volume units (#1065, #1131, thanks @shadowinlife), and get_market_data carries that provenance over MCP rather than dropping it at the transport boundary.
  • The Ollama runtime URL is normalized at source (#1074) instead of being patched at each call site.

Fixed

  • A run's grounding gate stops refusing answers it can support. The identity/price recovery path now re-fetches missing evidence within a bounded budget instead of rejecting outright (#1092, thanks @Shizoqua); identity constants in rate formulas are no longer read as unsupported quotes (#1083, thanks @AndyLongest); line-leading ordered-list markers are masked before number extraction (#1063, thanks @zzz607); ISO dates that run straight into CJK text are recognised as dates (#1132, thanks @Robin1987China); a report-style date cell parses as a date and a US CSV stem resolves; a dash-form trading day reads as a date and both bounds of a ranged level are masked; an order line is read as an instruction rather than an observed quote; and an "@" level is masked only when a quantity or an order label accompanies it, so a bare price is still held to the evidence standard.
  • Agent run reliability (#1105, thanks @wiliao) — grounding false-rejections, the final-answer gate and LLM timeouts, together with prompt wording and support/resistance masking (#1060).
  • build_registry() no longer returns a partial registry in silence (#1129, thanks @er-s-an): a construction failure is reported instead of leaving the caller with a short tool list that looks complete.
  • Backtest correctness: excess_return stays consistent with the corrected benchmark_return (#1058, thanks @Shizoqua) and the corrected benchmark fields are rounded to the metrics contract; the engine reports actual post-fill positions rather than the requested ones (#1082, thanks @AndyLongest); hold mode says so when it drops a requested resize; the archive no longer mixes two runs' artifacts into one bundle; and importing the runner module no longer loads .env as a side effect.
  • Quantlib numerics: xirr and money-weighted return survive long-horizon discount underflow instead of crashing (#1119, thanks @pengpengyi92); DCF refuses non-finite inputs rather than returning a silently negative share price (#1121, thanks @Robin1987China); a zero-volatility option discounts its forward value (#1066), the fixed-income curve keeps decay inside the requested bounds (#1076), event studies anchor to the prior session (#1078), and cross-validation aligns label ends to the prior observation (#1079) — all thanks @pengpengyi92. technical_indicator RSI uses Wilder EWM smoothing (#1056, thanks @Shizoqua).
  • Swarm: worker prompts are ordered for prompt-cache-friendly prefixes (#1057, thanks @Echoandelementwebsites); worker artifacts are isolated between retry attempts (#1053) and a path-shaped agent id is rejected before the retry rmtree; raw ok/success tool-result envelopes are rejected (#1052) and oversized tool results truncate with the shared notice (#1110) — thanks @Shizoqua; tool-less agents are no longer instructed to call write_file (#1144, thanks @Echoandelementwebsites); and the per-task ChatLLM is closed to stop a pooled-connection leak (#1145, thanks @cgycorey), with the same treatment for the one-shot clients in auto-title and image vision (#1153).
  • Connectors and market data: baostock volume is normalized to board lots behind a cross-source consistency guard (#1067, thanks @shadowinlife); the IBKR market-data tier is selectable and starved quotes report as no_data (#1075, thanks @jay79-boop), with the requested and the applied tier reported separately; eToro gains runtime UI parity for SDK connector status (#1051) and fixes crypto browse and flat market-data quotes (#1070, thanks @ofeksh-tr); the tencent loader builds its SSL context from the certifi CA bundle, since the new GlobalSign Atlas R3 root is absent from Python's default store and HK quotes failed every retry (#1113, thanks @x-lambda); and the East Money research-report endpoint gets the time parameters it now requires (#1077, thanks @zzz607).
  • CLI: connector orders renders broker_sdk rows — the renderer only knew the nested IBKR shape, so a flat row left every column but Account empty — and stringified SDK enums print as BUY rather than OrderSide.BUY, while class-B tickers such as BRK.B and decimal values are left intact (#1150, thanks @nstavros); direct SDK account diagnostics render (#1073); the show subcommand dispatches its run_id instead of the --show flag (#1147, thanks @cgycorey); and a Docker Codex OAuth EOF explains itself (#1054, thanks @zhiwuyazhe-fjr).
  • Providers: reasoning effort is honoured in chat completions (#1025, thanks @cgycorey) and passed through to the Anthropic adapter (#1115, thanks @straun-repo).
  • A remote MCP failure reports the server's own explanation instead of a bare status code, so a rejected token reads as a rejected token (refs #1126).
  • Scheduled research: an in-flight delivery is no longer overwritten (#1140, thanks @Shizoqua), and the scheduled-runs DELETE returns an empty Response for its 204 (#1068, thanks @ofeksh-tr).
  • Tools and agent: unsupported ticker-plus-name symbol queries are marked skipped rather than failed (#1114) and recovery steering is delivered as user messages with inline system tags (#1112) — thanks @lorenzozanee; prediction-market fields reject non-finite values and the envelope stays strict JSON (#1136, thanks @Shizoqua); gross_profit is derived from revenue minus COGS when the SEC tag is absent (#1111, thanks @cgycorey); every compacted message folds through the summarizer; and the autopilot preserves backtest validation evidence when linking hypotheses (#1139, thanks @Shizoqua).
  • Onboarding: .env.partial is created with owner-only permissions (#1086, thanks @lukiod) and written atomically, so a failed save cannot destroy recovery state.
  • Memory: the fallback tokenizer's minimum is aligned with the FTS5 sanitizer (#1071, thanks @Shizoqua).
  • Research reports reject a reversed window instead of reporting it as missing coverage.
  • Web: inferred strategy labels are marked as inferred in the run dashboard (#1134, part 1 of #1094, thanks @fixXxerTech).
  • The test suite no longer escapes its sandbox into the real config root (#1118, closes #1116, thanks @lorenzozanee). A full run had been appending synthetic order_rejected records to the live, hash-chained audit ledger at ~/.vibe-trading/live/audit.jsonl and writing a live/audit_chain.jsonl holding a single NUL byte. It was reported on Windows, but the escape was not Windows-specific — a default macOS environment leaked the same way. The config root is now sandboxed before collection, through one knob at the tail of the resolution chain rather than a per-test override, and os patches are confined to the module under test. hash locks are portable across supported platforms (#1102); Windows desktop packaging inputs are stabilized (#1104); the Docker build workflow is pinned and the image stays on hash-locked dependencies; the alpha-bench SP500 panel carries the GICS sector through; a frontend test waits on the sector re-render rather than on the spinner disappearing; and the Spanish locale says "demo", not "papel", for paper trading.

[0.1.13] — 2026-08-10

Rolls up 408 commits / 162 merged pull requests since 0.1.12.

