Skip to content

Soul-driven AI agent with permission-hardened tools, token budgets, and multi-channel access. Runs 24/7 from CLI, Telegram or More.

26.6 KB
Raw Blame

Mercury — Improvement Plan (October 2026)

Scope: make Mercury a top-tier personal agent harness across macOS, Linux, Windows and Termux, with the TUI as the flagship surface. Derived from a read-only audit of v1.3.0 (5aae3fb) covering the agent loop, TUI, channels, Mercury Code, Bots, security backlog, and the build/deploy pipeline. All file:line references are against that commit.


0. Executive summary — five root causes

Most user-visible pain traces back to five things. Fix these and the product changes character; everything else is polish.

# Root cause Symptom users report Where
1 Mercury Code's guards run in plain chat. The narration guard and verification gate fire whenever programmingMode is off and the message matches an implementation-intent regex (`mak add test
2 "Done" is not evidence-gated in practice. A failed or timed-out npm test satisfies the verification gate (string match, not exit code); remote channels never receive the verification line; pause banners are truncated to 80 chars and deleted on Telegram. Green "✓ Verified" over a broken build; bot "goes silent" on Telegram; permission prompts silently auto-deny after 120 s. execute-guard.ts:216; agent.ts:4707/4800/4829/5097; channels/telegram.ts:361-378, 680, 716
3 The TUI loses input and can't be interrupted. Multi-line paste strips \n; no bracketed paste; Ctrl+C hard-exits, Esc does nothing; chat mode re-parses the whole streaming message as markdown every 60 ms; the Ink patch can silently fail, after which new messages stop rendering past 100. Mangled code pastes, "it feels laggy", no way to stop a turn except /stop, CJK input broken (#41). src/ui/input-composer.tsx:250-259; src/ui/App.tsx:292, 681, 1705; package.json:23,66; tsup.config.ts:45
4 Standalone binaries ship without a working SQLite on any OS, and Termux is mis-detected (android ≠ linux). Second Brain silently off for every binary user (not just Linux #96); bots fall back to JSON; Termux exits after the wizard with "Unsupported platform". src/memory/second-brain-db.ts:15-34, src/bots/queue.ts:15-31; src/utils/platform.ts:8; src/cli/service.ts:495-499
5 Security backlog is mostly fixed but publicly "unpatched". 8 of 13 open security issues are fixed in code and not closed; 5 CVEs + 4 GHSAs list "Patched versions: None"; no SECURITY.md. One unreported shell injection remains in git_status/git_log/git_diff. Reputation risk any investor or enterprise buyer finds in minutes. src/capabilities/git/git-status.ts:13, git-log.ts:16, git-diff.ts:16; permissions.ts:477-498

Cross-cutting: src/core/agent.ts is 8,362 lines with a single 2,400-line handleMessage and onStepFinish copy-pasted four times; 699 any casts; 981 console.log; no ESLint; no end-to-end eval harness; 154 of 174 commits in 90 days by one person.


1. Scorecard

Area State One-line verdict
Conversation quality 🔴 Guards misfire in chat; history = last 10 texts, tool results never persisted; memory injected as fake dialogue
Latency 🟠 ~10K-token uncached prefix per step; blocking channel edits per step; 3 full session-JSON reads + ~9 SQLite queries before first token
TUI 🟠 Architecture is sound (Static scrollback, diff-render); input/interrupt/paste are below par; patch shipping fragile
Feedback (all channels) 🔴 No consistent started/step/waiting/done/failed contract; evidence lost on remote channels
Mercury Code 🟠 Good ideas (completion contract, watchdog) undermined by string-match verification and a watchdog that kills long builds
Bots (new) 🟠 Solid queue design; duplicate-execution race; lead synthesis never delivered; needsYou volatile
Security 🔴 (process) 🟠 (code) Code mostly patched; disclosure process absent; one unreported injection; safe-read classifier is a losing regex race
macOS 🟢 Reference platform
Linux 🟠 Fine via npm; binary lacks SQLite; arm64 untested in CI
Windows 🔴 Service install broken (#13), wizard can exit silently (#64), kill = TerminateProcess, upgrade race, update check dead
Termux 🔴 Mis-detected platform; npm-only; no service
Engineering foundations 🟠 Good docs/ADRs and OS matrix in CI; no lint, no eval harness, no binary smoke tests, monolith files

