autonomy-operate

autonomy-operate — the rhythm

Read autonomy.config.json for this repo's knobs. Read STATE.md first every session.

The non-negotiables

1. The baton is truth. Work only when LOOP-STATE.md turn: is your role. Never infer the turn from the commit log.
2. Write everything down. STATE.md / LOOP-STATE.md / REVIEW-FEEDBACK.md hold the thread — chat compacts and is lost. Never rely on memory.
3. The gate, every wave: gate.test + the frozen invariant intact (gate.frozenInvariant) + gate.build + gate.lint on touched files. If gate.coverage is set, also run node ${CLAUDE_PLUGIN_ROOT}/hooks/coverage-ratchet.mjs after the coverage command: a coverage drop below the .autonomy-coverage.json floor is a RED gate (the third gate, the drift guard). If gate.patchTarget is greater than 0, also run node ${CLAUDE_PLUGIN_ROOT}/hooks/patch-coverage.mjs --threshold=<patchTarget> so the wave's own changed lines are tested (the fourth gate; the reviewer passes --base=<last-reviewed-sha>). All green or REVERT — never commit red. If gate.envFidelity is set, the LOCAL gate must match CI (an env-naked pass is not proof CI is green).
4. Honesty mandate (honestyRule): no fabricated numbers; every rate carries N + CI or "building — N/30"; a capability with no real data ABSTAINS visibly — never faked, never silently dead.
5. Gate List (gateList): both terminals STOP and ask the human for risky/irreversible actions. Never self-authorize one — escalate to FOR-REVIEW.md, set turn: human.
6. Push the work branch only, never prod. (The gate-guard hook enforces this; don't fight it.)

Builder rhythm

Two modes — IDEATE (research lane) and BUILD. ultrathink the plan. Smallest high-signal diff; PURE module + RED-before-green test for new logic; zero deletions on shared files; re-read fresh. Pre-empt the reviewer by answering the 5-lens rubric (Correctness · Honesty · Regression/frozen-drift · Security · UX) in your build note, with a "red-team the opposite" paragraph.

Reviewer rhythm (5-lens panel)

ultrathink every review. Spawn 5 critic subagents (math · honesty · regression+frozen-drift · security · UX); unanimous PASS required. Red-team-the-opposite before any PASS. Bite-check the regression test (reintroduce the bug → RED → restore). Frozen-drift/protected-path → Builder→Critic→Judge debate, PARK to FOR-REVIEW. Fix what you safely can; flag the rest with severity.

Idle behavior — research & ideate, do NOT stand down

An empty bug/feature backlog is NOT a reason to set turn: human. When pending-for-builder is empty, the backlog is drained, and everything left is owner-gated, the builder runs the Research & Ideation lane (tasks/RESEARCH-LANE.md), it does not stand down:

• Start the next dated research cycle — one substantive, sourced theme per cycle (NEVER per-commit busywork; that regressive pattern is rejected). Honesty mandate holds: every stat carries N + CI or says "building".
• Write new feature/idea proposals to the ideas file (tasks/IDEAS.md): problem · sourced evidence · proposed feature · anchor files · risk tier · acceptance · gate plan.
• BUILD the safe additive (non-protected-path) winners behind the full gate → commit → push the work branch.
• PARK gated / irreversible / new-infra / strategic ideas to FOR-REVIEW.md as an approval-menu entry the human picks from — and KEEP GOING (next non-gated wave, else the next research cycle). Gated items are parked, never stops.

Set turn: human ONLY when truly blocked on all fronts at once (every remaining path needs a human answer AND research is genuinely dry) or the human explicitly pauses the loop — that should be rare. The reviewer mirrors this: never write "stand down" into pending-for-builder; point the builder at the next wave, else the research lane.

Durable learnings (append "what almost broke + the rule that caught it")

These generalized out of a real run — keep adding:

• A merge/replace of an existing function can silently SHRINK coverage — diff the old behavior (caps, overflow, edge branches) against the new; "refactor, output identical" must be PROVEN, not claimed.
• A bite that produces no RED is a no-op — if reintroducing the bug doesn't fail a test, the test isn't covering it. Re-verify with a clean, line-targeted edit.
• Securing a previously-open endpoint can break its callers — grep every caller before adding auth; a green test suite won't catch a cross-file UI regression.
• Units/scale mismatch at an API boundary is a silent killer — test the conversion against the consumer's validator, not just the arithmetic.
• "Renders only when present" must be strictly APPEND-ONLY — the empty branch must emit the ORIGINAL bytes; don't leave a structural line that lingers when the value is absent.
• Encoding into a shareable/indexable URL? Draw the PII boundary at the encode function — assert the EXCLUSION in a test, not just the inclusion.
• An env-naked local gate ≠ CI — if a guarded route or env-dependent code is touched, gate in BOTH env states (or wire gate.envFidelity so it's automatic).
• A deliberate deviation from a handed-down spec is fine when it's RIGHT — but surface it loudly (commit body + handoff) so the reviewer weighs the one judgment call instead of diffing for it.
• ripgrep silently skips files it deems "binary" — for a large generated/minified-looking source, use a binary-safe grep -a before concluding a symbol is unused.