Added

  • src/quantlib — a tested finance-math layer (265 functions across 19 modules): one implementation each of Black-Scholes price/greeks/IV inversion, bond math + Nelson-Siegel/Svensson, Altman Z + Merton/KMV, stationarity/cointegration/ GARCH/bootstrap, VaR/CVaR/EVT, Brinson-Fachler attribution, event studies, multiple-testing control, and purged cross-validation. Skills now import these instead of carrying formulas inside markdown code blocks. Reachable from the CLI, the Web UI, the REST API and MCP through the read-only, pure-compute quantlib_call tool, so the finance math works where bash is gated off — module allowlist, __all__-only, export_* refused. attribution and impact were allowlisted but carried no __all__, so the tool listed zero functions for them; both now export, and a guard test fails when any allowlisted module exposes nothing. Econometrics needs the stats extra (statsmodels/arch), which the functions lazy-import and name.
  • Valuation engine (src/quantlib/valuation/): run_dcf (FCFF bridge, WACC build, dual terminal value that cross-checks each method against the other's implied multiple and implied g, mid-year discounting, net-debt bridge, WACC×g grid), run_comps (EV bridge, LTM + calendar-year calendarisation, multiple matrix — a peer with a non-positive denominator is excluded and reported, never averaged in as a negative multiple), a linked three-statement projection with a hard balance assertion, explicit revolver plug and an iterated interest↔debt circularity that must converge or raise, plus input-hashed versioned artifacts with xlsx/pptx export. One shared rule in contracts.py: a missing input makes a model NOT RUNNABLE and is never silently defaulted.
  • Typed entity + irregular dated cash-flow substrate (src/entities/) — the non-bar ingestion path for NAVs, capital calls and coupons, deliberately parallel to the bar engines so a nav column can never reach one and get priced as a close. Surfaced by the read-only cashflow_performance tool (XIRR / MOIC / DPI / TVPI / TWR / Modified Dietz / MWR over an irregular dated series).
  • orderbook_depth (read-only): crypto L2 ladder via ccxt — spread bps, depth imbalance, and the impact cost of a stated notional walked through the real book in both directions. fully_filled and timestamp_source must be read before quoting anything; equities still have no depth source.
  • Governance wired into every run (src/governance/): a run manifest hashes the prompt, skill contents, tool registry and package versions so "what methodology produced that number" is answerable, and the audit ledger is hash-chained + fsynced so editing or deleting a record is detectable (an edit that recomputes its own hash is caught one record later via prev_hash_mismatch). All 30 swarm presets were re-audited — a deliverable no granted tool can compute is now declared as such instead of invented.
  • Four read-only data tools, all on free public sources: get_institutional_holdings (SEC 13F-HR; manager / ticker-holders / top-managers modes with quarter-over-quarter position diffs, and cover-page unit detection because pre-2023 filings report in thousands), etf_holdings (cross-market look-through — SEC N-PORT for US, and for A-shares the semi-annual/annual reports that carry the full book rather than the quarterly top ten: 510300 returns 342 rows / 98.66% of net assets vs 10 rows / 22.74%; coverage distinguishes full_portfolio from top_n_disclosed and every response stamps the report period), prediction_market (event-contract search/event/market/history with prices surfaced as labelled implied probability), and research_papers (arXiv + OpenAlex search/read with source-anchored claim extraction that marks what a source does not state instead of inferring it).
  • Institutional research surface: six slash commands — /comps /dcf /attrib /memo /earnings /screen — each carrying a step skeleton and an arithmetic-consistent worked example (the Brinson decomposition sums exactly to active return; the earnings bridge sums exactly to the EPS delta).
  • Investor lenses as a standalone skill (src/skills/investor-lenses/): named-investor reasoning frameworks as stackable analysis overlays, decoupled from the data layer. Each lens is an operating procedure — priority signals, disqualifying conditions, typical misuse — not a biography, and names no tool.
  • Five ready-to-schedule research playbooks (premarket brief, earnings-season tracker, portfolio checkup, A-share money flow, institutional-holdings diff), reachable three ways: auth-gated REST routes beside the existing CRUD, a vibe-trading playbook CLI subcommand, and a /playbook slash command. Templates state their data needs in natural language rather than naming tools, so coverage can grow without editing them, and every one mandates naming a missing input instead of filling it from memory.
  • Desktop shell — a source-first Electron host that owns the existing backend lifecycle (random loopback port, per-launch secret, five-locale startup recovery, owned-process cleanup) (#923). Windows packaging assembles a checksum-pinned embedded Python 3.12 runtime with x64 NSIS review/signing paths, plus Electron safeStorage for an allowlisted credential set — the renderer can set or clear secrets but never read them, plaintext config migrates once, decrypted values reach only the owned backend, and both unsigned-review and signed builds fail closed on the wrong signature state (#1015). No installer artifact was published from that PR.
  • Canadian equities end to end (#1024, #1019, #1037, closes #952): .TO/.V symbols classified in CAD, routed Yahoo → yfinance → local, executed under Canada-specific GlobalEquity rules, benchmarked against XIC.TO, with mixed-currency aggregation refused.
  • Korea equity (KRX: KOSPI/KOSDAQ) as the 9th backtest engine (#693, thanks @JungHoonGhae) — execution-time ±30% band on a unified tick grid, structurally long-only, config-driven 2026 0.20% securities transaction tax, optional pykrx loader that rejects non-daily intervals so the runner falls through.
  • eToro connector with path-separated demo/real profiles (#989); live risk-increasing actions remain mandate-gated and audited.
  • OpenBB Workspace bridge (#817, thanks @shugaoye) — optional openbb extra registering /agents.json + an authenticated /v1/query, each request replaying the supplied history in an ephemeral session. pro.openbb.co is not a default CORS origin; opt in via VIBE_TRADING_EXTRA_CORS_ORIGINS.
  • Read-only Taiwan snapshot tool (#848, thanks @TSENGCHIENFENG), registered only when VIBE_TW_STOCK_DB points at a schema-valid snapshot.
  • Options strategy analytics (#946, rebuilt from #883, thanks @he-yufeng) — analytic expiry P&L extrema, exact breakevens including continuous zero-P&L intervals, engine-aligned entry commissions, and spot × IV scenarios through Agent and MCP.
  • Read-only sentiment tool (#939, thanks @Robin1987China): local lexicon scoring for arbitrary text plus the crypto Fear & Greed Index from Alternative.me's free, no-auth API.
  • Read-only technical_indicators tool (#921, refs #920, thanks @Robin1987China) — RSI/MACD/Bollinger/SMA/EMA through the existing loaders.
  • Timezone-aware scheduled research (#954, closes #953, thanks @ngoanpv): jobs take an optional IANA timezone and evaluate cron on that zone's wall clock, so a cadence survives DST — a spring-forward gap is skipped and a fall-back ambiguous time runs once. Cron fields gain comma lists and ranges (1,3-5); jobs without a timezone keep UTC semantics; the web UI gains a Scheduled page in all five locales.
  • vibe-trading update (#1020) — checks PyPI for a newer release and upgrades in place, installing the exact release it checked and verifying fresh metadata without downgrading. Editable/source checkouts are detected and answered with the right manual instruction (git pull + reinstall) instead of a pip upgrade that would silently replace the dev install.
  • ModelScope joins the built-in providers through its official OpenAI-compatible hosted-inference endpoint, default Qwen/Qwen3.5-27B (#1011, thanks @yrk111222).
  • Live model discovery in Settings (#924, thanks @QCYTSN) — configured providers are queried on demand with stable warning codes and five-locale controls, and each reply records and reloads the immutable provider/model/ reasoning identity that actually served it.
  • alpha_zoo + bounded alpha_bench on MCP (#979, thanks @cgycorey) with horizon/result/output-path limits and safe report creation; QVeris discovery/inspect/execute also join the MCP surface (#976, closes #964, thanks @shadowinlife), with the cost quote read from the marketplace instead of trusted from the caller.
  • The MCP surface grows to 70 tools. Six read-only analytics tools that had reached the agent but never MCP are now mirrored onto it: quantlib_call, cashflow_performance, orderbook_depth, sentiment, technical_indicators and get_fundamentals. Order-placing tools stay structurally un-exposed — the mirrored registration path refuses any class whose is_readonly is not True, so an order tool cannot reach MCP even if someone adds it to the source list.
  • Memory Tier 2: structural organization (#815, thanks @shadowinlife) — four independently-gated modules (H-MEM hierarchical routing VT_MEMORY_HIERARCHY, A-MEM BM25 semantic linking VT_MEMORY_LINKS, TF-IDF auto-compression VT_MEMORY_COMPRESSION, FTS5 index with CJK bigram tokenization VT_MEMORY_FTS_INDEX), a one-line VT_MEMORY=off|on|full preset replacing seven individual flags, dependency-conflict warnings, lazy FTS5 rebuild on first search, and 9 integration + 23 benchmark tests. All off by default (#733, closes #732).
  • SDM lifecycle tools sdm_register / sdm_status / sdm_decay_scan (#457, thanks @shadowinlife) backed by a SQLite artifact store (UNIQUE(name, universe), WAL) and an IC/Sharpe decay state machine (active → monitoring → decayed → disabled).
  • alpha bench --strict (#796, closes #773, thanks @he-yufeng) finally wires the strict same-universe random-control + OOS gate that shipped unreachable since 0.1.9; alpha bench also discloses CSI300/SP500 sources, counts, degraded fallbacks and survivorship bias (#859, closes #845), with the _meta disclosure forwarded to CLI/HTML (#841, closes #797, thanks @AmirF194).
  • Risk x-ray artifacts (risk_xray.json/.md) emitted by every portfolio backtest with headline concentration/vol/drawdown metrics (#900, thanks @he-yufeng), plus rebalance-notes artifacts and turnover metrics (#795).
  • Composable optimizer weight constraints (#818, thanks @he-yufeng); calc_metrics reports tracking error and benchmark beta.
  • Opt-in atomic same-direction rebalancing in backtests with immutable fill evidence (#951), and opt-in position_adjustment=rebalance for strict USD-M historical backtests preserving collateral, funding, fees, realized P&L, liquidation behavior and fill evidence across increases and reductions (#1019, thanks @honginp).
  • USD-M perpetual realism: perpetual_strict settles historical funding before fills and executes isolated/cross margin breaches as real liquidations (#903, #889, thanks @honginp), margin state contracts (#798), historical funding rates actually consumed rather than fetched-and-ignored (#819, thanks @g0rdonL), and ordered fill/funding/risk/liquidation event persistence plus a fidelity summary (#936).
  • Final report aggregator stage for the quant_strategy_desk swarm preset (#1048, thanks @Robin1987China).

