@fusengine/harness

A governance layer for AI coding agents: gates backed by cited evidence (not vibes), a decision-time lesson memory, verification receipts, and a one-shot gate metric — ported from a Claude Code plugin into one Bun-native npm package. The policy core is harness-agnostic; how much of it actually enforces depends on the harness's own hook system — see the compatibility matrix below before assuming parity across harnesses.

detect → init (writes that harness's hook wiring) → `harness hook` → guards + APEX gates → native deny/ask/context

Install

npm i -g @fusengine/harness     # for the CLI (harness init/hook/check)
# or, as a library:
bun add @fusengine/harness      # Bun reads the TS source directly — no build step

Quickstart

cd your-project
harness init                              # detects the harness, writes its pre+post hook wiring
export FUSE_HARNESS_MARKETPLACES=fusengine-plugins   # (optional) which marketplaces to auto-scan
export FUSE_HARNESS_REFS=.claude/skills              # (optional) explicit refs dir, overrides auto-discovery

init writes the wiring file for the detected harness (.claude/settings.json, .codex/hooks.json, .cursor/hooks.json, .gemini/settings.json, or .clinerules/hooks/PreToolUse+PostToolUse) — for Claude Code, Codex, Cursor, Gemini CLI and Cline this covers PreToolUse + PostToolUse only (src/init/templates.ts); Hermes has no init runner because its config lives outside the project at ~/.hermes/config.yaml (src/adapters/hermes/index.ts:3-4). The richer lifecycle (SessionStart, SubagentStart/Stop, Stop, PreCompact/PostCompact, TaskCompleted, TeammateIdle, PostToolUseFailure, InstructionsLoaded — 14 distinct Claude Code hook events in total, src/runtime/lifecycle/dispatch.ts + src/runtime/handle.ts:56-80) is implemented in the runtime but only fires when something wires those extra hook events into .claude/settings.json — either by hand, or via the fusengine-plugins marketplace this repo is itself developed under (its ~11 sibling plugins each add their own event, src/runtime/burst-window.ts:5-11).

Where state lives (two separate roots — this changed across 0.1.34→0.1.51, don't assume either the old or the "everything in one place" story):

• Session track, deny-loop/one-shot sidecars: out-of-tree, under ~/.fuse-harness/state/<8-char-md5-of-project-path>/ (src/runtime/paths.ts:5-6,41-43) — deliberately outside the repo so the protected-path guard never has to bless writes to its own enforcement state.
• MCP/WebFetch response cache and the curated MEMORY/LESSON.md-equivalent lessons file: still in-tree, under <project>/.harness/{cache,memory} (src/config/layout.ts:15-26) — .harness/ is gitignored except memory/LESSON.md, which is meant to be committed (STATE_GITIGNORE, src/config/layout.ts:10).

CLI

Command	What it does
harness init [id]	Write the pre+post hook wiring for the detected (or named) harness.
harness hook <id>	Runtime: read a hook payload on stdin, gate (pre) or record (post), print the native response. (Hooks call this — you don't.)
harness check	cli-mode: check staged files in a pre-commit step, exit non-zero on a violation. For harnesses without hooks.
harness doctor	Print the version + resolved path of the harness actually executing, and compare it to npm's latest — the fast way to catch a stale global (see Pinning).
harness --version	Print the running version (bare, on stdout) and exit.

Every invocation writes a @fusengine/harness vX.Y.Z banner to stderr (never stdout — the hook JSON contract stays clean) so you can see which version ran.

cli-mode (Aider / Windsurf / OpenHands), as a pre-commit step:

# .husky/pre-commit
npx harness check

Pinning (required for hook consumers)

Any consumer hooks.json / settings.json that runs the harness via bunx MUST pin an exact version — @fusengine/harness@X.Y.Z — never a range (^ / ~) and never the bare or @latest spec:

// .claude/settings.json — correct: exact pin
"command": "bunx @fusengine/harness@0.1.56 hook claude"
// WRONG — may silently keep running a stale global install:
"command": "bunx @fusengine/harness hook claude"

Why. A still-open bun bug (oven-sh/bun#5791) makes an unpinned bunx <pkg> prefer an already-installed global copy over npm-latest, and publishing a new version updates neither that global nor the bunx cache (bun pm cache rm does not remove the global). A hook wired without an exact pin can therefore keep executing an old harness indefinitely after you publish a fix. Pinning @X.Y.Z forces the exact version to resolve.

