/audit "<entry-point>"

Runs the adversarial audit pipeline. Accepts any entry point: feature slug, file list, spec reference, or prior audit report path. Finds bugs, proves them with failing tests, fixes the code, and leaves the codebase clean.

Subagent contract — MANDATORY for every dispatch

Every subagent this skill dispatches MUST be given the following preamble at the top of its prompt — no exceptions:

Subagent contract: Honor rules/completeness-contract.md (load-bearing — no silent deferrals; trigger phrases = escalation signals, not completion modes). If you cannot complete assigned scope, escalate via AskUserQuestion with user-validatable proof. A return claiming COMPLETE alongside deferred items is a contract violation.

This applies to Classification, Exploration, Suspect, Prove, Fix, Regression, Report, Reconciliation, and any other subagent dispatched by /audit.

When any subagent returns, the orchestrator MUST run the validation script BEFORE accepting the return:

mkdir -p /tmp/vallorcine
return_file=/tmp/vallorcine/audit-return-"<job-name>".txt
printf '%s\n' "$FULL_RETURN_TEXT" > "$return_file"
bash .claude/scripts/validate-subagent-return.sh "$return_file" 2>/tmp/vallorcine/validator-stderr.txt
rc=$?

• rc=0 → accept the return.
• rc=1 → trigger phrase detected (sub-agent shipped partial work as complete). Surface to user via AskUserQuestion with validator stderr. Do not advance to the next job until resolved.
• rc=2 → tooling error. Log and treat as rc=0.

Orchestrator discipline — MANDATORY

You are a state machine. Your ONLY job is to dispatch subagents serially and relay their one-line summaries. You use TodoWrite to show progress.

What you may do:

• Check file existence (does audit-report.md exist? yes/no)
• Read one-line return summaries from subagents
• Parse counts from return summaries (cluster count, finding count, lens count)
• Display progress via TodoWrite
• Decide which subagent comes next based on return summaries
• Run mechanical pipeline scripts (reconcile-cards.py, extract-views.py) — these are data transformations, not analytical work

What you must NEVER do — these are HARD RULES, not suggestions:

• DO NOT read any pipeline output file (scope-definition.md, cluster--packet.md, suspect-.md, prove-fix-.md, audit-report.md, exploration-decisions.jsonl, construct-cards.yaml, analysis-cards.yaml, lens--cards.yaml, active-lenses.md)
• DO NOT read source code files
• DO NOT run build/test commands (./gradlew, npm test, cargo test, etc.)
• DO NOT run git commands (git diff, git log, etc.)
• DO NOT parse XML test results, grep test output, or count test files
• DO NOT write any pipeline output file
• DO NOT make analytical decisions about scope, bugs, tests, or fixes
• DO NOT do the work of ANY subagent yourself, even if it seems simple

How to detect a violation: If you are about to use Bash, Read, Write, Edit, or Grep for anything other than checking file existence or reading a return summary, STOP. That work belongs in a subagent.

Why: Every tool call result accumulates in your context window. You exist for the entire pipeline. A single file read adds ~500 tokens that you carry through every subsequent turn. Subagents absorb work in isolated context windows that are discarded after they return.

Pipeline overview

Five jobs, executed serially. All subagents run one at a time. Mechanical scripts run between subagent phases (no LLM cost).

Scope:      Classification → Exploration → Card Construction
            (interactive)    (discovery)    (assertion sweeps)
                                  ↓
            [reconcile-cards.py] → [extract-views.py detect]
                                  ↓
            Domain Pruning → [extract-views.py project] → Assembly
            (lens challenge)                               (domain-lens clustering)
                                  ↓
Suspect:    Lens1/C1 → Lens1/C2 → ... → Lens2/C1 → ... → LensK/CN
                                  ↓
Prove-Fix:  Finding 1 → Finding 2 → ... → Finding M
            (sequential — each finding: write test → confirm/impossible → fix)
                                  ↓
Report:     [aggregate-results.py] → Single subagent → audit-report.md + audit-prior.md
            (pre-aggregate, then cross-domain finding combination)
                                  ↓
Reconcile:  Single subagent → spec-updates.md + kb-suggestions.md
            (optional — only when .spec/ exists)