Changed

  • Rebuilt WebUI — the guided-minimalism overhaul: no first-frame flash, one durable activity object per turn with a live reasoning whisper and a reload-safe tool trail, LLM-written session titles, full five-locale parity.
  • Sessions, runs, swarm runs and uploads now live under ~/.vibe-trading (relocatable via VIBE_TRADING_HOME) with a one-time automatic migration (#925, closes #904, thanks @MuggleJinx).
  • The frontend moves to Node 22 + React Router 8, clearing a high-severity advisory; Electron transitive advisories were patched separately.
  • Daily price bands are judged at execution time — from a pre-fill base price against the prospective fill price — never from the decision bar's own close.
  • A session runs one attempt at a time (a second concurrent send is refused with HTTP 409), and a user stop is its own terminal state distinct from a failure, both live and on reload (#676, thanks @tyj147454413-cmd).
  • Trace redaction is now sink-aware: content is released only in the tool-RESULT sink and stays redacted in the fail-closed ARGUMENTS sink used by tool-call arguments and the live audit ledger (env is never released); results also get their string leaves pattern-scrubbed (#675, #913, #911, thanks @santhreal). Trace records are fsynced and sidecars made durable before the record that references them (#662).
  • Oversized tool results are paged by whole record with an explicit total instead of being cut mid-JSON; the system prompt drops redundant tool prose.
  • Dependency locks refreshed and verified across Python and the frontend (#949, #948, #850–#852, #1021, #1023, #1026, #1027); the breaking MCP 2.0 bump remains unmerged pending a complete lock/runtime migration (#950).