Run harness doctor to see which version is actually executing and whether npm has a newer one:

harness doctor
# @fusengine/harness doctor
#   running:    0.1.56
#   package:    /Users/you/.bun/install/global/node_modules/@fusengine/harness
#   runtime:    /Users/you/.bun/bin/bun
#   npm latest: 0.1.57
#   ! stale — npm serves 0.1.57. Pin "@fusengine/harness@0.1.57" in hooks.json.

If doctor reports a stale global, clear it and re-pin:

bun remove -g @fusengine/harness && bun pm cache rm

Compatibility

There is no "runs the same on any harness." Each row is a real ceiling, not a formatting nuance — read it before assuming a gate that works on Claude Code also works elsewhere.

Harness	PreToolUse coverage	Lifecycle (Session/Subagent/Stop/Compact/…)	Known limit
claude-code	Full: evaluate + APEX gates via handleHook (src/adapters/claude/index.ts)	14 event types implemented (dispatch.ts) — fires once wired into .claude/settings.json beyond the init default	None found; richer lifecycle needs manual/marketplace wiring (see Quickstart)
codex	Bash gated reliably; apply_patch edits are now gated too — the patch text is parsed per file (adapters/codex/apply-patch.ts), each hunk runs the file gates and ONE violating hunk denies the whole patch (runtime/apply-patch-gate.ts, sim scenario 22 incl. the multi-file smuggling case). ask prompts are downgraded to explicit deny (respond.ts, sim scenario 23) because Codex fails open on unsupported shapes — the deny now carries a CONFIRM <code> recourse (see "CONFIRM code" under Beyond gating below). On the PostToolUse side, an allowed patch is fanned into one synthetic per-file event so the SOLID/tracking/post-edit handlers see every touched file (runtime/post-fanout.ts), and a non-blocking security advisory fires on the first qualifying file (securityAdvisoryForPatch).	Not wired by harness init codex (PreToolUse Bash|apply_patch + PostToolUse only, src/init/templates.ts:29-38) — Stop (session cleanup + the SOLID/receipt completion check, since Codex never emits SessionEnd/TaskCompleted, runtime/lifecycle/stop-core.ts) and SessionStart (plugin agents/commands cache resync, runtime/lifecycle/codex-resync/) ARE implemented but, same caveat as claude-code's extra lifecycle above, only fire once the codex-plugins marketplace's own hooks.json wires them.	Upstream caveat: Codex itself does not always enforce a correct apply_patch deny (openai/codex#27833) — we emit the right verdict; enforcement is theirs. No interactive ask. Codex has no native PostToolUseFailure; a failure is inferred from the ordinary PostToolUse result shape (non-zero exit or an explicit error field only — never a guess) and journaled into the one-shot failure tally (src/tracking/codex-post-failure.ts).
cursor	beforeShellExecution can deny/ask (shell only, cursor/index.ts:16-21)	none	File edits are advisory only: afterFileEdit always returns allow + a user_message correction on violation — a deny there has no proven effect (hook was "informational only" at launch, and Cursor's deny-enforcement for file ops is confirmed broken upstream, forum.cursor.com/t/154377). Human sees the message; the model is never re-informed. Platform ceiling, documented in cursor/index.ts.
gemini-cli	BeforeTool denies via {decision:"deny",reason} (gemini/index.ts:22-36)	none	Thin stateless adapter — no session track, no APEX gates wired through it.
cline	PreToolUse only; block → {cancel:true}, non-block → contextModification (cline/index.ts:24-36)	none	Same as gemini-cli: stateless guard only.
hermes	pre_tool_call proven: reuses the Claude stdin reader, blocks via {decision:"block",reason} (hermes/index.ts:12-36)	untested — no lifecycle dispatch wired for Hermes in this repo	ask/inform degrade to non-blocking {context} — Hermes "has no interactive ask state" (hermes/index.ts:27-28).
kimi	PreToolUse denies via the camelCase JSON channel {"hookSpecificOutput":{"permissionDecision":"deny","permissionDecisionReason":"…"}} on stdout at exit 0 (verified live against kimi-code v0.27.0 — exit 2 with stderr = reason also blocks, but is not required) — only deny is documented. ask is downgraded to deny prefixed [downgraded from ask — Kimi Code has no interactive approval], with a CONFIRM <code> recourse appended (see "CONFIRM code" under Beyond gating below); inform rides plain stdout text at exit 0, never wrapped in JSON. Blocking events: UserPromptSubmit, PreToolUse, Stop (kimi/index.ts).	Observation only: PostToolUse, PostToolUseFailure, PermissionRequest, PermissionResult, SessionStart, SessionEnd, SubagentStart, SubagentStop, StopFailure, Interrupt, PreCompact, PostCompact, Notification — Kimi delivers them but ignores any response, so no verdict can be returned from them.	A hook cannot request approval — Kimi's ask lives in a parallel, hook-unreachable system ([[permission.rules]] decision = "ask" in config.toml), hence the ask→deny downgrade. Hooks are configured only in the global ~/.kimi-code/config.toml (no project-local hooks file), so harness init writes no kimi wiring — copy the TOML snippet from docs/adapters.md, and emit no field beyond event/matcher/command/timeout or the whole config fails to load. Fail-open by design: any exit code other than 0/2, a timeout, or a crash lets the call through. Stdin payload is snake_case and carries only hook_event_name, session_id, cwd, tool_name, tool_input.command and an undocumented tool_call_id (unused here) — no transcript_path, no permission_mode, no tool_response. Verified live against kimi-code v0.27.0 for PreToolUse/Bash; validate hand-written config with kimi doctor. Instructions file is AGENTS.md, not CLAUDE.md.