2. Priorities

P0 — this week (release as 1.3.1 hotfix)

# Change Evidence Size Acceptance
P0.1 Gate narration guard and verification gate to programmingMode.isExecute() only agent.ts:4459-4476, :4745-4747 S "Make me a workout plan" and "write my notes to notes.md" produce a direct answer with zero forced rounds; regression test in execute-guard.test.ts
P0.2 Verification must be ok === true and after the last mutation; timeout = ✗; failed verification → pause banner, never "complete" execute-guard.ts:216, agent.ts:4707/4800, run-command.ts:113 S Test: npm test exit 1 → no completion banner; timed-out test → ✗
P0.3 endTask() before terminal sends; flush deferred text; never truncate pause/failure banners agent.ts:4829, telegram.ts:361-378, discord.ts:445, slack S Telegram test: step-budget pause arrives as a full persistent message
P0.4 Fix git_status/git_log/git_diff shell injection: execFileSync('git', ['-C', path, …]) + checkFsAccess(path,'read') git-status.ts:13, git-log.ts:16, git-diff.ts:16 (pattern already correct in git-add.ts:16) S Test with a path containing "; echo pwned
P0.5 Close the 8 fixed-but-open security issues with commit refs; add SECURITY.md; publish advisories for the 5 CVEs / 4 GHSAs with fixed versions; CHANGELOG security entry issues #101/#71/#77, #82/#72, #80/#76, #99/#75 (residual race noted) S (admin) GitHub shows a security policy and advisories with patched versions
P0.6 Termux detection: platform === 'linux' || 'android'; assert in Termux CI job platform.ts:8, index.ts:2249-2275 S Termux job asserts process.platform and that autoDaemonize doesn't exit
P0.7 Bots: running.has(job.id) guard in retry/pause timers; deliver lead mailbox turns; make needsYou durable (journal row) bot-manager.ts:521-548, 565, 154 S Test: retry + due-sweep never runs a job twice; lead synthesis reaches the thread
P0.8 Permission/choice prompt timeouts write the outcome back ("⏱ no answer → denied") and keep the card; editMessageReplyMarkup after approvals (#23) telegram.ts:262-275, 680, 716, 760 S Manual + unit
P0.9 Ink patch: pin "ink": "5.2.1" exactly; ship scripts/apply-ink-patch.cjs in files; assert patch markers at boot and abort with a clear message package.json:23,66, tsup.config.ts:45, cli.ts:604 S Fresh npm i --ignore-scripts → clear error, not a silent 100-message cap
P0.10 Drop minute-resolution time/budget from the system prompt (append at the end at day resolution); add Anthropic cache_control; log cachedInputTokens agent.ts:5201-5215 S Cached-token ratio visible in completionMeta

P1 — next 2–3 weeks (1.3.2 / 1.4)