Determine entry point

Parse the argument to determine what kind of audit this is:

• Feature slug (e.g., "float16-vector-support"): look for .feature/<slug>/ directory
• File path or glob (e.g., "src/index/*.java"): direct file audit
• Spec reference (e.g., "spec:F01"): audit code against spec
• Prior report (e.g., "audit-report.md" path): incremental audit

If the argument is ambiguous, ask the user using AskUserQuestion:

• Question: "What would you like to audit?"
• Options:
  1. "Feature" — audit a feature by slug (you'll ask for the slug next)
  2. "Specific files" — audit files by path or glob pattern
  3. "Spec coverage" — audit code against a spec (provide spec ID)
  4. "Prior report" — incremental audit from a previous report

Store the entry point type and value for the Classification subagent.

Spec-state gate (spec: entry points only)

Before proceeding, verify the target spec is APPROVED. /audit on a DRAFT or INVALIDATED spec cannot do conformance checking — there is no authoritative contract to prove code against. Adversarial exploration alone would produce misleading findings that the user will interpret as conformance bugs.

# Only for "spec:" entry points. Skip for feature / file / prior-report.
bash .claude/scripts/audit-state-gate.sh "<spec-id>"

If the script exits non-zero, stop immediately. It prints a clear error message and the recommended next step (usually /spec-verify to promote DRAFT → APPROVED). Do not ask the user to override — a non-APPROVED spec makes /audit structurally unsound for this entry point.

For feature-slug or file entry points, do not run this check — those entry points don't target a specific spec, and if specs are resolved during classification, that phase handles state awareness.

Budget

After determining the entry point, ask the user for a budget using AskUserQuestion:

• Question: "What's your budget? Budget is split proportionally: ~30% discovery, ~45% prove-fix, ~15% reporting. Discovery costs ~$8-12 per file; prove-fix costs ~$5 per finding."
• Options:
  1. "$300 (default)" — full scope for most features
  2. "$150" — may narrow to 4-5 files on larger features
  3. "$80 (focused)" — 2-3 files, ~7 findings
  4. "Unlimited" — no cap on spending

The user can type a custom dollar amount via "Other". If the custom amount is below $50, display a note: "Budgets under $50 typically cover only 1-2 files. Consider $80+ for meaningful results." Still allow the user to proceed.

• If the user picks a dollar amount: store as AUDIT_BUDGET
• If the user picks "Unlimited": leave AUDIT_BUDGET empty

Budget allocation model (computed when AUDIT_BUDGET is set):

DISCOVERY_ALLOC  = AUDIT_BUDGET * 0.30
SUSPECT_ALLOC    = AUDIT_BUDGET * 0.10
PROVE_FIX_ALLOC  = AUDIT_BUDGET * 0.45
POST_RESERVE     = max(AUDIT_BUDGET * 0.15, 10)

The budget is a planning target, not a hard wall. Each phase scopes its work to fit its allocation. Post phases (test cleanup, report, reconciliation) always run regardless of budget. See "Budget control" in Job 3 for the soft cap mechanism.

Resume detection

Before starting any work, run the state management script to determine pipeline state. This is mechanical, not LLM judgment.

For new audits

After determining the entry point, locate or initialize the audit state:

# Locate the audit directory
AUDIT_DIR=$(python3 .claude/prompts/audit/audit-state.py locate <type> <value>)

# Check for existing state
python3 .claude/prompts/audit/audit-state.py resume "$AUDIT_DIR"

The resume command outputs one of:

• RESUME_AT=init — no prior runs, start fresh
• RESUME_AT=new_run — prior run is complete, start a new run
• RESUME_AT=<step> + RUN_DIR=<path> — interrupted run, resume at step

When starting fresh (init or new_run)

python3 .claude/prompts/audit/audit-state.py init <type> <value>

This creates the run directory and state.json. Parse INIT_RUN_DIR from the output — all pipeline artifacts for this run go in that directory.

