Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
docs(structure): map the Recording Pipeline surface to its page, API and shared derivation
nirholasa9ab44af7022
diff --git a/STRUCTURE.md b/STRUCTURE.mdindex 08720615d..d1a8ef60e 100644--- a/STRUCTURE.md+++ b/STRUCTURE.md@@ -236,6 +236,7 @@ bottom.| Oracle conviction model (three heads, learned on a clock) | Fitter [api/_lib/oracle/fit.js](api/_lib/oracle/fit.js), scorer [api/_lib/oracle/conviction.js](api/_lib/oracle/conviction.js), DB-backed weights [api/_lib/oracle/model-store.js](api/_lib/oracle/model-store.js), learning loop [api/cron/oracle-refit.js](api/cron/oracle-refit.js) (every 6h), CLI [scripts/oracle-fit.mjs](scripts/oracle-fit.mjs) | Live, v3 | v2 ranked "spikes 3x or graduates", so a coin that ran 3x and went to zero counted as a hit: Prime-tier coins hit that target 58.2% of the time and rugged 53.0% of the time. v3 fits three heads over one design matrix (`win` = ran and the holder kept it, `rug` = holder down more than half, `moon` = ran at all) and anchors the public score on `win`, publishing `giveBackRisk = 1 - P(win)/P(moon)` beside it: given it runs, how often a coin like this hands the run straight back. Weights moved out of the container image into `oracle_model_versions`, so the refit cron can promote a better model between deploys; it was frozen at its 2026-08-09 fit for nineteen days because nothing but a human could refit it. Every candidate faces a promotion gate (complete fit, absolute AUC floor, every tier band still earning its claim, no feature-set collapse, a material gain over the incumbent, no regression on the other heads) and is stored either way with the reason. Thin buckets are shrunk toward zero by `n/(n+200)` so 34 samples cannot reverse a verdict. Docs: [docs/oracle-model.md](docs/oracle-model.md) || Price-independent outcome labels | [workers/agent-sniper/intel/learn.js](workers/agent-sniper/intel/learn.js) `deriveOutcome`, migrations [20260828170000](api/_lib/migrations/20260828170000_oracle_price_independent_labels.sql) + [20260828190000](api/_lib/migrations/20260828190000_oracle_recompute_rug_flag.sql) | Live | The rug flag was a readout of the SOL price. It was decided by "market cap under $3,000", and an empty pump.fun bonding curve is worth a fixed `30 * 1e9 / 1073000191 = 27.958993 SOL`, so the test amounted to asking whether SOL was above about $107.30 that day: of 206,428 coins labeled rugged, 206,419 were under $3,000; of 25,180 labeled survivors, the cheapest was exactly $3,000. Same dead curves, sorted by a price feed. Labels now use `retained = last_mc/ath_mc` and `hold_multiple = ath_multiple * retained`, both taken from the same reading so the price cancels; the real rug rate is 11.4%, not 91%, and it holds within two points from 60 minutes to three days. `label_version` marks which rule judged a row and the fitter trains on version 2 only. The training-set upsert also stopped being `do nothing`, which had made a written label permanently uncorrectable. || Coin Intelligence (the read the sniper trades on) | [pages/coin-intel.html](pages/coin-intel.html) `→ /coin-intel` (self-contained page: markup, styles and module script in one file), read API [api/pump/intel.js](api/pump/intel.js) `→ /api/pump/intel`, serverless firehose [api/cron/coin-intel-observe.js](api/cron/coin-intel-observe.js), shared finalize path [workers/agent-sniper/intel/finalize.js](workers/agent-sniper/intel/finalize.js) | Live | The public dashboard over the Coin Intelligence engine: every pump.fun launch observed for its first ~90 seconds, classified, scored 0..100, and labeled by outcome after the fact. Four views off one endpoint (`view=feed` live radar with category/verdict/quality/search filters, `view=leaderboard` best of 24h plus confirmed graduated/pumped winners, `view=traders` cross-coin wallet reputation with real win-rates, `view=learning` the per-signal weights the model learned). A coin's card opens a drawer with its launch signals, trade footprint, risk flags, a funder-cluster bubble map, the classified order book, and Oracle conviction fetched in parallel from [api/oracle/coin.js](api/oracle/coin.js). Reads `pump_coin_intel` / `pump_coin_wallets` / `pump_coin_outcomes` / `pump_intel_weights`, all written by the engine; the endpoint never writes and degrades to a designed "engine warming up" state (200 with `degraded:true`) when those tables are absent. `traders` and `learning` are whole-history aggregates and are edge-cached for minutes; the radar stays on a 5s cache and auto-refreshes every 15s. Watchlist stars share the `ld_watchlist` key with `/oracle`, `/launches` and `/watchlist`. |+| Recording Pipeline (is the data loop actually running?) | [pages/pipeline.html](pages/pipeline.html) `→ /pipeline` (self-contained page: markup, styles and module script in one file), one-call status API [api/pipeline.js](api/pipeline.js) `→ /api/pipeline`, shared derivation [api/_lib/pipeline-summary.js](api/_lib/pipeline-summary.js) | Live | One board for the whole closed loop that turns a pump.fun launch into a trade signal: the always-on recorder ([workers/agent-sniper](workers/agent-sniper)), signal intel, ground-truth outcomes, Oracle conviction, the