# Change Evidence Size
P1.1 Feedback contract implemented behind one TaskSurface interface used by CLI, web, Telegram, Discord, Slack, Signal (see §3C) four divergent status-card implementations in src/channels/* M
P1.2 Context builder v2: persist compact tool-step records (tool, args head, result head) in the session; token-budgeted window with rolling summary; memory moved from fake dialogue into a system section with a score threshold agent.ts:3063-3065, 3110-3118, 4917-4934; user-memory.ts:255-266 M
P1.3 Per-step channel updates fire-and-forget + coalesced; Promise.all cloud pool search with memory retrieval; cache session reads per turn; batch SQLite writes agent.ts:3130, 3678-3739; sessions/repository.ts:132→458 M
P1.4 Mode-aware activeTools: no Spotify/GitHub/scheduler schemas in a plain chat turn (~6.5K tokens of tool schema per step today) capabilities/registry.ts:277 S
P1.5 TUI input: keep \n\t, enable bracketed paste (?2004h), strip ESC[200~/201~, Shift+Enter multi-line; grapheme-aware cursor (Intl.Segmenter + string-width); persistent history input-composer.tsx:250-259, 284; App.tsx:368, 700 M
P1.6 TUI interrupt semantics: Esc = interrupt turn, Ctrl+C = clear input / double-tap quit; prompt resolver map keyed by prompt id App.tsx:292, 681; cli.ts:1757-1845 S
P1.7 TUI perf: reuse Mercury Code's fence-aligned tail renderer in chat/coding; one shared 100 ms tick store replacing five tickers; async readGitStateQuick; reset log-update baseline on resize App.tsx:1705, 1834, 1749, 1925, 2296; cli.ts:2183; patch log-update M
P1.8 Tool progress + abort plumbing: run_command streams stdout to markProgress, honours abortSignal, stall clock suspended while a tool with explicit timeout runs run-command.ts, agent.ts:397, 2278, 4345 M
P1.9 Standalone SQLite: bun:sqlite backend when process.versions.bun; node:sqlite fallback for toolchain-less npm installs; mercury doctor --storage second-brain-db.ts:15-34, queue.ts:15-31, 475, board-db.ts:135, pool-search.ts:10-12 M
P1.10 Windows: fix installWindows quoting + getDistPath via realpathSync(argv[1]); HKCU Run fallback for non-elevated; graceful shutdown over local HTTP instead of process.kill; installer waits on parent PID; execFile('npm') with shell:true; os.homedir() for ~ service.ts:59-64, 495-499, daemon.ts:219-223, index.ts:4740-4757, update-notice.ts:96, agent.ts:7168,7817 M
P1.11 Wizard robustness (#64): single readline, reject on close, global error handlers before configure() index.ts:152-160, 2325-2350 S
P1.12 Read-side path canonicalisation + nlink>1 prompt + fstat dev/ino (#104 and symlink reads); strip list_agents/stop_agent/delegate_task from children by default + lineage check (#74) permissions.ts:477-498; stop-agent.ts:14-22, supervisor.ts:338 S/M
P1.13 /new on every channel (fresh session, memory kept) (#38); /trace <id> per-turn trace; daemon log level info; mercury doctor --health agent.ts:7582, telegram.ts:313-329, utils/logger.ts:5, index.ts:3852 M
P1.14 Provider layer: DeepSeek thinking by prefix (deepseek-v4*, reasoner), pass stored reasoning back (#24); capability tiers so fallback never silently downgrades; remove sticky markSuccess override; sane maxOutputTokens per provider; set temperature providers/deepseek.ts:14, providers/registry.ts:166-176, agent.ts:4409, config.ts:295/328 M

P2 — this quarter (1.5)