When resuming

If resuming at prove-fix, run extract-findings.sh first to get the accurate remaining count (the script filters out findings with existing output files):

bash .claude/scripts/extract-findings.sh <RUN_DIR>

Parse the "N remaining" count from the script output. Display to the user with the actual remaining count:

Resuming audit — picking up at <RESUME_AT>.
Run directory: <RUN_DIR>
<N already processed>, <M remaining> findings to prove-fix.

Use AskUserQuestion to confirm:

• "Resume" — process the remaining findings
• "Restart fresh" — discard this run and start a new audit
• "Cancel" — stop

Initialize TodoWrite with completed steps already checked. Start from the resume step.

After each step completes

Update state immediately:

python3 .claude/prompts/audit/audit-state.py complete "$AUDIT_DIR" <step> [key=value ...]

Include metadata when available:

• After assembly: clusters=9 lenses=4
• After suspect: findings=39
• After prove-fix: fixed=20 impossible=10 fix_impossible=0 deferred=0

All pipeline artifacts go in the run directory (RUN_DIR), not in .feature/<slug>/ directly. This keeps runs isolated — no stale files from prior runs can confuse the pipeline.

Progress tracking

Before launching the first subagent (or after resume detection), initialize the TodoWrite checklist:

◻ Step 1.1: Classification
◻ Step 1.2: Exploration
◻ Step 1.3: Card Construction
◻ Step 1.4: Reconciliation + Lens Detection
◻ Step 1.5: Domain Pruning
◻ Step 1.6: View Projection
◻ Step 1.7: Assembly
◻ Job 2: Suspect
◻ Job 2b: Pre-prove gates
◻ Job 3: Prove-Fix
◻ Job 3b: Test Cleanup
◻ Job 4: Report
◻ Job 5: Reconciliation (if .spec/ exists)

Update rules:

• Steps 1.1–1.7: Mark each step complete immediately when it finishes. Do NOT mark the next step complete until it actually finishes.
• Jobs 2–3: Update the label with running counts as subagents return. Examples: "Suspect (3/8 clusters, 12 findings)", "Prove-Fix (5/12 findings, 4 fixed, 1 impossible)". Mark complete only when all subagents for that job finish.
• Job 4: Mark complete when Report returns.
• Job 5: Only shown if .spec/ exists. Mark complete when Reconciliation returns.

Job 1: Scope

Step 1.1 — Classification subagent

Launch a subagent. Tell it:

You are the Classification subagent for an audit pipeline. Read .claude/prompts/audit/classification.md for your complete instructions.

Entry point: <type> = <value> Working directory: <cwd>

Resolve scope deterministically. Do NOT ask the user questions. Write your outputs and return a summary.

The subagent will:

• Resolve the entry point to initial file paths
• Detect prior audit reports and load prior work
• Gather specs, KB entries, ADRs
• Detect language and project structure
• Write classification.md with scope summary

Expected return format: "Classification complete — <n> initial files, <n> specs resolved, prior_round=<yes|no>, language=<lang>"

After Classification returns, apply the budget scope gate:

Parse the file count from the return summary. If AUDIT_BUDGET is set, compute the discovery allocation and maximum affordable files:

max_files = floor(DISCOVERY_ALLOC / 10)   # ~$10 per file

Case 1: Budget fits the scope (max_files >= file_count or no budget):

Display the return summary and use AskUserQuestion:

• Question: "Audit scope: <n> files, budget $<budget>. Estimated total: ~$<file_count * 30> (within budget)."
• Options:
  1. "Proceed" — explore all files
  2. "Adjust scope" — change which files to include
  3. "Cancel" — stop the pipeline

Case 2: Budget requires narrowing (max_files < file_count):

The orchestrator narrows scope automatically to the top max_files files from classification's priority-ordered list. Display:

Audit scope: <n> files, budget $<budget>
  Budget covers ~<max_files> of <n> files for full analysis.
  Focusing on the <max_files> highest-risk files:
    1. <file> — <rationale from classification>
    2. <file> — <rationale>
    ...
  Remaining <n - max_files> files deferred to a follow-up round.