Fixed

  • The identity/grounding gate stops refusing answers it has the evidence for. This was the single most user-visible defect in the 0.1.12 line — a well-formed run would spend minutes on real tool calls and then return "cannot safely confirm instrument identity or price evidence". Root causes, each fixed and covered by two-sided guard tests:
    • .SS and .SH were treated as different instruments, so every Shanghai ticker was permanently ambiguous. Symbols are now canonicalized (.SS≡.SH, sh600519≡600519.SH, 700.HK≡00700.HK, BTC/USDT≡BTC-USDT) before comparison, and candidates are collapsed by canonical symbol before the ambiguity test.
    • A failed side query could demote an already-locked identity; aggregate status now keeps a lock unless something genuinely conflicts.
    • Yahoo returns HTTP 400 for every CJK query, which was recorded as a source failure and escalated to blocking invalidated. Non-ASCII queries are now declined up front with an explicit skip marker, and a resolution with clean sources and no failures resolves to not_found rather than invalidated. All resolver skip markers were unified on one prefix, with a cross-module contract test.
    • A hardcoded per-tool whitelist decided which bare tickers could be matched, blocking 11 of the 17 documented argument spellings; matching is now by uniqueness against the authorized set.
    • Chinese-language answers were rejected for writing 雅虎/腾讯 instead of the ASCII loader name, and 元 instead of 人民币. Source and currency aliases now cover the user's language.
    • Thousands separators split a clause mid-number, so ¥1,309.22 was compared as 1 against the observed range and reported as a price conflict.
    • Conceptual questions with no instrument at all, and comparison reports, were dead-ended by the gate.
    • Follows the narrower 0.1.12-line fixes for numbers that were never prices — confidence scores, indicator readings, moving-average windows, year-less dates like 8/5, percentage ranges, and a plan's own trigger levels (close ≥ 6.45 is a condition, not a quote) (#1001, #983, #955) — and a many-candidate shortlist now counts as an answer rather than a stalled resolution. A quote outside recorded OHLC evidence is still refused.
    • Ordering, evidence and serialization repairs from the same family: a canonical symbol/venue must be locked before a market-sensitive tool-call batch starts so a resolver and its consumer cannot race (#887/#886); tool envelopes with ok: false / success: false are no longer recorded as successful calls; NaN/inf volume and close no longer crash format_grounding_block (#1043, thanks @santhreal); the latest close is paired with its correct date; and bash-written run-dir OHLC CSVs count as observed evidence (#1037, thanks @wiliao).
  • Sandbox gap closed: generated strategy code can no longer import the broker layer, nor reach socket/subprocess/os.system/ctypes through a renamed binding — both were accepted before. src.quantlib still imports.
  • SEC reporting periods are keyed on their (start, end) span. A 10-Q files the true quarter and the year-to-date frame under the same end date and fiscal period, so period="annual" had been returning a single quarter for AAPL FY2018–2020 (a 4.2× understatement) and every fiscal-Q4 slot in a quarterly series carried the full-year figure. get_fundamentals("AAPL.US") no longer answers ok:true with an all-null panel. PIT fundamentals also dedup restated rows and stop the snapshot regressing to an older fiscal period on a late restatement (#772, closes #771, thanks @klmtseng).
  • Tushare A-share prices are corporate-action adjusted in both the factor bench and backtests — a raw close-to-close return across an ex-date was off by up to 47 percentage points (300750.SZ, 2023-04-26) — and the CSI300 bench masks each date to its point-in-time index membership.
  • Cross-market composite backtests refuse a mixed-currency code set instead of summing CNY, USD and KRW into one equity curve; Shadow splits mixed markets by settlement currency without invented FX aggregation (#997).
  • Option legs are marked at the volatility they were opened at, removing a fabricated day-zero P&L of up to +93% of premium; options partial-close honors the requested quantity instead of flattening the lot.
  • bar_returns no longer erases the real move across a trading halt longer than the forward-fill window — the resumption move was silently recorded as 0, understating volatility and inflating Sharpe — and an inf prior price can no longer read as a clean −100% (#895, thanks @darkknight4563). Buy-and-hold returns are sign-safe (#872), the rolling correlation matrix no longer forward-fills missing closes (#873, thanks @ddy4633), indicators use consecutive unsampled history (#1005), negative-equity drawdown and empty insolvent cross accounts are handled (#958, #959), and inf from pct_change on zero equity is replaced (#1041, thanks @santhreal).
  • Annualisation now covers all 24 data sources at every interval, with a coverage test that fails CI when a loader lands without entries (#891, closes #884, thanks @Robin1987China). A 19-PR interval-normalization sweep accepts lowercase 1h/4h/1d/1w everywhere, fails fast on unsupported intervals instead of silently returning daily bars, maps Yahoo 4H→1h, and keeps 1H/4H as hour bars across the Tiger/Alpaca/OKX/Shoonya/Longbridge connectors (#812–#838, #778–#794, thanks @santhreal); lowercase 4h returns true four-hour bars (#1013).
  • Market-data routing: partial results complete the missing symbols through the fallback chain and fail closed instead of silently shrinking the backtest universe (#689, closes #681, thanks @xkam7ar); HK fallback routing repaired with a new Tencent HK source (#1000); yfinance crypto routes to the crypto engine (#970); OKX bars use history-candles with rate-limit retry (#644); a Binance loader joins the crypto fallback chain (#643).
  • Providers: Claude models that deprecate temperature (opus-4-7, opus-5, sonnet-5) work — the adapter drops the field when the API rejects it, retries once, and remembers the model, so no per-release patch is needed (#890, closes #856, thanks @yagnikpipaliya). The whole Kimi K-series auto-forces temperature=1 (#701, thanks @sambazhu); endpoint resolution falls back to each provider's canonical base URL; mapping-shaped Responses stream events are accepted (#1034, thanks @cgycorey); OpenAI Codex OAuth gets a separate synchronized credential store and one-shot 401 recovery (#1014); proxy opt-out covers sync and async clients (#995); long model slugs stay readable (#1006).
  • MCP: dataclass results no longer crash on a false Circular reference detected (#849, #922, thanks @Echoandelementwebsites); factor_analysis is realigned to the registered tool's real CSV contract so calls no longer die on KeyError (#715, closes #635, thanks @Robin1987China); list/dict arguments tolerate JSON-string clients (#993); the network guard accepts IPv6 and case-variant hosts (#750); the specs cache key includes remote server identity (#1049, thanks @Shizoqua); nested results serialize cleanly.
  • Swarm: ok:false / success:false tool failures are detected, not just status:error (#1028, thanks @Shizoqua); MCP discovery is cached (#704).
  • Scheduled research isolates malformed records (#1003), fixes interval-timezone validation (#1004), and retries transient dispatch failures with capped exponential backoff.
  • Memory: entries are written and recovered with their .md suffix (#984); FTS5 ranking applies importance decay (#1032, thanks @Shizoqua); exact index-anchor matching and a respected result bound (#956, #957, thanks @santhreal).
  • Two quantlib modules were allowlisted but unreachable. attribution (Brinson-Fachler) and impact (market-impact models) carried no __all__, and quantlib_call dispatches on __all__ alone, so the tool listed zero functions for both while the package docstring advertised them. Both now export, and a guard test fails when any allowlisted module exposes nothing — the previous test only asserted each module imports, which could never catch this.
  • The published MCP manifest under-reported the server. SKILL.md's tool count was derived by counting @mcp.tool decorators, which ignores every tool registered through the mirrored path — so four institutional-research tools were live over MCP and absent from the manifest, and the contract test asserted that absence was correct. Both tests now measure mcp.list_tools().
  • Resource leaks and robustness: the HTTP throttle sweeps stale bucket entries interval-aware, preserving rate-limit spacing (#1047), the rate limiter no longer grows unboundedly with unique client IPs (#1039), the event bus notifies and removes subscribers on clear (#1046), _json_loads is guarded against corrupted JSON in database columns (#1045), and Shoonya tolerates empty-string numeric fields (#1038) — all thanks @santhreal.
  • Connectors: the connector CLI loads ~/.vibe-trading/.env so env-sourced broker credentials resolve again (#902, closes #901, thanks @MuggleJinx); IBKR moves to a thread-local connection pool with snapshot quotes, fixing hangs under parallel agent runs (#636, thanks @MikeCer); MT5 trading_history coerces numpy scalars so JSON serialization no longer dies on int64 (#776, closes #774, thanks @shadowinlife); CLI balances show a real account label (#843, closes #846).
  • Docker's hash-locked install works again with a new CI lock check (#858, closes #847), and saving Agent LLM settings from the Web UI no longer returns HTTP 500 on Windows — the POSIX-only os.fchmod hardening is platform-guarded (#561, thanks @CRui5in).
  • Non-interactive vibe-trading run injects a host session id: research-goal tools previously failed on every call while the run still reported success (#885). The agent now stops when evidence is sufficient (#1010).
  • The vn.py export skill is repaired for the vn.py 4.x layout, where vnpy.app.cta_strategy no longer exists upstream — templates import from vnpy_ctastrategy (#869, thanks @y85998607).
  • An encoding and parsing batch: UTF-16 BOM decoding in the document reader and trade-journal CSVs, currency symbols stripped before numeric coercion, BTCUSDT-style symbols inferred as crypto, CJK characters preserved in skill directory slugs, Eastmoney/Futu/Tonghuashun Excel-serial dates normalized, blank/NaN symbol rows skipped, and skill frontmatter parsed at EOF (#862–#868, #811, #749, #861, thanks @santhreal and @Robin1987China).
  • QQ replies retain source message IDs (#1008); Feishu/CLI markdown table edges and indent-preserving channel message splits are fixed (#867).
  • The daily-bar validator can opt in to non-positive prices — opening on negative bars while still rejecting zero (#816, closes #571, thanks @darkknight4563).
  • Sandboxed runs retain their canonical root (#1012, #1017); persisted run state is fsync-durable (#645); Portfolio Studio artifacts are surfaced in run detail (#980, #982, #966, #973).