# Change Why
P2.1 Split handleMessage into prepareContext → generate → guards → deliver with one shared onStepFinish; break agent.ts, index.ts, App.tsx, cli.ts into modules Velocity and bus factor; every P1 item is harder inside an 8K-line file
P2.2 Auto-approval argv lane: POSIX-tokenise, allowlisted argv[0], per-command flag policy, execFile(absBinary, argv, {env: minimal}) from a PATH pinned at startup; shell only in the approval lane. Immutable per-agent permission context (capability tokens, child ⊆ parent) Closes the whole safe-read bypass class (#103, residual find -fprint0, rg --pre, git --output, --ext-diff) and the #99/#75 race
P2.3 Replay-based eval harness: fixtures (history, message, recorded tool outputs) → mocked provider; assert no forced guard in chat, non-empty answer, verification correctness, per-phase timings; run in CI Today nothing drives handleMessage; quality regressions are invisible
P2.4 Vendor the patched Ink (noExternal + alias to vendor/ink), drop postinstall; upstream Yoga hygiene (facebook/yoga#1818) and Static.itemKey; real hardware-cursor positioning for IME (#41) Deterministic TUI on every install path incl. Termux
P2.5 Declarative keymap table driving App.tsx, /help and attach TUI; attach consumes TuiState diffs over SSE instead of reconstructing from events Removes ~135 lines of drifting duplicate key handling
P2.6 CI: Bun job per OS building the host binary and running --version, doctor --platform, doctor --storage; Node 24/25; linux-arm64; installer scripts against a local release dir; shellcheck/PSScriptAnalyzer; unit tests for plist/unit/schtasks generation No binary or service path is tested today
P2.7 ESLint (strict), any budget with a ratchet, console.log → logger, files/tarball verification test Hygiene that compounds
P2.8 Termux first-class: ~/.termux/boot/start-mercury from mercury service install, node:sqlite path, narrow-width TUI rules (cols-based instead of fixed 60/50/34+56/26) Makes the fourth platform real

3. Workstreams (detail)

A. Conversation quality and latency

Architecture: enqueueMessage (661) → processQueue (2383) → handleMessage (2805-5179) → buildSystemPrompt (5181); AI SDK streamText with stepCountIs(75), maxOutputTokens: 32768, all 53 tools every step; history = last 10 message entries from JSON sessions; memory via FTS.

Problems and fixes, ranked:

  1. Guards in chat (P0.1). Also isTextDeliverableRequest should short-circuit the verification gate.
  2. Cross-turn amnesia: only finalText is stored; the [Using: …] loop-detector regex at 3072 scans a marker never written. Persist tool-step summaries; budget by tokens (P1.2).
  3. Memory as fake dialogue, injected even with zero matches (profile fallback). System section + threshold (P1.2).
  4. Uncached ~10K prefix (system ~3.8K + tool schemas ~6.5K) re-sent on up to 75 steps; prompt changes every minute. P0.10 + P1.4.
  5. Blocking per-step channel edits serialise the loop (P1.3).
  6. Hot path before first token: awaited cloud pool search (800 ms timeout) in series; presentChoice modal for ambiguous skills; Mercury Code handoff modal (45 s) on coding-looking CLI messages; 3 full session reads; ~9 SQLite queries. P1.3.
  7. Non-answer endings: loop-detector aborts end with a banner and no synthesis; empty text becomes "(no text response)" and is delivered; intermediate narration is concatenated with the final answer. Fix: on abort, run one final no-tools round to summarise; never deliver "(no text response)"; separate narration stream from answer stream.
  8. Provider layer (P1.14).

Metrics to add to completionMeta and /trace: prompt-build ms, memory ms, first-token ms, per-step ms, cached/uncached input tokens, guard decisions, provider switches.

B. TUI

Architecture is right (single Ink root, TuiState via useSyncExternalStore, <Static> scrollback, diff-rendered live region). The gaps are input, interrupt, perf and patch shipping.

Order: P0.9 (patch safety) → P1.6 (interrupt) → P1.5 (paste/IME/history) → P1.7 (perf) → narrow-width rules (P2.8) → P2.4/P2.5.

UX features, in priority order once the above lands: permission prompts showing the diff/command inline with "always for this tool"; ? keybinding overlay and rotating status hints; Ctrl+R history search; /copy via OSC 52 and OSC 8 links; real LCS diff with intra-line highlight (tool-block.ts:133 is prefix/suffix only); OSC 11 theme query (today COLORFGBG only, mis-detects Windows Terminal / VS Code / tmux); bell when a prompt needs you; Ctrl+O expand in place.

Known-bad specifics to fix along the way: Delete acts as backspace (App.tsx:750); setSlashSelIdx(0) effects add two commits per keystroke (App.tsx:128,150); 120 ms tail memo has no trailing flush (App.tsx:2652); attach streaming only clears on text_done/error so tool-only turns show "working…" forever (attach-tui.tsx:207-213).