Use AskUserQuestion:

• Question: (the text above)
• Options:
  1. "Proceed" — explore the selected files (recommended)
  2. "Override" — choose different files to include
  3. "Raise budget to $<suggested>" — increase budget to cover all files
  4. "Cancel" — stop the pipeline

If "Override": ask which files to include/exclude and adjust. This is the escape hatch for when the user knows something classification doesn't.

If "Raise budget": update AUDIT_BUDGET and recompute allocations. Proceed with the full file list.

Passing scope to Exploration: When scope was narrowed, pass ONLY the selected files to the Exploration subagent (not the full list). Include a note in the subagent prompt: "Scope narrowed to <n> of <total> files for budget. Explore only these files."

Mark Step 1.1: Classification complete.

Step 1.2 — Exploration subagent

Launch a subagent. Tell it:

You are the Exploration subagent for an audit pipeline. Read .claude/prompts/audit/exploration.md for your complete instructions.

Read the classification output at .feature/<slug>/classification.md for your starting files, prior work summary, and context package.

Write exploration-decisions.jsonl as you work — every tiering decision, clearing check, and frontier stop must be logged.

Return a summary line when done.

Expected return format: "Exploration complete — <n> analyze, <n> boundary, <n> ignore, <n> domain signals, frontier=<n> stops"

Mark Step 1.2: Exploration complete.

Step 1.3 — Card Construction subagent

Launch a subagent. Tell it:

You are the Card Construction subagent for an audit pipeline. Read .claude/prompts/audit/card-construction.md for your complete instructions.

Read the exploration output at .feature/<slug>/exploration-graph.md for the construct list, locations, and boundary contracts. Read the classification context at .feature/<slug>/classification.md.

Build construct cards using assertion-first sweeps.

Return a summary line when done.

Expected return format: "Cards built — <n> constructs, <n> batches, <n> with assumptions, <n> with empty contracts"

Mark Step 1.3: Card Construction complete.

Step 1.4 — Reconciliation + View Detection (mechanical scripts)

Run the reconciliation script to add incoming edges and derive co_mutators/co_readers:

python3 .claude/prompts/audit/reconcile-cards.py .feature/<slug>/

Expected output: "Reconciled <n> cards — <n> inconsistencies, <n> with co_mutators"

Then run view extraction phase 1 to detect candidate domain lenses:

python3 .claude/prompts/audit/extract-views.py detect .feature/<slug>/

Expected output: "Detected <n> candidate lenses, <n> eliminated — active: <lens names>"

Mark Step 1.4: Reconciliation + Lens Detection complete.

Step 1.5 — Domain Pruning subagent

Launch a subagent. Tell it:

You are the Domain Pruning subagent for an audit pipeline. Read .claude/prompts/audit/domain-pruning.md for your complete instructions.

Read .feature/<slug>/active-lenses.md for candidate lenses. Read .feature/<slug>/exploration-graph.md for domain signal corroboration.

Challenge each candidate lens. Update active-lenses.md with CONFIRMED or PRUNED status.

Return a summary line when done.

Expected return format: "Domain pruning — <n> confirmed, <n> pruned: <pruned names or 'none'>"

Mark Step 1.5: Domain Pruning complete.

Step 1.6 — View Projection (mechanical script)

Run view extraction phase 2 to produce per-lens card projections:

python3 .claude/prompts/audit/extract-views.py project .feature/<slug>/

Expected output: "Projected <n> lenses, <n> total card projections — <details>"

Mark Step 1.6: View Projection complete.

Step 1.7 — Assembly subagent

Launch a subagent. Tell it:

You are the Assembly subagent for an audit pipeline. Read .claude/prompts/audit/assembly.md for your complete instructions.