[0.1.12] — 2026-07-22

Added

  • User swarm-presets directory: preset YAMLs dropped into ~/.vibe-trading/swarm/presets/ are discovered alongside the bundled roster (same-name files override it — the same rule as user skills) and survive pip install -U. list_presets() entries now carry a source: "user" | "bundled" field; explicitly named user presets run through run_swarm(preset_name=...), while keyword auto-routing stays limited to the curated table. Preset names are validated to a single path segment before any filesystem lookup.
  • Security hardening: all 10 findings from the 2026-07-10 external audit closed (#476, tracking discussion #468) — Docker multi-stage rebuild with digest-pinned base images, AST-hardened backtest sandbox (blocks network/subprocess/eval/os.environ/unsafe-open reachable from generated code, including inside nested function bodies), short-lived single-use SSE auth tickets replacing a long-lived key in the URL/logs, hardened Compose (read_only, dropped capabilities, no-new-privileges, resource limits), auth + rate limiting on /correlation, security headers (CSP Report-Only, X-Content-Type-Options, Permissions-Policy), /live + /ready health split, hash-locked dependencies wired into the Docker build, GitHub Actions pinned by commit SHA, and an HMAC-authenticated factor cache.
  • Opt-in TAP mode for Alpaca (#377, thanks @0xZKnw) — routes all broker egress through a self-hosted TAP proxy so the agent process never holds the raw API key, with writes blocking on human approval.
  • Realized portfolio turnover (avg_turnover / total_turnover) surfaced in backtest metrics for every optimizer (#478, thanks @Robin1987China).
  • Frazzini-Pedersen betting-against-beta academic factor (#480, thanks @YogeshModi24) — Alpha Zoo: 460 → 461.
  • MetaTrader 5 connector + data source (Exness-style MT5 brokers, Windows-only pip install "vibe-trading-ai[mt5]"). Broker connectors: 11 → 12 — full read surface plus order placement against a locally running terminal, with a bidirectional identity guard (paper profile ⇔ demo trade_mode, login pinned, contest rejected), connector-level max_order_volume/max_order_notional_usd guards on demo AND live, and hedging-safe position close by ticket. The live mandate gate gains forex/cfd instrument vocabulary (schema v1 unchanged) and a lot-aware quantity_notional_usd sizing hook so USD caps bind on lot-sized orders (0.1 lot EURUSD ≈ $10,800, never 0.1 × quote). Market-data sources: 20 → 21 — the mt5 loader heads the forex fallback chain (broker-exact symbols with Exness suffix discovery, 1m–1D bars), get_market_data learns forex/metal symbol routing (EUR/USD, XAUUSD.FX previously fell through to tushare), and akshare's forex path accepts the canonical slash form so degradation off-Windows keeps working.
  • Strategy Development Manager skill (#457, thanks @shadowinlife, closes #455) — sdm_register / sdm_status / sdm_decay_scan turn academic papers and broker research into registered factors/strategies with a persistent SQLite artifact store (UNIQUE(name, universe)) and automated IC/Sharpe decay monitoring driving an active → monitoring → decayed → disabled lifecycle. Pluggable OCR for read_document (local RapidOCR by default; cloud Qwen-VL is explicit opt-in only via VIBE_TRADING_OCR_ENGINE=qwen-vl, never auto-selected). Skills: 86 → 87.
  • Requesty as an OpenAI-compatible LLM gateway provider (#474, thanks @Thibaultjaigu) — same provider/model naming and capability shape as OpenRouter, wired through CLI onboarding, provider menu, and Settings.
  • Binance USD-M perpetual routing, slice 1 of #462 (#470, thanks @honginp) — explicit BTC-USDT-PERP symbol contract with execution/mark price separation, fail-closed when the two aren't timestamp-synchronized.
  • Correlation regime timeline (#756, thanks @ebujinovch, closes #719) — a new additive GET /correlation/regime endpoint plus an opt-in "Regime timeline" strip on the Correlation tab: rolling pairwise correlations reduce to an edge-density series, causally smoothed, and run through a two-threshold hysteresis state machine that marks FUSED episodes ("when did the market fuse into one bloc?"). Descriptive risk context, not a trading signal; shares /correlation's auth + rate-limit budget. Backed by the correlation-regime skill (#557, thanks @ebujinovch).
  • Three new LLM providers — SiliconFlow CN + Global (#565, thanks @UNHNQ), iFlytek Spark (#537, thanks @FenjuFu), and a native Anthropic Messages API adapter (#695, thanks @jelech; pip install "vibe-trading-ai[anthropic]"). MiniMax now exposes its regional API endpoints (#731, thanks @octo-patch).
  • Historical USD-M funding settlements for Binance perpetuals (#716, thanks @honginp); maintenance brackets are supplied as a validated, versioned artifact rather than a live authenticated fetch (#757, thanks @honginp), so a plain -PERP backtest stays zero-credential.
  • Pluggable OCR for read_document with optional LLM-vision extraction and a configurable text-density threshold (#548, thanks @shadowinlife) — local RapidOCR by default; cloud engines are explicit opt-in, never auto-selected.
  • New academic factor academic_corr_rewire (#705, thanks @ebujinovch) and the fundamental zoo wired into the _VALID_ZOOS whitelist (#707, thanks @sambazhu). Binance crypto fallback loader (#643) and bounded OKX history fetches with rate-limit handling (#644, thanks @tyj147454413-cmd).
  • QVeris premium-track hardening — session budget applied to backtest data calls (#685) and atomic credit accounting (#686, thanks @xkam7ar).

Changed

  • Correlation tab accepts bare tickers like AAPL,SPY and walks the full loader fallback chain instead of failing with Fetched: [] (#472, thanks @yxhuang, closes #471).
  • local loader honors the requested interval via OHLCV resampling instead of silently returning daily bars (#467, thanks @Shizoqua).
  • Provider credentials are resolved through one centralized path, fixing a gateway misroute (#563, thanks @shadowinlife, closes #549/#553). When no *_BASE_URL is set, the backend now falls back to the provider catalog's canonical default_base_url (the same default Web Settings already used), so a CLI / manual-.env user reaches the right endpoint instead of defaulting to api.openai.com.
  • Signal alignment is vectorized for an ~80× speedup on wide panels (#698, thanks @shadowinlife); swarm workers cache MCP tool-discovery specs to avoid redundant RPC round-trips (#704, thanks @shadowinlife).