wallet reputation graph, trained per-signal weights, and the agents trading on the result. Seven stage cards, each linking into the surface that owns it, over one endpoint that queries every stage independently so a cold worker or a table that does not exist yet reports an honest zero rather than failing the whole response. The recorder is network-scoped: `bot_heartbeat` is keyed by worker alone, so asking about devnet while the worker records mainnet answers `offline` with a reason instead of lending one network's liveness to the other. The board polls every 8s and patches its cards in place, so a card the user tabbed to keeps focus and only the one-line health summary is announced. `next_action` is the same machine-readable recommendation the page renders as its call to action, so an agent reading the JSON and a person reading the page get one answer. Docs: [docs/api-reference.md](docs/api-reference.md) (Recording Pipeline API). Tests: [tests/pipeline-summary.test.js](tests/pipeline-summary.test.js) || Agent voice cloning (your voice, on your agent) | Capture + BYOK panel [src/voice/voice-setup.js](src/voice/voice-setup.js) (`VoiceSetup`), mounted as step 5 of the create wizard ([src/create-agent.js](src/create-agent.js) `→ /create-agent`), in the agent editor, and in the talk-mode shell [src/voice/voice-clone-modal.js](src/voice/voice-clone-modal.js); server [api/agents/_id/voice.js](api/agents/_id/voice.js) `→ GET/PUT/DELETE /api/agents/:id/voice` + `POST /api/agents/:id/voice/clone`; credential resolver [api/_lib/agent-voice-key.js](api/_lib/agent-voice-key.js); playback config mapper `agentVoiceConfig` in [src/runtime/speech.js](src/runtime/speech.js), consumed by the hosted embed [pages/agent-embed.html](pages/agent-embed.html) `→ /agent/:id/embed` and the SDK runtime [src/runtime/index.js](src/runtime/index.js) | In-repo · BYOK live, platform lane awaits a key | Record or upload 30+ seconds and the agent speaks in that voice everywhere on three.ws. Three credential lanes, and they are not interchangeable: a clone only exists inside the ElevenLabs account that made it, so the winning lane is recorded on `agent_identities.voice_key_source` and replayed on every synthesis. **BYOK is the activation path today** (no platform `ELEVENLABS_API_KEY` is set on the service), and the panel renders inline key entry rather than a dead "not configured" message; setting that one variable lights up the platform lane, metered to $THREE credits. `/api/tts/eleven` only spends an owner's key when the request names the agent, which is what lets a bound voice speak to signed-out visitors inside a cross-origin embed. Docs: [docs/tutorials/voice-and-lipsync.md](docs/tutorials/voice-and-lipsync.md), [src/voice/README.md](src/voice/README.md). Tests: [tests/api/agents-voice.test.js](tests/api/agents-voice.test.js), [tests/runtime/agent-voice-config.test.js](tests/runtime/agent-voice-config.test.js) || Trading Copilot (conversational) | client [src/agent-copilot.js](src/agent-copilot.js), hub tab [src/agent-wallet-hub/tabs/copilot.js](src/agent-wallet-hub/tabs/copilot.js), server [api/agents/copilot.js](api/agents/copilot.js) `→ POST /api/agents/:id/copilot` (SSE) | In-repo | Owner-only chat over an agent's Solana wallet: text/voice in, markdown replies, live **data cards** (portfolio, firewall, quote, smart-money, intel) and **confirm-before-execute** trade/limits cards. Read-only tools run server-side; state-changing intents return as proposals the browser re-quotes and runs through the existing spend guards + firewall + custody audit. Slash commands, copy/regenerate, localStorage persistence. See [docs/trading-copilot.md](docs/trading-copilot.md) || Alpha Co-pilot (the agent narrates its own read) | page [pages/alpha-copilot.html](pages/alpha-copilot.html) `→ /alpha-copilot`, client [src/alpha-copilot.js](src/alpha-copilot.js), styles [src/alpha-copilot.css](src/alpha-copilot.css), profile cross-link [src/shared/alpha-copilot-link.js](src/shared/alpha-copilot-link.js), server [api/agents/alpha.js](api/agents/alpha.js) `→ GET /api/agents/:id/alpha/candidates` + `POST /api/agents/:id/alpha/read`, grounding + scrub [api/\_lib/alpha-read.js](api/_lib/alpha-read.js) | Live | The intelligence and voice layer on the same trade rails the Trading Copilot uses: pick a public agent from a real gallery, pick a real live pump.fun mint off the candidate feed, and the agent reads it in character while its `<agent-3d>` avatar speaks the verdict aloud (agent TTS provider, then the free server lane, then browser speech). Every figure it says is validated against the `signals` bundle the server actually fetched (bonding-curve state, a real non-binding reference buy quote for true price impact, graduation progress, smart-money wallets, and for the owner the wallet balance and spend limits); a mismatch replaces the line and raises `hallucination_guard.line_replaced`, which the page surfaces rather than hides. The narrator never signs: acting on a read is an owner-only drawer that re-quotes through `previewAgentTrade` and executes through `executeAgentTrade`, so the spend policy, firewall, and custody audit all still apply. Docs: [docs/alpha-copilot.md](docs/alpha-copilot.md) |