Read the reconciled cards at .feature/<slug>/analysis-cards.yaml, the confirmed lenses at .feature/<slug>/active-lenses.md, the per-lens projections at .feature/<slug>/lens-*-cards.yaml, and the classification context at .feature/<slug>/classification.md.

DO NOT read source code. Work only with structured data.

Return a summary line when done.

Expected return format: "Assembly complete — <n> lenses, <n> total clusters, <n> unique constructs clustered, <n> constructs in multiple lenses, <n> excluded"

Parse the lens count and total cluster count from the return. You need these to dispatch Suspect subagents.

Cost checkpoint: If AUDIT_BUDGET is set, run:

bash .claude/scripts/audit-budget.sh

Display to the user:

Discovery complete — $<cost> of $<budget> spent
  <n> clusters across <n> lenses identified
  Remaining: ~$<budget - cost - POST_RESERVE> for suspect + prove-fix

Mark Step 1.7: Assembly complete.

Job 2: Suspect

For each cluster packet (organized by lens), launch a subagent serially. The Assembly return tells you the total cluster count across all lenses. Cluster packets are named cluster-<lens>-<N>-packet.md.

For each cluster, tell the subagent:

You are the Suspect subagent for <lens>/cluster <N>. Read .claude/prompts/audit/suspect.md for your complete instructions.

Your cluster packet is at .feature/<slug>/cluster-<lens>-<N>-packet.md. It contains everything you need — construct cards, domain lens, analysis guidance, boundary contracts, specs, KB. Read source files directly using the file paths and line ranges from the construct cards.

Your analysis is scoped to the <lens> domain. Focus your attack reasoning on <lens> concerns as described in the packet's domain guidance section.

DO NOT read any file not referenced in your construct cards. DO NOT query specs, KB, or ADRs — they are embedded in your packet.

Return a summary line when done.

Expected return format: "Suspect <lens>/C<N> — <n> findings, <n> cleared, <n> boundary observations, <n> card-driven"

After each subagent returns:

1. Parse finding count from return
2. Accumulate total findings across all clusters and lenses
3. Update Job 2 label: "Suspect (<completed>/<total> clusters, <accumulated> findings)"

After all Suspect subagents complete:

Cost checkpoint: If AUDIT_BUDGET is set, run:

bash .claude/scripts/audit-budget.sh

Display to the user:

Suspect complete — $<cost> of $<budget> spent
  <n> findings to prove-fix (~$<findings * 5> estimated)
  Prove-fix capacity: ~<floor((budget - cost - POST_RESERVE) / 5)> of <n> findings within budget

Mark Job 2: Suspect complete. If zero total findings, skip to Job 4 (Report).

Job 3: Prove-Fix

Process findings ONE AT A TIME, sequentially. Each finding gets its own subagent with a fresh context window. The subagent writes a test to prove the bug, then fixes the source code if confirmed.

Build the finding list

After Suspect completes, run the finding list extraction script:

bash .claude/scripts/extract-findings.sh .feature/<slug>/audit/<run-dir>

This produces finding-list.txt — one line per finding, pipe-delimited: <finding-id>|<title>|<construct>|<lens>|<cluster>|<suspect-file>

Pre-prove gates (mechanical scripts — no LLM cost)

Run two scripts to optimize the dispatch queue before prove-fix:

# Gate 1: Reorder findings so cross-lens duplicates run after their primary
python3 .claude/prompts/audit/dedup-findings.py .feature/<slug>/audit/<RUN_DIR>/

# Gate 2: Check for existing adversarial tests that cover findings
python3 .claude/prompts/audit/check-test-coverage.py .feature/<slug>/audit/<RUN_DIR>/ .

Dedup identifies behavioral duplicates across domain lenses and reorders the dispatch queue so the primary finding runs first. After the primary confirms and fixes, Phase 0 catches duplicates in ~3 turns instead of ~35. No findings are skipped. If dispatch-order.txt exists, use it as the dispatch order instead of the default lens/cluster ordering.