C. Feedback contract (every surface, every turn)

  1. started — one line within 2 s: intent restatement + mode (chat/code/bot), edited in place.
  2. step — n/N · tool target · ok|err · Δt in one live region; never a new message per step.
  3. waiting-on-you — explicit card with options and a visible timeout; the timeout outcome is written back ("no answer → denied").
  4. done-with-evidence — files changed, verification command + exit code + 1-line output tail, tokens/cost/time. No "complete" without verified.ok === true or an explicit "unverified because: …".
  5. failed/paused-with-reason — which guard/provider/limit fired, what was preserved, the one command to resume; delivered as a persistent message, never a status-card notice.
  6. trace handle — every terminal message carries an id resolvable by /trace <id> and the web cockpit.

Implement once behind TaskSurface; channels render, they do not decide.

D. Mercury Code and Bots

Mercury Code: P0.2 (verification semantics), P1.8 (progress/abort so a 10-minute build isn't killed at 4 minutes with "stalled provider"), plan checklist reconciled against actual tool success instead of model self-report (cli.ts:1437), update_plan on all channels (today CLI-only, agent.ts:2534), checkpoint/undo beyond /code diff.

Bots: P0.7 (duplicate run race, lead delivery, durable needsYou), sub-agent results not sliced to 500 chars (supervisor.ts:315), "needs you" sends retried, retention caps verified under load.

E. Security

Code order: P0.4 (git helpers) → P1.12 (reads, lineage) → P2.2 (argv lane + capability tokens) → PATH pinning in service files (service.ts:297, 398). Windows blocklist additions: pwsh, Remove-Item, Set-ExecutionPolicy (blocklist.ts). .env written with mode: 0o600 (index.ts:723-734).

Process (P0.5): SECURITY.md with private reporting via GitHub advisories and an SLA; advisories published for CVE-2026-22169, -32010, -30741, -28463, -26322 and GHSA-796m-2973-wc5q, -jccr-rrw2-vc8h, -943q-mwmv-hhvh, -qrp5-gfw2-gxv4 with fixed versions; close #101/#71/#77, #82/#72, #80/#76, #99/#75 with commit refs; add regression tests for find -exec, #99/#75, #100, #74 lineage. Until P2.2 ships, describe safe-read auto-approval as best-effort, and do not claim "asks before it acts" for git_status/git_log/git_diff until P0.4 is released.

F. Cross-platform

macOS Linux Windows Termux
Install OK OK (glibc) OK x64; arm64 refused npm only
Service OK OK P1.10 P2.8
TUI OK OK P1.5/P1.6 (Ctrl+C hard-kill in arrow-select.ts:117) P2.8 narrow widths
SQLite (binary) P1.9 P1.9 (#96) P1.9 n/a
Upgrade OK OK P1.10 npm
Channels Signal needs Java 17+ all no Signal no Signal

Also: install.sh:49 calls err before it is defined; no post-install --version smoke or rollback. CI additions in P2.6. Release workflow is manual (publish.sh, CI publish job disabled, ci.yml:138-163); binaries are built on one Mac with Bun and never executed in CI.

G. Engineering foundations

P2.1 split, P2.3 eval harness, P2.6 CI, P2.7 lint. Keep ADRs current: add ADR-015 (feedback contract / TaskSurface), ADR-016 (argv-lane auto-approval and capability tokens), ADR-017 (vendored Ink), ADR-018 (SQLite backend selection: better-sqlite3 → bun:sqlite → node:sqlite → JSON).