What it enforces

Guard/gate chain evaluated before a tool runs (src/policy/guards/index.ts, src/policy/apex-gates.ts, src/policy/evaluate.ts):

Guard / gate	Fires on
security	rm -rf /, fork bombs, curl | sh; sudo (asks)
protected-path	edits to .claude/plugins|logs|cache, .git/, the harness's own state dirs
bash-write	sed -i / redirects to code files; python3 -c judged on content (mutating script → block, read-only → pass, same as node -e)
interface-separation	top-level interface/type/protocol in a component/controller
install	npm/pip/brew/... installs (asks)
git	destructive git (push --force, reset --hard, …) — block; routine git — ask
file-size (SOLID)	a code file over FUSE_SOLID_MAX_LINES (default 100)
APEX brainstorm	creating a new file without brainstorming (when flagged)
APEX freshness	explore-codebase + research-expert not run within the window
APEX doc-consulted	Context7 and Exa not consulted this session (a web-only fallback also passes)
APEX solid-read	required SOLID refs (auto-discovered, or FUSE_HARNESS_REFS) not read within the TTL
framework sub-skill	framework / shadcn / Tailwind code whose required skill wasn't read this session
Gemini MCP (opt-in)	hand-written Tailwind UI without a mcp__gemini-design__* call — only when FUSE_ENFORCE_GEMINI_MCP is set
MCP verbosity / cache	caps exa numResults; serves a fresh cached MCP/WebFetch result

A trivial-edit fast path lets a few tiny (< 5-line, non-replace_all) Edits through per window without the full APEX gates (Write is never trivial).

