audit

name	audit
description	Use when a loen loop stage — plan, act, check, or result — must be validated and gated before the next one. Mode-aware for delivery/repair/research; the execution-loop analog of check-chain.

loen:audit — loop stage validator + live report

Invoke as loen:audit <stage> where stage ∈ plan | act | check | result. Read the active run from docs/loen/current and mode from its loop.yaml (delivery | repair | research). Every stage returns a verdict OK / needs_work, gates the next stage, and regenerates docs/loen/<run-id>/report.html (via the html-report skill) plus appends to state.md.

Stage checks (all modes)

• plan — loop.yaml parses; objective measurable; mutable_scope/protected_scope non-empty and disjoint; quality_gates non-empty; budget present; human approval recorded. verifier_isolation, when present, MUST be subagent or microvm — validate it AND the host capability with this plugin's scripts/verify_microvm.sh preflight <run-dir>/loop.yaml (resolved from the skill base dir): microvm on a host without KVM/Firecracker/images → needs_work with the script's "install microVM support or drop to verifier_isolation: subagent" hint. No silent downgrade at plan time. needs_work blocks Act.
• act — the latest iterations/iter-NN/diff.patch exists and touches only mutable_scope; no protected_scope path present (cross-check with this plugin's scripts/guard_protected.sh via the run's loop.yaml, resolved from the skill base dir); and the run dir passes this plugin's scripts/check_layout.sh — the deterministic net that catches any Bash-written non-canonical artifact that bypassed the PreToolUse hook.
• check — dispatch the verifier per the contract's verifier_isolation key (subagent when absent). subagent → dispatch the verifier subagent (isolated), exactly the MVP path. microvm → write the mode-specific checklist (the same text the subagent dispatch prompt would carry) to a temp file OUTSIDE the run dir, then run this plugin's scripts/verify_microvm.sh check <run-dir> <iter-NN> <checklist-file> (resolved from the skill base dir): it snapshots the tree, runs the verifier headless inside an iclaude Firecracker microVM, and writes the returned text to iterations/iter-NN/verifier.md unchanged — downstream consumers are agnostic to where the verdict was produced. A non-zero exit (VM boot/provision failure, silent host fallback, missing verdict, host-tree tripwire) → verdict needs_work quoting the script's failure log path — NEVER fall back to the in-session subagent; the human may edit the contract to subagent to proceed un-isolated. In both dispatch modes the verdict lands at iterations/iter-NN/verifier.md; confirm gates.log shows the gates ran. OK iff the verdict is APPROVE and gates are green.
• result — every plan step is done, gates green, verifier APPROVE across the final iteration. On OK: finalize report.html, ensure pr-summary.md exists, and mark the docs/TODO.md row (Result: OK, Status: done, Closed: <today>) keyed by <topic>.

Mode: repair — additional checks

• plan — quality_gates include the failing command recorded in the Baseline section of state.md at bootstrap.
• check — when the diff claims a NEW/extended regression test, gates.log carries the worker's logged inversion evidence (stash the fix → the regression test FAILS → pop → it PASSES).
• result — regression coverage evidenced in the final diff: case (a) a new/extended test, or case (b) the originally-failing test itself — the verifier states which case applies; the originally-failing command exits 0; the diff is minimal (every non-test hunk required for that command to pass).
• verifier dispatch prompt carries the repair checklist: regression coverage case (a/b), validation of the LOGGED inversion evidence, minimal-diff confirmation. The verifier stays read-only — it never mutates the tree.

Mode: research — additional checks

• plan — eval_command non-empty; EXACTLY ONE metrics.primary entry of the form <name>:max or <name>:min; stop_conditions carry one target: <primary-name> <op> <number> line (<op> ∈ >=/<= matching the direction) and only well-formed tolerance: <name> regression <= <number>[%] lines; protected_scope COVERS the eval assets (the eval_command script, eval datasets, ground truth) — non-empty alone is not enough.
• check — gates green (they run on every experiment); iterations/iter-NN/metrics.jsonl has exactly one summary line; experiments.jsonl has this iteration's record; for every keep decision the verifier RE-RUNS eval_command — exporting LOEN_METRICS_PATH to a throwaway temp path, never appending to canonical artifacts — and confirms the claimed delta. revert records are trusted as logged.
• result — kept changes are metric-backed (primary improved in its declared direction, secondaries within their stated tolerances) OR the budget is exhausted with a best-result report; the experiments.jsonl stream is consistent end-to-end with the per-iteration metrics.jsonl summaries.
• verifier dispatch prompt carries the research checklist: delta re-check via a throwaway LOEN_METRICS_PATH, stream cross-check (experiments.jsonl vs metrics.jsonl), protected eval assets untouched. The verifier stays read-only.

report.html (every stage)

Invoke the html-report skill targeting docs/loen/<run-id>/report.html with: the contract (loop.yaml), an iterations table (diff summary, gates pass/fail, verifier verdict), metrics before/after — in research mode an experiments table (hypothesis, before/after, delta, decision) — budget spend, current stage/verdict, and handoff reasons. Self-contained, opens by double-click.

Rules

• Never edit the diff you are judging. Never weaken a gate to pass.
• All writes land at canonical docs/loen/<run-id>/ paths (the loop-guard hook enforces this); the report is report.html, nothing else.