4. Targets

Metric Today (estimated) Target
Forced tool rounds in plain chat up to 10 0
"Complete" banners with failed/absent verification possible 0 (test-enforced)
Fixed prefix tokens per step ~10K, uncached ≤6K, ≥80% cached on Anthropic/OpenAI
Time to first token (local, warm) prompt build + 3 session reads + ~9 queries + serial pool search < 400 ms before provider latency; measured in /trace
Pause/failure messages lost on Telegram yes 0
Standalone binary with working Second Brain 0 of 4 targets 4 of 4
Windows mercury service install broken works for standard users
Termux mercury after wizard exits runs; boot script optional
Open security issues already fixed 8 0; advisories published
Tests driving handleMessage / TuiApp.useInput 0 / 0 eval harness + key-handling tests in CI

5. Sequencing

  • Week 1 — 1.3.1 hotfix. P0.1–P0.10. Mostly one-line conditions, admin, and small guards. Publish advisories the same day the git-helper fix ships.
  • Weeks 2–4 — 1.3.2 then 1.4. P1.1 (feedback contract) and P1.2 (context v2) first because everything else renders through them; then TUI P1.5–P1.7, Windows P1.10–P1.11, SQLite P1.9, provider P1.14, security P1.12, /new + /trace P1.13.
  • Weeks 5–12 — 1.5. P2.1 split (do it early in this window; later work gets cheaper), P2.2 argv lane, P2.3 eval harness in CI, P2.4 vendored Ink + IME, P2.6 binary CI, P2.7 lint, P2.8 Termux.

Hiring/ownership note: with one author on 90% of commits, the P2.1 split and the eval harness are what let a second engineer contribute safely. Prioritise them over new features.


Appendix — condensed findings by area

Agent loop (src/core/agent.ts, execute-guard.ts, providers/*, sessions/*, memory/*)

  • handleMessage 2805–5179; onStepFinish duplicated at 3473, 3917, 4563, 2430.
  • Narration guard 4459–4476 (chat mode), verification gate 4745–4747; FORCED_ACTION_TOOLS 4545; WORK_NOT_STARTED_BANNER; second-wind rounds 4619.
  • History 3063–3065 (10 messages, 4 in Saver); only finalText stored 4917–4934; dead [Using: …] detector 3072.
  • Memory injection 3110–3118, 3161–3168; profile fallback user-memory.ts:255-266.
  • Prompt: time/budget 5201–5215; soul+guardrails+persona ~5.8 KB; 25 skill summaries ~5.5 KB; 53 tool schemas ~25 KB incl. 14 Spotify tools always present.
  • Blocking step I/O 3678 (Telegram), 3692 (Signal), 3720 (Discord), 3734 (Slack), 3739 ([Using…]).
  • Hot path: pool search 3130; skill choice modal 3240; Mercury Code handoff 766; research prompt 752; sessions.get re-reads 132→458.
  • Loop-detector aborts 3552–3582, 4289–4295; "(no text response)" 4445, 4882; narration concatenation stepAwareTextStream 103–140; auto-continue only in execute 3831.
  • Providers: default deepseek-chat (config.ts:295/328); thinking only for exact deepseek-reasoner (deepseek.ts:14); sticky markSuccess (registry.ts:166-176); chain doubled 4409; maxOutputTokens: 32768 393; no temperature.
  • Tests: 20 core/*.test.ts, pure helpers only.

TUI (src/ui/*, src/channels/cli.ts, patches/ink+5.2.1.patch)

  • Root cli.ts:1067; TuiState coalescing 712–741; token throttle 1640–1680; Mercury Code tail 32 KB / 120 ms memo App.tsx:2645.
  • Hook census App.tsx: 25 useState, 15 useEffect, 6 setInterval; useInput closure 485 lines (281–766).
  • Paste: cleanPrintableInput input-composer.tsx:250-259; no ?2004h.
  • IME: hidden cursor log-update.js:9; fake cursor App.tsx:2130, 2430; UTF-16 editing input-composer.tsx:284, App.tsx:368, 700.
  • Patch: tsup.config.ts:45 external; package.json:23 postinstall || echo; scripts/ not in files (66); caret range; 100-message window App.tsx:914, 1052, 2691 + ignored itemKey on stock Ink.
  • Markdown per frame App.tsx:1705; tickers 1834, 1749, 1925, 2296, use-terminal-size.ts:21.
  • Interrupt: App.tsx:292 → index.ts:2711; Esc 681; resolver slot cli.ts:1757-1845.
  • Resize vs diff-render ink.js:91; readGitStateQuick cli.ts:2183; fixed widths App.tsx:2110, 938, 772-786, 1101, 2259; attach attach-tui.tsx:207-213, 323-457.