Beyond gating: memory, receipts, one-shot metric

Features shipped since 0.1.44, each with its own test:

• CONFIRM <code> — recourse for a degraded ask — Codex and Kimi both ignore permissionDecision: "ask" (Kimi's own binary shortcuts on hookSpecificOutput?.permissionDecision !== "deny"; Codex fails a hook open on the unsupported shape), so the harness downgrades every ask there to a hard deny. That deny now appends a short 4-hex-char code; retyping CONFIRM <code> in the very next prompt authorizes that exact action once (Claude Code is untouched — its native ask still shows an interactive confirmation, no code ever appears there, src/runtime/confirm/confirm-gate.ts). Guardrails, each independently testable (test/confirm.test.ts, test/confirm-g0.test.ts): G0 no token can be placed while a sub-agent is active for this session (a monotone max-write timestamp, window tunable via FUSE_CONFIRM_SUBAGENT_WINDOW_SEC, default 300s); G1 a token is consumed on first use; G2 a token expires after 5 minutes; G3 the token is scoped to the action's full SHA-256 hash, never the 4-char display code (which collides by design — it exists only for a human to retype); G4 irreversible commands (push --force, reset --hard, rm -rf, git clean -fd, branch -D, …) are never confirmable, regardless of a valid token; G5 an explicit refusal in the next prompt drops any pending token. This is a guard against accidental/hasty denial with no recourse, not a security control against an adversarial agent — an agent with arbitrary shell access can write the token straight into the session-state file it authorizes from and self-approve, same as it could bypass any other stateful gate this harness keeps outside a sandbox.
• Deny-loop breaker — an identical retried call that was already denied gets a rewritten [REPEAT] … STOP message forcing a different approach, instead of looping silently (src/policy/deny-loop.ts, test/deny-loop.test.ts).
• Burst-window dedup — every deployed plugin registers its own hook, so one real tool call can fan out to ~11 sibling processes; a same-op record within 2s is folded into the first instead of re-counted (src/runtime/burst-window.ts, test/burst-dedup.test.ts).
• One-shot gate metric — every gate outcome (deny or its later fix) lands in a 7-day sidecar keyed by a content-free op hash, so a deny→allow transition is visible (src/tracking/one-shot.ts, test/one-shot.test.ts).
• Verification receipts — a tsc/test run is captured from PostToolUse Bash output; TaskCompleted refuses a "done" over modified code files without a fresh passing receipt (src/tracking/receipts.ts, test/receipts.test.ts).
• Decision-time lessons — a MEMORY/LESSON.md bullet tagged with [TRIGGERS tool:… path:… error:… keyword:…] is injected as additionalContext the moment a matching call is about to repeat a known mistake, cooldown-guarded (src/policy/lessons/lesson-gate.ts).
• Failure lessons (Claude-Code-only — no PostToolUseFailure hook on Codex/Hermes, src/runtime/lifecycle/failure-lesson.ts:8-9) — a tool failure's error message is matched against error:-triggered lessons and injected on the spot (test/failure-lesson.test.ts, test/sim/scenarios/17-failure-lesson.json).
• TeammateIdle anti-false-done (Claude-Code-only, teammate-idle-check.ts:9-10) — files a teammate announced as changed are checked against disk; a missing file warns the lead before it's treated as done (test/teammate-idle-check.test.ts).
• PostCompact re-injection — after context compaction, the reconciliation snapshot is re-sent with a "reread files before editing" reminder, deduped per session (src/runtime/lifecycle/post-compact.ts, test/post-compact.test.ts).
• Reconciliation snapshot at SessionStart — git state, running harness version + drift vs. npm, .claude/BOARD.md, and the one-shot summary, each collector isolated so one failure can't blank the rest (src/runtime/lifecycle/snapshot/index.ts, test/snapshot.test.ts).
