agent-output-audit

Agent Output Audit

You are the independent evaluator. Answer one question — "Did the implementing agent actually do what task_NN.md says it did?" — from files, public behavior, tests, and CI. A self-report is not evidence. (Whether a real user can succeed at the product is qa-execution; run both on a Compozy slug and keep their outputs separate.)

Step 1: Discover the Repository Verification Contract

1. Read root instructions, repository docs, and CI/build files before running commands.
2. Run python3 scripts/discover-project-contract.py --root . to surface candidate install/verify/build/test/lint/start commands and E2E signals.
3. Prefer repository-defined umbrella commands (make verify, just verify, CI entrypoints) over language defaults. When discovery surfaces more than one plausible gate or mixes ecosystems, read references/project-signals.md before choosing, and state the tie-breaker.
4. Read references/e2e-coverage.md before classifying any flow's coverage.
5. Resolve the audit artifact directory: the audit-output-path argument if given, else repository conventions, else /tmp/agent-output-audit-<slug>. Create its audit/ subdirectory; store all bugs and reports under <audit-output-path>/audit/.
6. Detect Compozy mode. If .compozy/tasks/<slug>/ exists, record the slug and:
   • Read state.yaml read-only — scripts/update-state.py owns its mutation per the cy-codex-loop contract.
   • Read _techspec.md (deliverable source of truth) and _tasks.md (task roster) when present.
   • List every task_NN.md and capture its frontmatter status: (pending | in_progress | completed). When frontmatter disagrees with state.yaml, frontmatter is the source of truth.
   • Note the memory slot .compozy/tasks/<slug>/memory/qa-execution.md — Step 5 writes it before any status flip.

Step 2: Run the Baseline Verification Gate

1. Install dependencies with the repository-preferred command.
2. Run the canonical gate once before any audit work, fastest-first: lint and type-check → build → unit tests → integration tests.
3. If the E2E command is separate from the umbrella gate, decide whether to run it now or after runtime prerequisites are ready, and record that plan.
4. On a baseline failure, read the first failing output and determine whether it is pre-existing or introduced by current work. Exclude a failure from audit scope only after a clean reproduction proves it unrelated.
5. Flaky-failure protocol. Before classifying any baseline failure, run the failing test in isolation 3-5 times on the same SHA. If it passes at least once without a code change, record it as flaky-suspect in the SUITE HEALTH SNAPSHOT (test name, attempts, retry outcome, suspected category) rather than promoting it to PASS. Read references/flaky-triage.md before assigning a suspected category or proposing a quarantine.

Step 3: Audit Task Implementations

Skip this step only when no task, phase, PRD, tech spec, or implementation-plan artifacts exist.

1. Read references/independent-evaluator-protocol.md in full before forming any verdict — it owns what does and does not count as evidence, and the transcript classification (genuine-failure / grader-bug / ambiguous-task / bypass-exploit). In Compozy mode, read the implementer's memory/<phase>.md artifacts and record anomaly classifications in memory/qa-execution.md → Errors / Corrections before judging the task.
2. Summarize each task_NN.md and its body into a Task Implementation Matrix (columns mirror cy-codex-loop frontmatter):
   • task_path, declared_status (literal frontmatter status:)
   • title, type, complexity, dependencies — mirrored from frontmatter
   • techspec_deliverable — linked _techspec.md section when present
   • Requirements, subtasks, checklist items, success criteria, dependent files
   • implementation_evidence — files, modules, routes, commands, migrations, seeds, tests
   • verification_evidence — commands executed, exit codes, output summaries
   • qa_verdict — PASS | PARTIAL | FAIL | REOPEN | BLOCKED (distinct from declared_status)
   • ai_audit_findings — red flag IDs that fired in Step 4 with verdict
   • action — none | fixed | reopened-frontmatter | BUG-NNN.md filed
   • linked_bugs — BUG IDs
3. Verify every completed or claimed-complete task against actual files, public behavior, automated tests, and acceptance criteria. Re-execute the smallest public proof against the current repository state.
4. Assign qa_verdict:
   • PASS: every material requirement and success criterion has implementation and fresh verification evidence.
   • PARTIAL: implementation exists but one or more non-critical requirements, tests, or evidence are missing.
   • FAIL: claimed behavior does not work or a critical requirement is absent.
   • REOPEN: frontmatter says status: completed but the QA verdict is PARTIAL or FAIL.
   • BLOCKED: a concrete prerequisite is missing. Validate every local boundary that does not need the missing dependency and report the blocked live validation separately.