Coverage check matches findings against existing adversarial tests from prior audit rounds. Covered findings get stub prove-fix-*.md files with Result: COVERED_BY_EXISTING_TEST — the existing "skip findings with output files" mechanism handles the rest.

Report gate results to the user:

Pre-prove gates:
  Dedup: N duplicates in M groups (reordered for Phase 0 cascade)
  Coverage: K findings covered by existing tests
  Active: J findings to prove-fix

Build the dispatch queue

Read finding-list.txt to build the dispatch queue. If dispatch-order.txt exists (from dedup), use its ordering. The extraction script automatically filters out findings that already have prove-fix-*.md output files — so covered findings and resumed findings are skipped automatically. Do NOT read suspect files directly — the script already extracted what you need and reading suspect files would add thousands of tokens to your context.

Lens → test class mapping

Each domain lens has ONE shared adversarial test class. Determine the test directory from the project's build structure (e.g., src/test/java/<package>/ for Java). Test class naming:

Add rows for any other active lenses following the same pattern.

Processing order

Process findings grouped by lens, then by cluster, then by sequence number. Shorter lenses first for early signal:

1. Count findings per lens
2. Sort lenses by finding count (ascending)
3. Within each lens, process clusters in order (C1, C2, ...)
4. Within each cluster, process findings in sequence order

Dispatch loop

For each finding, launch ONE subagent:

Read .claude/prompts/audit/prove-fix.md for your full instructions, then execute them.

Your assignment:

• Finding: <finding ID> — "<finding title>"
• Construct: <construct name and location>
• Domain lens: <lens>
• Test class: <test class path from mapping above>
• Suspect file: .feature/<slug>/suspect-<lens>-cluster-<N>.md (your finding only — <finding ID>)
• Cluster packet: .feature/<slug>/cluster-<lens>-<N>-packet.md
• Write output to: .feature/<slug>/prove-fix-<short-id>.md

For the short-id in the output filename, use the finding ID with dots replaced by dashes (e.g., F-R5.cb.1.1 → prove-fix-F-R5-cb-1-1.md).

Wait for the subagent to return its summary line. Record the result.

Expected return format: "<finding ID>: <CONFIRMED_AND_FIXED | IMPOSSIBLE | FIX_IMPOSSIBLE> — <one-line summary>"

Error handling

If a subagent fails (API error, timeout, 0 tokens returned, no output file written):

1. Retry once — transient API errors (500, 529) usually resolve. Re-launch the same subagent with the same prompt.
2. If the retry also fails: mark the finding as DEFERRED with reason "subagent failure — <error description>". Do NOT stop the pipeline.
3. Continue to the next finding. Deferred findings can be retried in a subsequent run.
4. Report the failure: [N/total] <finding ID>: DEFERRED (API error)

After each subagent returns successfully:

1. Parse the result from the return summary
2. Accumulate totals (fixed, impossible, fix_impossible, deferred)
3. If AUDIT_BUDGET is set, run bash .claude/scripts/audit-budget.sh to get the running cost.
4. Report progress to the user (include cost if budget is set): [N/total] <finding ID>: <result> — <summary> ($<cost>/$<budget>)
5. Update Job 3 label: "Prove-Fix (<completed>/<total>, <fixed> fixed, <impossible> impossible) — $<cost>/$<budget>" (omit cost portion if no budget set)

After all findings are processed, mark Job 3: Prove-Fix complete. If zero confirmed findings, skip to Report.

Budget control

If AUDIT_BUDGET is set:

The budget is a soft cap. The goal is to stop prove-fix dispatch near the budget while guaranteeing post phases always run.

STOP_THRESHOLD = AUDIT_BUDGET - POST_RESERVE - 5

The -5 allows one more prove-fix finding (~$5) before stopping. This means overshoot is bounded to at most one prove-fix cycle plus post costs.