• Target-aware root-doc injection — SessionStart and UserPromptSubmit inject the target's own root instructions doc, not a hardcoded CLAUDE.md: apexDocName(id) + harnessHomeSegment(id) (src/policy/apex-target.ts) resolve AGENTS.md under .codex/.kimi-code for Codex/Kimi, CLAUDE.md under .claude for every other target (byte-identical to the pre-target-aware default). The systemMessage label follows suit ("AGENTS.md injected" under Codex/Kimi) (src/runtime/lifecycle/session-start.ts, src/runtime/inject-context.ts).
• Injection budget cap — harness-produced context fragments (lessons, snapshot, APEX task context) are capped at ~8000 chars each; owner-authored content (CLAUDE.md/rules) is never capped (src/runtime/inject-budget.ts, test/inject-budget.test.ts).
• Hook simulator — 18 end-to-end scenarios (payload in, expected verdict out) replayed against the real CLI in both src and built dist modes in CI (test/sim/README.md, test/sim/scenarios/).
• Codex multi_agent_v2 custom-agent evidence — a spawn_agent call carrying tool_input.agent_type is recorded as subagent-<agent_type> in the same SessionTrack as Claude Task evidence, so a Codex explore-codebase/sniper satisfies the APEX freshness gates exactly like a Claude subagent (src/tracking/session-state.ts, sim scenario 32b). Recognized agent-launcher tools across targets: Task (Claude), Agent/AgentSwarm (Kimi Code) — each credits its subagent_type/agent_type toward the same freshness evidence (src/runtime/activity.ts, src/freshness/agent-evidence-record.ts).
• File-size gate judges the real Edit outcome, not the on-disk count — an Edit on a file already over FUSE_SOLID_MAX_LINES used to be denied unconditionally (the on-disk count always won); the gate now computes the actual post-edit line count and allows a shrinking Edit while still denying a computable growth past the limit (src/policy/file-size.ts, sim scenario 33).
• Codex apply_patch fan-out to post-handlers — each file in a patch envelope becomes a synthetic per-file PostToolUse event (add→Write, update→Edit, delete no-ops on the Write/Edit-only handlers) instead of the whole-patch shape whose own filePath/content are always undefined (src/runtime/post-fanout.ts, test/apply-patch-post.test.ts).
• Non-blocking multi-file security advisory — the same PreToolUse "read the security skill" nudge a single Write/Edit gets is now evaluated per file across a Codex apply_patch envelope, firing on the first qualifying add/update code file (delete ignored outright) — additionalContext only, never a deny (src/runtime/lifecycle/security/check-skill.ts::securityAdvisoryForPatch, test/security-advisory-patch.test.ts).
• Codex SessionStart agent/command resync — the plugin agents/commands cache under ~/.codex/{agents,prompts} is re-materialized only when its sha256 fingerprint changed (or a command symlink dangles) since the last apply, otherwise a cheap no-op; agents are copied (Codex won't load symlinked TOMLs), commands are symlinked; a best-effort inter-process lock avoids a torn cache under concurrent session starts (src/runtime/lifecycle/codex-resync/, test/resync-agents.test.ts).
• Codex failure tracking inferred from PostToolUse — Codex reports every outcome, success or failure, on the same PostToolUse event; a failure is classified from an explicit signal only (non-zero exit, an error field — never "can't tell") and journaled into the existing one-shot failure tally (src/tracking/codex-post-failure.ts, test/codex-post-failure.test.ts).
• Shell-credited ref reads — a .md SOLID/skill reference read through a read-only shell command (cat/head/tail/sed/rg/less/more/bat, including inside a sh -c wrapper) now credits the same ref-read evidence a native Read would, so a Codex teammate's shell-first consultation still satisfies solidReadGate; sed -i/--in-place and any non-whitelisted command are never credited (src/policy/shell-read-refs.ts, test/shell-read-refs.test.ts).