Bots / Mercury Code / channels

  • Verification: execute-guard.ts:216; vOk regex agent.ts:4707/4800; timeout text run-command.ts:113; CLI-only note 5097; channel completions 4985–5065.
  • Pause swallowed: agent.ts:4829 → telegram.ts:361-378 → finalizeChannelTask 2796; discord.ts:445.
  • Watchdog: MAX_STALL_MS agent.ts:397, 2278; ambiguous end 4345; run-command ignores abortSignal.
  • Bots: retry race bot-manager.ts:521-548 vs resumeDueJobs 261; lead delivery 565 / sendToBot 370; needsYou 154; sub-agent slice supervisor.ts:315.
  • Prompts: telegram.ts:680, 716, 760; keyboards 262–275; statusNotices 372 unbounded.
  • /reset agent.ts:7582; Telegram commands telegram.ts:313-329; logger utils/logger.ts:5; doctor index.ts:3852; plan progress agent.ts:2534, cli.ts:1437.

Security

  • Fixed-not-closed: permissions.ts:681 (redirection), :686 (find flags), :696-697 ($, backticks, ~); tests permissions.test.ts:181-262 (no explicit -exec case).
  • Sub-agent approvals sub-agent.ts:172-175; shared mutable PermissionManager supervisor.ts:183; internal auto-approve window agent.ts:2929-2933 → 5174; isGlobalAutoApproveActive permissions.ts:386-390; "full permissions" prompt sub-agent.ts:504.
  • Lineage: stop-agent.ts:14-22, supervisor.ts:338; tool-filter.ts:5.
  • Reads lexical only permissions.ts:477-498; realpath for writes 484–492; read-file.ts:21-38.
  • Unchecked exec: git-status.ts:13, git-log.ts:16, git-diff.ts:16, git-push; registry.ts:244-246; run-command.ts:32-36 shell:true; PATH in service.ts:297, 398.
  • Residual auto-approved primitives: find -fprint0, find -fls, tree -o, git log --output, rg --pre, --ext-diff.
  • Process: no SECURITY.md; advisories unpublished; CHANGELOG security section incomplete; only docs/security/ssrf-108-response.md.

Cross-platform

  • Build: tsup.config.ts:40-62 externals; post-build.cjs; build-bin.cjs:75-81, 138-146, 173-207; verify-standalone-release.cjs never runs a binary; ci.yml:138-163 publish disabled; files package.json:67-71.
  • SQLite in binaries: better-sqlite3/lib/database.js:48 via bindings; probes second-brain-db.ts:15-34, queue.ts:15-31, board-db.ts:135, pool-search.ts:10-12; warn only index.ts:2460-2463.
  • Windows service: service.ts:495-499, 59-64, 95; stop daemon.ts:219-223; shutdown index.ts:3632-3666; upgrade index.ts:4740-4757, install.ps1:204-206; update check update-notice.ts:96; ~ via HOME agent.ts:7168, 7817, cli.ts:2047, 2203, bot-manager.ts:1387.
  • Termux: platform.ts:8; autoDaemonize index.ts:2249-2275; binary refused install.sh:198; Signal binary.ts:32.
  • Wizard: index.ts:152-160, handlers 2325–2350; .env mode index.ts:723-734 vs config.ts:657, web/auth.ts:22.
  • install.sh:49 err before definition; install.ps1:147 arm64 refused.
  • Tests: 109 files; it.skipIf(!sqliteAvailable) ×17; live-region-guard.test.tsx:68; none for service.ts, daemon, installers, binaries.