Step 4: AI Test-Hygiene Scan (RF-1..RF-6)

1. Read references/ai-implementation-audit.md in full before scanning the test diff of any task with declared_status: completed — it owns the RF-1..RF-6 scanners, the Requirement→Test mapping, and the verdict matrix.
2. Run the scans against the diff since the task baseline (git log --follow <test_file>, git diff <baseline_sha>..HEAD).
3. Emit the verdict the matrix assigns. RF-1 (skip/only/xit/t.Skip inserted), RF-2 on a P0/P1 criterion (weakened assertion), RF-3 (mock on a dependency the TC declared Integration/E2E), and RF-4 on P0/P1 (unjustified snapshot drift) are automatic FAIL.
4. Record findings in the matrix column ai_audit_findings and in the per-task block of audit-report.md.
5. Apply the Requirement→Test mapping: for every Success Criterion in task_NN.md and every linked _techspec.md bullet, mark the matching test covers / weak / missing. A checked item or status: completed without a covers row is an audit failure.

Step 5: Reopen, File Bugs, Write Memory

1. Mark every incomplete completed task REOPEN.
2. In Compozy mode, write memory/qa-execution.md with the cy-codex-loop canonical sections (Objective Snapshot, Important Decisions, Learnings, Files / Surfaces, Errors / Corrections, Ready for Next Run) before flipping any task_NN.md frontmatter (memory-precedes-status invariant).
3. Edit the offending task_NN.md frontmatter status: back to pending (or in_progress if salvageable). Leave state.yaml alone — update-state.py owns it, and the next iteration reconciles from frontmatter.
4. File BUG-<num>.md under <audit-output-path>/audit/issues/ using assets/issue-template.md, including: the task path (Reopens task:), the failed Success Criterion (Summary:), the original strict assertion when RF-2 fired (Root cause:), the red flag ID and verdict (Automation Follow-up:), and any transcript anomaly classification (Related:).
5. When the gap is a bounded root-cause fix inside the audit scope, implement it, add regression coverage, and rerun the task proof. Otherwise reopen the task.

Step 6: Quality Gates Verdict

1. Re-run the canonical verification gate from scratch after the last code change made during the audit.
2. Compile the Quality Gates section of audit-report.md, each PASS / FAIL / N/A:
   • Flaky rate <2% in the canonical suite.
   • Zero FAIL from the AI test-hygiene scan on P0/P1 tasks.
   • Zero Critical / High issues open.
   • Coverage delta ≥ baseline (no regression).
   • Zero unresolved flaky-suspect on P0 flows.
3. A FAIL on any gate blocks an unconditional PASS verdict for the run.

Step 7: Write the Audit Report

1. Write the report to <audit-output-path>/audit/audit-report.md using assets/audit-report-template.md, with all mandatory sections:
   • Claim / Command / Exit code / Verdict per command executed in Steps 2 and 6.
   • AUTOMATED COVERAGE — support detected, harness, canonical command, required flows with classification, specs added or updated.
   • TASK IMPLEMENTATION AUDIT — Compozy slug, plan sources, matrix totals, per-task verdicts, reopened/fixed/blocked tasks, links to bugs.
   • SUITE HEALTH SNAPSHOT — flaky rate, flaky events, mutation score (when a harness exists), coverage delta vs baseline, blocked count, manual-only count, AI audit findings count.
   • QUALITY GATES — PASS/FAIL/N/A per gate.
   • ISSUES FILED — total, by severity, with Reopens task: annotations.
   • Report each blocked scenario, missing credential, or environment gap with the exact command or prerequisite that stopped execution.
2. In a Compozy slug, a final PASS feeds cy-codex-loop's verify.last_status=PASS precondition for Phase E — leave update-state.py to cy-codex-loop.
3. Before declaring the audit complete, confirm every item in references/checklist.md — it is the exhaustive completion criterion across all steps.