1. The cost check runs as part of the per-finding progress reporting (step 3 above: bash .claude/scripts/audit-budget.sh).
2. Compare the returned cost against STOP_THRESHOLD:
   • cost >= STOP_THRESHOLD: Dispatch the current finding (already in progress), then stop. Mark all remaining findings as DEFERRED. Report: Budget reached ($X of $Y). N findings deferred.
   • cost < STOP_THRESHOLD: Continue dispatching normally.
3. In the completion summary, include: Budget: $X spent of $Y limit (N findings deferred)
4. Post phases always run. Jobs 3b (test cleanup), 4 (report), and 5 (reconciliation) execute regardless of budget. The budget controls only prove-fix dispatch — it never skips post activities since those ensure a clean codebase.

If AUDIT_BUDGET is not set, omit budget checks entirely — the pipeline runs unbounded.

Job 3b: Test Cleanup (after prove-fix, before report)

After all prove-fix agents complete, pre-existing tests may assert old (buggy) behavior that the fixes changed. These are not regressions — they are tests written against the previous behavior that need updating to match the corrected behavior.

Launch a subagent:

You are the Test Cleanup subagent for an audit pipeline.

The prove-fix stage has fixed <N> bugs. Some pre-existing tests may now fail because they were asserting the old (buggy) behavior.

Run the full project test suite. For each failing pre-existing test (NOT the adversarial tests written by this audit):

1. Read the failing test method and its assertion
2. Read the prove-fix output that changed the behavior it tests
3. Check: do the failing test and the passing audit test both reference the SAME spec requirement (via covers: comments, finding IDs, or construct/method under test)?
4. Classify the failure into one of three categories:

STALE — test asserts OLD (buggy) behavior the fix corrected:

• Update the test to assert the NEW (correct) behavior
• Add a comment: "// Updated by audit <finding ID>: <old behavior> was a bug, now correctly <new behavior>"

SPEC AMBIGUITY — both the failing test and the audit test are valid interpretations of the same spec requirement, but they expect opposite behavior. The spec allows both readings:

• Do NOT modify either test
• Flag: "SPEC AMBIGUITY: <test method> and audit test <method> interpret <spec>.<requirement> differently. Failing test expects: <behavior A> Audit test expects: <behavior B> The spec requirement allows both — it needs tightening."
• Include the spec requirement text and both interpretations

REGRESSION — the fix genuinely broke unrelated behavior:

• Do NOT modify the test
• Flag: "REGRESSION: fix <finding ID> broke <test method> — the fix may need revision"

Return: "Test cleanup — <N> stale updated, <N> spec ambiguities, <N> regressions"

Expected return: stale test count + regression count.

If spec ambiguities found, ask the user using AskUserQuestion for each ambiguity. Include the spec requirement and conflicting test behaviors as context in the question text:

• Question: "<N> spec ambiguity detected — a requirement allows conflicting interpretations:\n\n <spec>.<requirement>: "<requirement text>"\n Test A (<method>) expects: <behavior A>\n Test B (<method>) expects: <behavior B>\n\nWhich interpretation should the code follow?"
• Options:
  1. "<behavior A>" — update Test B to match this interpretation
  2. "<behavior B>" — update Test A to match this interpretation
  3. "Defer" — leave both tests, resolve in /spec-author

Apply the user's decision: update the losing test and add a comment noting the decision. If deferred, note both tests as conflicting in the audit report.

If regressions found, ask the user using AskUserQuestion before proceeding to Report. Include the regression details as context in the question text:

• Question: "<N> pre-existing tests broke as regressions (not stale assertions):\n <test method> — broken by <finding ID>\n\nThese fixes may need revision."
• Options:
  1. "Continue to Report" — proceed to the report phase as-is
  2. "Review regressions" — examine the regressions before continuing

If zero failures in the full suite, skip this step entirely — no subagent needed.

Mark Job 3b: Test Cleanup complete.