Fixed

  • Explicit source: local backtests now route US/HK equities to the global-equity engine instead of the crypto default, and explicit benchmarks are fetched through the configured source's loader — local fails closed (no yfinance fallback) so offline runs stay offline (#550).
  • Loading .env now invalidates an EnvConfig singleton cached during early CLI imports, so the welcome panel, /settings, and dotenv diagnostic report the configured provider and model consistently (#541).
  • FastMCP transport imports work across both module layouts (#469, thanks @roberttidball).
  • Portfolio optimizers no longer include the decision bar's close-to-close return in weights executed at that bar's open (#487, thanks @YZY0108).
  • Backtest turnover metrics now use actual filled and rounded position sizes; targets rejected by market rules no longer inflate reported turnover.
  • End-of-backtest liquidations now apply exit slippage and include their commission in the final reported equity.
  • Open-price rebalances no longer use the decision bar's close for sizing or depend on whether a replacement symbol sorts before the position it closes.
  • Preflight (vibe-trading run) no longer resolves provider/model against a stale EnvConfig snapshot cached before dotenv loads (#479, thanks @ananaymital, closes #477).
  • Switching providers no longer leaves a stale OPENAI_BASE_URL from a previous configuration silently overriding the newly-resolved endpoint (#484, thanks @Bortlesboat, closes #482).
  • Strict-JSON / finite-number hardening across the backtest + tools stack (thanks @santhreal): risk ratios stay finite when equity crosses zero mid-path (#765) or annualizes an explosive path (#739/#740); scalar backtest metrics (#766), factor IC std (#767), and pattern trend-slope (#764) emit strict RFC-8259 JSON (null, never bare NaN/Infinity); Black-Scholes helpers treat non-positive spot/strike as intrinsic (#744).
  • Loader / data correctness — yahoo 1m stays minute bars instead of being uppercased to monthly (#761, @santhreal); the composite engine falls back to the first available sub-engine for unknown symbols (#734, thanks @Marnie0415); mootdx history that doesn't reach the requested start is rejected (#692, thanks @xkam7ar).
  • Session / journal robustness — one corrupt session.json (#762) or a schema-bad messages.jsonl line (#763) no longer aborts listing/reading; Excel float-stringified A-share codes (#770), unicode-dash PDF page ranges (#769), and export KEY= dotenv lines (#768) all parse correctly (@santhreal).
  • Native zai provider on glm-5.1 (#758) — endpoints that stream zero chunks fall back to a non-streaming invoke instead of raising, and an HTML error page surfaces an actionable base-URL hint.
  • Partial market-data results are completed through the loader fallback chain instead of silently shrinking the universe (#689, closes #681).
  • Cancellation is honored before the first AgentLoop iteration (#641, thanks @xkam7ar, closes #638); streaming output no longer triggers an insertBefore DOM race in the frontend (#717, thanks @Marnie0415); codex stream HTTP failures are classified for correct retry (#663, thanks @tyj147454413-cmd).
  • Robinhood connector account_number wiring and remote-MCP display shape (#726, thanks @nareshkps).
  • Broad reliability fixes across packaging, web, scheduler, swarm, and CLI (#584, thanks @xkam7ar).

[0.1.11] — 2026-07-11

Added

  • Indian equity (NSE/BSE) as a first-class market (#305, thanks @muku314115). A dedicated IndiaEquityEngine — T+1 delivery, no overnight shorts (opt-in intraday), configurable circuit bands, 1-share lots, and a config-driven STT / stamp-duty / exchange / SEBI / GST cost stack — with .NS/.BO symbol routing (yahoo → yfinance → india_broker → local), an opt-in read-only Shoonya/Dhan india_broker data bridge, and 255 alpha101/qlib158 factors opted into the new equity_in universe. Backtest engines: 7 → 8; market-data sources: 19 → 20.
  • Fundamental factor layer, Phase 1. PIT-safe SEC fundamentals flow into dense daily fund:* factor panels — filed-date anchoring, first-filed restatement policy, true-quarter (start, end) frame selection with Q4 synthesis (so YTD/annual frames can't contaminate TTM), and rolling TTM — plus a get_fundamentals tool and 4 quality/value factors in a new fundamental zoo family. Alpha Zoo: 456 → 460 across 5 families.
  • Research Autopilot Phase 3 — the loop closes (#267, thanks @Robin1987China). scaffold_signal_engine writes a contract-correct signal engine from a hypothesis and link_autopilot_backtest runs it, completing hypothesis → signal-engine → backtest end to end.
  • 4 canonical academic alphas (#277, thanks @Robin1987China) — Jegadeesh short-term reversal, George–Hwang 52-week high, Amihud illiquidity, and Harvey–Siddique co-skewness join the academic family (452 → 456), with a central OHLC-invariant guard at the runner fetch boundary dropping malformed bars from every loader (#274, thanks @Shizoqua).
  • Scheduled research runs end to end (#278 closing #254, thanks @mvanhorn). A default-off background executor (VIBE_TRADING_ENABLE_SCHEDULER) fires due interval/cron jobs through the session runtime, on top of a crash-safe atomic job store, 3 auth-gated /scheduled-runs REST routes, a Reports library, and post-backtest attribution. Route test coverage followed in #452 (thanks @Robin1987China).
  • IM channel runtime — research delivery over 16 adapters. The agent session runtime now attaches to 16 built-in message adapters (WebSocket, Telegram, Slack, Discord, Matrix, WhatsApp, Signal, QQ/NapCat, WeChat/WeCom, Feishu, DingTalk, email, MS Teams, MoChat), dependency-gated with install hints, configurable via AgentConfig.channels, and surfaced in REST (/channels/*), CLI (vibe-trading channels ...), and Web Settings — in all 5 UI locales.
  • QVeris optional premium data track. The 19 free sources stay the default; an explicit-only QVeris mode (Settings → QVeris or vibe-trading data mode paid) unlocks 63+ providers behind 3 key-gated tools (qveris_search / qveris_inspect / qveris_execute) with preview-by-default and a session budget gate. Never enters auto-fallback.
  • Trading 212 read-only connector (#321, thanks @mvanhorn) — 11 brokers total. Trading 212 exposes no runtime paper/live discriminator, so the connector is fully read-only: place_order/cancel_order hard-refuse every order, paper included. Live order guards also gained an opt-in, broker-agnostic PreTradeAdvisoryInterface that records advisory reviews without bypassing the mandate gate.
  • Turnover-aware portfolio optimizer (#466, thanks @Robin1987China) — fifth optimizer: mean-variance utility with an L1 penalty on weight changes versus the previous rebalance (SLSQP, long-only simplex), so the portfolio only trades when expected improvement outweighs churn. Optimizers: 4 → 5.
  • analyze_image vision tool (#464, thanks @fei-moss) — send a local chart / K-line screenshot / app screenshot to the session model as a multimodal message and get a semantic read (complements read_document's OCR). Path-validated against the allowed file roots; requires a vision-capable model. Tools: 71 → 72 free-mode (75 with QVeris).
  • Value-investing toolkit (thanks @sambazhu): financial-rigor + report-audit tools, 4 skills, and a value_investing_committee swarm preset. Swarm presets: 29 → 30.
  • CN-friendly search fallbacks — web_search gains China-reachable backends in the ordered no-key engine chain.
  • Provider roster additions: Kimi for Coding as a distinct provider (#435, thanks @yxhuang), opencode provider mappings (#444, thanks @imsankz), Codex OAuth default model bumped to openai-codex/gpt-5.4 (#446, thanks @morluto).
  • SKILL.md manifest guard test (#461, thanks @asahikiko) — the packaged skill's capability counts (skills / presets / zoo / sources / MCP tools / engines) are now asserted against source, so distribution paperwork can't silently drift again.