Notification sounds

A native OS sound plays on two lifecycle events: the core-scope Stop under Codex (stopCore, since Claude Code already voices its own turn-end via a native afplay hook — this harness deliberately does not double it) and TeammateIdle under Claude Code when there's an actionable notice (a missing deliverable or a sniper reminder). On by default; opt out entirely with FUSE_HARNESS_SOUND=0.

Resolution per event kind (stop/permission/human) cascades: (1) a per-kind env override path (FUSE_HARNESS_SOUND_STOP / _PERMISSION / _HUMAN) if it exists on disk; (2) the package's own bundled assets/song/{finish,permission-need,need-human}.mp3 (located by walking up to the running package.json, so it survives the flat/hashed dist/ build); (3) $CLAUDE_PLUGIN_ROOT/song/<file>.mp3 as a last-resort fallback. No file resolves → silent no-op. The permission kind and its .mp3/env var exist in the module for symmetry but have no current call site — permission prompts stay native-only on every harness today.

Playback is native per platform (afplay on darwin, paplay on linux, PowerShell's SoundPlayer on win32) and absolutely fail-open: an unsupported platform, a missing player binary, an undecodable file, or a non-zero exit is swallowed — a broken or absent player can never break a hook (src/runtime/notifications.ts, src/runtime/notification-sound.ts, test/notifications.test.ts).

Environment

Var	Effect
FUSE_SOLID_MAX_LINES	SOLID file-size limit (default 100).
FUSE_HARNESS_REFS	Explicit path.delimiter-list of .md SOLID-reference dirs → activates solidReadGate. Overrides auto-discovery.
FUSE_HARNESS_MARKETPLACES	Comma-list of marketplace names whose solid-* skill refs are auto-discovered when FUSE_HARNESS_REFS is unset (default fusengine-plugins).
FUSE_ENFORCE_TTL_SEC	APEX freshness window in seconds (default 120).
FUSE_LESSONS_THROTTLE_MIN	Lessons-injection throttle, minutes (default 5).
FUSE_ENFORCE_GEMINI_MCP	Opt-in (default off). Blocks hand-written Tailwind UI (.tsx/.jsx/.vue/.svelte) until a mcp__gemini-design__* call is made this session. Read fresh per call (src/policy/gemini-mcp-gate.ts).
FUSE_DESIGN_GEMINI	Opt-in (default off), a different gate from the one above. Enables the design-pipeline's own Gemini gates (create_frontend validation + "generate before hand-writing HTML/CSS") — inert unless a design agent is active (src/policy/design/gates.ts:58-60, see docs/design.md).
FUSE_MCP_TTL_SEC	MCP (Context7/Exa) cache freshness, seconds (default 48h, src/runtime/mcp-key.ts).
FUSE_WEBFETCH_TTL_SEC	WebFetch cache freshness, seconds (default 24h — pages stale faster than docs).
FUSE_CONFIRM_SUBAGENT_WINDOW_SEC	G0 cool-down (seconds, default 300) for the CONFIRM <code> mechanism above — no token can be placed within this window of the last SubagentStart/Stop seen for the session.
RALPH_MODE	Opt-in (default off). Exempts safe git commands (add/commit/checkout -b/status/diff/log) from the confirmation ask and auto-approves project installs. Destructive git (force-push, reset --hard) and system installs still gate.
CLAUDE_PROJECT_DIR	Overrides the project root used to hash the out-of-tree state dir (src/runtime/paths.ts:20).
FUSE_HARNESS_SOUND	On by default. Set to 0 to disable every lifecycle notification sound.
FUSE_HARNESS_SOUND_STOP / _PERMISSION / _HUMAN	Per-kind override: an absolute path to a sound file to play instead of the bundled assets/song/*.mp3 for that kind (see Notification sounds).

Library usage

import { detectHarness } from "@fusengine/harness/detect";
import { evaluate } from "@fusengine/harness/policy";
import { gate } from "@fusengine/harness/runtime";

const { id, mode } = detectHarness();            // { id: "cursor", mode: "hook" }

// stateless guards (file-size, git, security, …)
const verdict = evaluate({ tool: "Write", filePath: "src/big.ts", content });
if (verdict.decision !== "allow") console.error(verdict.prompt?.reason);

// full gate (stateless + stateful APEX, fed from the session track)
const prompt = await gate({ sessionId, framework: "react", tool: "Write",
  filePath: "src/Button.tsx", content, now: Date.now(), trackFile });

The Prompt it returns ({ kind: "block" | "ask" | "inform", title, reason, actions? }) is portable; each adapter maps it to the harness's native shape — but, per the compatibility matrix above, not every harness can act on every kind.