Changed

  • api_server.py modularization completed — 1,103 → 371 lines (#424 closing #331, thanks @shadowinlife), after route slices for channels, settings, and the remaining route groups, plus a shared compat layer with session-service writeback fixes.
  • Centralized environment variable management (#440 closing #438, thanks @shadowinlife) — every env var flows through a single Pydantic EnvConfig schema, enforced by an AST-based CI gate that rejects raw os.getenv/os.environ outside agent/src/config/.
  • Factor engine acceleration — hot rolling operators use bottleneck/NumPy fast paths, and alpha-bench parallelism stops resending large panel payloads to workers.
  • Robinhood Agentic MCP refresh — current MCP tool names across generic reads, live-runner plumbing, default read-only seeds, and mandate-gate tests; interactive OAuth holds the handshake open through multi-minute broker sign-ins (VIBE_LIVE_AUTHORIZE_TIMEOUT_SECONDS).
  • Loader fetch() signatures now match the loader protocol across OKX / Tushare / yfinance (#437, thanks @shadowinlife).
  • Timezone-aware UTC timestamps across session, goal, channel, and API paths (#397, thanks @mustafakamal88).
  • Inbound IM media now lands under ~/.vibe-trading/uploads/<channel>/ (fixes #465, thanks @fei-moss for the report) — inside the default allowed file roots, so the agent can read what users send over IM channels with zero configuration. The Matrix E2E store moves to the runtime dir (legacy path honored) so credentials never enter the readable root.

Fixed

  • Docker/server startup crash when FastAPI route iteration hit an included-router entry without path (#450, thanks @Penn-Live).
  • GLM thinking models on the zhipu provider no longer lose their reasoning stream (#458).
  • Trading mandate UX guards — a second-confirmation dialog before committing a real mandate, unified error toasts, and clarified inputs (#453, thanks @wison1717-maker).
  • trading_place_order treats zero quantity/notional as unset instead of passing a zero-size order to the broker (#417, thanks @irfanallana-oss).
  • Longbridge Decimal values serialize as floats across quotes, bars, balances, positions, orders, and executions (#459, thanks @fanfpy).
  • NapCat private messages now trigger pairing codes (#463, thanks @fei-moss).
  • Backtest validation artifacts: validation.json no longer requires a pre-existing artifacts dir (#429, thanks @isaveall), and nested NaN/Infinity values are normalized before writing, so strict JSON parsers don't choke.
  • CLI: resume preserves the first user message (#448 closing #447, thanks @morluto); --swarm-run rejects extra tokens with a clear error (#428, thanks @isaveall); the interactive CLI prints the session-id on exit with a copy-paste resume hint.
  • Shadow Account: extracted rules carry RSI / prior-return entry bounds computed from PIT-safe context fetched through the loader registry, so generated engines enter on real conditions (#302/#314/#316, thanks @Robin1987China); tushare ETF/index/HK symbol routing fixed along the way.
  • Content-filter resilience — event-driven and swarm runs skip individual LLM content-moderation hits and warn when filter rates are high; Gemini safety finish-reasons recognized.
  • IM channel reply timeout is configurable (#413, thanks @dpersek).
  • Provider preflight no longer follows redirects (#404 closing #402, thanks @dpersek).
  • Windows baseline green — vibe-trading setup/dev handle Windows TypeScript builds, correct cwd, the Vite 5899 port, and child-process shutdown; mootdx batch pulls let KeyboardInterrupt/SystemExit propagate.
  • Security hardening — loopback API CSRF protection (cross-site POSTs can no longer drive side effects on the local API), SSRF guards on interactive fetch paths, tightened API/Docker/frontend dev defaults, and cleared frontend dependency/CSP alerts.
  • Reverted the IRR-AGL reliability/governance stack (#405/#416) after it broke session chats on day 1 (#433 — thanks @yxhuang for the precise diagnosis); the evidence-bound research-pipeline direction continues in reviewable slices on #442.

[0.1.10] — 2026-06-19

Roll-up release; see the v0.1.10 release notes for the full narrative.

Added

  • Global data layer — market-data sources 10 → 18 (direct-API Eastmoney / Sina / Stooq / Yahoo + key-gated Finnhub / Alpha Vantage / Tiingo / FMP) with ban-risk-ordered fallback chains behind a shared throttled HTTP gate, plus 18 read-only data tools (fund flow, dragon-tiger, northbound, margin, block trades, shareholder count, lockup, sector, research, news, SEC filings, financial statements, options chains, institutional holdings, screening, symbol search, FRED macro, iwencai) — all MCP-exposed.
  • 10 broker SDK connectors — Tiger / Longbridge / Alpaca / OKX / Binance / Futu / Dhan / Shoonya join IBKR (local read-only) and Robinhood (Agentic MCP); direct-SDK live orders pass the fail-closed bounded-autonomy gate; brokers without a runtime paper/live discriminator are structurally capped at paper + read-only.
  • Alpha Zoo alpha compare across CLI, REST, Web UI, and agent tool.
  • Research Autopilot Phase 1 (run_research_autopilot, generate_backtest_config) and the local Data Bridge loader.
  • Opt-in local data cache (VIBE_TRADING_DATA_CACHE) for settled bars under ~/.vibe-trading/cache/.
  • Per-run token usage (llm_usage.json) + progressive Run Detail charts; CLI resume <session-id>.

Changed

  • Provider reliability overhaul — DeepSeek hang fixes, Kimi access, streaming liveness watchdog, Gemini 3.x multi-turn tool-calling fix.
  • Swarm workers pull market data through the loader layer; live swarm status cards stream in the chat timeline.
  • Baseline install slimmed (pyharmonics/ta behind vibe-trading-ai[harmonic]).

Fixed

  • Community security-hardening wave (#241–#258): Settings write auth, shell tools opt-in, LAN-access 403 clarity, Docker-to-host Ollama URL rewrite, web_search multi-engine fallback, and more.

[0.1.9] — 2026-06-01

Added

  • Connector-first broker profiles (IBKR + Robinhood). Trading access now starts from a selectable connector profile rather than separate broker/live entry points; vibe-trading connector list/use/check/account/positions/orders/quote/history and the MCP trading_* tools share the selected profile, with paper/live as a property under the connector. IBKR is usable immediately as a local read-only TWS / IB Gateway profile; the official IBKR remote MCP path is seeded as an OAuth mcp.read probe until stable read tool names ship. Robinhood Agentic Trading is a bounded connector behind OAuth, a committed mandate, an order guard, an audit ledger, and an instant halt switch.
  • Research Goal runtime. Long-running, research-only goals with auditable checklist criteria, budgets, and a /goal CLI command, plus REST + MCP endpoints (start_research_goal, get_research_goal, add_goal_evidence, update_research_goal_status) and a Web GoalDrawer.
  • Swarm retry_run. Re-launch a failed/stale/cancelled run with the original preset + variables; exposed as both POST /swarm/runs/{id}/retry and an MCP retry_run tool (the list_runs → retry loop). 36 MCP tools now.
  • Operator-configured external MCP tools in swarm workers (#142) and remote MCP transports for the built-in agent.
  • mootdx A-share OHLCV loader — native 通达信 TCP, no token, sits between tushare and akshare in the fallback chain. CCXT loader now reads proxy env for restricted networks (#126).
  • Hypothesis Registry CLI — list / show / invalidate.
  • Strict alpha-bench mode with a mandatory random control (#143).

Changed

  • CLI split into the agent/cli/ package (from a 3216-LOC single file), with a refreshed interactive terminal UI (figlet banner + activity rail) and a single cli/_version.py version source.
  • Swarm status reconciles from live task files on every read; run_swarm sends MCP progress heartbeats, and the stale-run reaper uses per-run thresholds (#132).
  • Refreshed provider default model ids; bumped langgraph for CVE-2026-28277.

Fixed

  • --version no longer drifts (#156). The version derives from package metadata, falling back to reading pyproject.toml directly — no hardcoded constant left to forget on release.
  • Session running-status indicator now survives reconnect / page reload / sidebar navigation; swarm DAG blocks downstream tasks when an upstream task fails (#145).
  • Robustness pass: pre-flight validation for LLM-generated signal engines with clean JSON errors (#149), graceful agent-loop exit at the iteration budget instead of an output-less failed (#148), flush + fsync session message writes that skip corrupted JSONL lines on read (#147), and IME Enter handling in the Web composer (#146).
  • Full Report link now always renders when a runId exists, even cross-browser (#150); SSE idle timeout is configurable via VIBE_TRADING_SSE_TIMEOUT (#157); cross-market correlation normalizes timestamps so crypto-vs-equity pairs align (#158).

[0.1.8] — 2026-05-17

Added — Alpha Zoo (450+ pre-built quant alphas)

  • agent/src/factors/ — base operators (rank, scale, ts_*, delta, decay_linear, signed_power, safe_div, market-aware vwap) and a registry that AST-extracts metadata from each alpha module without importing it. Lookahead is enforced at the operator level (delta(d>=1)), and registry sanity checks reject +/-inf and outputs that are more than 95 % NaN.
  • 4 zoos shipping 452 alphas total:
    • qlib158 (154 alphas) — port of Microsoft Qlib's Alpha158 feature handler under Apache-2.0, with pinned commit SHA per file.
    • alpha101 (101 alphas) — implementation of Kakushadze (2015) "101 Formulaic Alphas" (arXiv:1601.00991), written from the paper appendix; the relevant trademarked string is intentionally absent.
    • gtja191 (191 alphas) — implementation of Guotai Junan's 2014 "191 Short-period Trading Alpha Factors" research report.
    • academic (6 factors) — Fama-French 5 + Carhart momentum, shipped as honest price-based proxies (not the canonical FF series).
  • vibe-trading alpha {list,show,bench,compare,export-manifest} CLI subcommand. show and export-manifest enforce path-traversal guards.
  • New agent tools: AlphaZooTool (browse) and AlphaBenchTool (orchestrator with Jinja2 autoescape + strict CSP HTML report).
  • ZooSignalEngine.from_zoo(...) — composite multi-factor signal engine with cross-sectional standardisation, weighting, and optional top-N / bottom-N long-short conversion.
  • wiki/scripts/build_alpha_library.py — Alpha Library renderer. Reads manifest.json produced by vibe-trading alpha export-manifest and emits 452 per-alpha HTML pages plus 4 per-zoo overviews, each with script-src 'none' CSP. The landing page hydrates per-zoo counts from content/index.json.
  • New blog post: "Which of the 191 GTJA alphas still work in 2026?" with aggregate IC statistics, theme breakdown, and the top alphas that survive eight years of out-of-sample data.

Added — Web UI for Alpha Zoo

  • New page at /alpha-zoo in the Vite + React frontend with three views: browse (4 zoo cards + filter bar + paginated table), detail (formula, metadata, collapsible source code), and bench-runner (form → SSE-streamed progress + Alive/Reversed/Dead stat cards + Top-5-by-IR table + by-theme breakdown chart). "Alpha Zoo" nav entry added to the layout.
  • Four new REST routes in the FastAPI server:
    • GET /alpha/list — filterable alpha catalogue
    • GET /alpha/{alpha_id} — meta + source code
    • POST /alpha/bench — kicks off a background bench job and returns a job_id
    • GET /alpha/bench/{job_id}/stream — Server-Sent Events with progress, result, done, and error event types. In-memory job state with a 1-hour TTL; no Redis/Celery dependency.
  • Bench math is refactored into agent/src/factors/bench_runner.py so the CLI driver (agent/scripts/w4a_run_benches.py) and the new API worker share a single implementation.

Added — Safety floor

  • agent/tests/factors/test_alpha_purity.py — AST allowlist scan over every zoo/**/*.py module (whitelist: pandas, numpy, scipy.*, src.factors.base, __future__, typing, math, dataclasses; banned: os, sys, subprocess, socket, urllib, requests, httpx, pathlib, Path, open, eval, exec, compile, __import__, and getattr(_, "__*")).
  • agent/tests/factors/test_lookahead.py — sentinel future-row injection on a 300-row synthetic panel; corrupting rows after the probe must leave the probe value unchanged within 1e-9.
  • tools/ci_grep_gates.sh — CI gate that rejects yaml.load( without safe_load, any trademarked-name leak in shipped artifacts, and any per-stock-code data leak in wiki/**/*.{json,csv,html}.
  • agent/tests/factors/conftest.py — opt-in pytest-socket integration that hard-fails any test attempting outbound network during the factors test suite.

Added — Community governance

  • CONTRIBUTING.md — Developer Certificate of Origin sign-off requirement and a contributor checklist for new alpha PRs (purity, lookahead, __alpha_meta__ shape, LaTeX-matches-code, per-zoo LICENSE.md, DCO).
  • NOTICE (repo root) — Apache-2.0 attribution for Qlib and a declaration that the bundled formulas from Kakushadze, GTJA, and the academic baselines are mathematical content (paper prose, tables, and figures are not reproduced here).
  • Per-zoo LICENSE.md for each of qlib158/, alpha101/, gtja191/, and academic/, plus an upstream NOTICE for qlib158/.

Changed

  • agent/src/tools/factor_analysis_tool.py extracted its IC/IR and layered-backtest helpers to agent/src/factors/factor_analysis_core.py so the new alpha_bench_tool reuses the same maths. Public tool signature is unchanged; _compute_ic_series and _compute_group_equity remain importable as backward-compatible aliases.
  • agent/cli.py grew by 7 lines to register the alpha subcommand; all handler logic lives in agent/src/factors/cli_handlers.py.
  • Packaging: pyproject.toml now ships zoo/**/*.yaml, zoo/**/*.md, and zoo/**/NOTICE as package data; MANIFEST.in recursively includes agent/src/factors.

Known limitations

  • The btc-usdt universe is single-asset; cross-sectional IC requires ≥2 instruments, so the bundled alpha101_btc bench run returns alive/reversed/dead = 0/0/0 by construction. Use a multi-symbol crypto basket (e.g. BTC + ETH + SOL + the top-N perpetuals) for meaningful cross-sectional results; a curated crypto-majors universe is planned for 0.2.

Internal

  • wiki/alpha-library/manifest.json and wiki/alpha-library/content/ are generated artifacts and gitignored. Run vibe-trading alpha export-manifest --out wiki/alpha-library/manifest.json --force followed by python wiki/scripts/build_alpha_library.py to regenerate the static site.