Extend it

Add your own project rules without forking — they run after the privileged core chain (two-tier), and the chain is fail-closed (a guard that throws blocks, never silently passes):

import { registerGuard } from "@fusengine/harness/policy";

registerGuard(({ tool, command }) =>
  tool === "Bash" && command?.includes("kubectl delete")
    ? { kind: "ask", title: "Confirm cluster change", reason: command }
    : null);

Subpath exports

Subpath	What
./detect	detectHarness() / detectMode() — 13 harnesses, hook vs cli.
./policy	evaluate(ctx), the guard chain, evaluateApex, framework detection.
./runtime	handleHook, gate, recordActivity, activityFor, per-harness storage + MCP intercept.
./tracking	Session track: recordAgent/Doc/RefRead, agentsFresh, receipts, one-shot metric.
./refs	Frontmatter parse, loadRefs(dir), SOLID ref scoring/routing.
./prompt	The portable Prompt type + formatPrompt.
./cache	MCP/WebFetch cache: key, lookup/store, compaction, response extraction.
./memory	Per-project "never reproduce" lessons.
./config ./util ./state ./statusline ./freshness ./init ./cli	env config, project-root, locks, statusline, doc-freshness, wiring templates, staged checks.
./adapters/{claude,codex,cursor,cline,gemini,hermes}	Thin per-harness adapters — see Compatibility for what each can actually enforce.

Documentation

Guide	What
docs/index.md	architecture overview + map
docs/detect.md	harness detection (hook vs cli)
docs/policy.md	evaluate, the guard chain, framework, APEX gates
docs/guards.md	the guard chain, registerGuard, fail-closed
docs/runtime.md	handleHook, gate, tracking, MCP intercept, state paths
docs/config.md	env config (TTL, max-lines, refs dir, Gemini opt-ins)
docs/modules.md	cache · refs · state · memory · statusline · util
docs/adapters.md	adapters, compatibility, harness init/hook wiring
docs/design.md	design-agent pipeline — state machine, gates, opt-in Gemini
CHANGELOG.md	release history

Run bun run docs:api for the generated typedoc API reference.

Known limitations

• Codex apply_patch enforcement depends on Codex itself. The harness parses the patch and emits the right verdict per file, but Codex does not always enforce a correct deny on its end (openai/codex#27833, see Compatibility) — enforcement beyond the verdict is theirs.
• Cursor file edits are advisory-only. afterFileEdit fires after the edit already happened — a platform limit, not something this harness can work around.
• Hook fan-out is mitigated, not eliminated. The ~11-sibling-plugin burst is deduped within a 2s window (BURST_DEDUP_MS), but that window is a heuristic, not a protocol guarantee — an unusually slow fan-out could in theory land outside it.
• Sidechain hook reliability is a platform issue, worked around, not fixed. Sub-agent PostToolUse hooks don't always fire on Claude Code (documented platform issues #43612/#27655/#34692); SubagentStop transcript harvesting (src/freshness/evidence-harvest-io.ts) compensates, but only at that checkpoint, not continuously.
• Hermes coverage beyond pre_tool_call is unverified — no lifecycle events have been proven against a live Hermes install in this repo.
• CONFIRM <code> is not a security boundary. It guards against an ask being silently dropped by Codex/Kimi's degrade-to-deny, i.e. against precipitation and lost recourse — not against an adversarial agent. Any agent with shell access can write its own confirm token into session state and self-approve; this is true of every stateful gate this harness keeps outside a sandbox, not specific to this mechanism.

Develop

bun test            # 662 tests (120 files)
bunx tsc --noEmit   # typecheck (isolatedDeclarations)
bun run build       # dist + .d.mts via tsdown (for Node/bundler consumers)
bun run docs:api    # generate the typedoc API reference

CI runs test + typecheck on every PR. MIT licensed.