Job 4: Report

4a. Aggregate results (mechanical script — no LLM)

Run the aggregation script to pre-compute summaries for the report subagent. This replaces ~80 individual file reads with 2 summary files.

python3 .claude/prompts/audit/aggregate-results.py .feature/<slug>/audit/<RUN_DIR>/

Expected output: "Aggregated N findings → prove-fix-summary.md"

If the script fails, fall back to the original behavior (the report subagent can still read individual files, just at higher cost).

4b. Launch report subagent

Launch a single subagent:

You are the Report subagent. Read .claude/prompts/audit/report.md for your complete instructions.

Read the pre-aggregated pipeline output files in .feature/<slug>/audit/<RUN_DIR>/:

• scope-definition.md (or classification.md)
• scope-exclusions.md
• active-lenses.md
• prove-fix-summary.md (pre-aggregated — do NOT read individual prove-fix-*.md files)
• boundary-summary.md (pre-aggregated — do NOT read individual suspect-*.md files)
• finding-list.txt

Perform cross-domain finding combination for constructs that appear in findings from multiple domain lenses.

Write audit-report.md and audit-prior.md.

Return a summary block for display.

Display the Report subagent's summary as the final output:

── Audit complete ─────────────────────────────
  Scope: <n> constructs, <n> clusters, <n> domain lenses active
  Findings: <n> suspected, <n> fixed, <n> impossible, <n> fix_impossible
  Cross-domain compositions: <n>
  Deferred (budget): <n>
  Spec conflicts: <n> (or "none")
  Fix-impossible needing resolution: <n> (or "none")
  Pipeline health: fix=<n>% impossible=<n>%
  Report: .feature/<slug>/audit-report.md
───────────────────────────────────────────────

If the report summary mentions spec conflicts, display the resolution flow:

── Spec conflicts detected ────────────────────
A fix from this audit contradicts an existing spec requirement.
This is a design tradeoff, not a bug — both the fix and the spec
had valid reasons. A decision is needed.

  CONFLICT-1: <description>
    Fix: <finding ID> — <what the fix changed>
    Spec: <spec>.<req> — <requirement text>
    Tradeoff: <why this is a tension>

Ask the user using AskUserQuestion for each conflict:

• Question: "CONFLICT-1: <description>\n Fix: <finding ID> — <what the fix changed>\n Spec: <spec>.<req> — <requirement text>\n Tradeoff: <why this is a tension>\n\nHow should this conflict be resolved?"
• Options:
  1. "Keep fix, update spec" — update the spec requirement to match the new behavior and check if other specs depend on the old behavior
  2. "Revert fix, keep spec" — revert the source change and mark the finding as FIX_IMPOSSIBLE with the spec requirement as the reason
  3. "Defer" — log the conflict as an open obligation on the spec

For option 1: read the conflicting spec, update the requirement to match the fix, and run spec-resolve.sh to check if any other spec's requires references the changed requirement. If so, flag the downstream specs.

For option 2: revert the fix (git checkout the changed lines), update the prove-fix output to FIX_IMPOSSIBLE, and note the spec requirement as the architectural constraint.

For option 3: add [UNRESOLVED] marker to the spec requirement and add an open_obligations entry. The spec becomes DRAFT if it was APPROVED.

If the report summary mentions Fix-impossible needing resolution: N with N > 0, walk each RELAX-<N> entry from the report's Fix Impossible — needs resolution section. Each entry represents a confirmed bug whose fix conflicts with an existing test that pins the old behavior; without explicit resolution the bug stays in the codebase invisibly.

── Fix-impossible findings need resolution ────
A fix was blocked by an existing test that pins the old behavior.
The bug is real (Phase 1 confirmed it). The decision is whether the
test, the spec, or the bug is authoritative.

  RELAX-1: <finding ID> — <one-line summary>
    Confirmed bug: <what the test proved>
    Blocking test: <test method> in <test class>
    Relaxation request: <what the fix would change + which test pins old behavior>
    Suggested route: <relax-test | wontfix | spec-author | defer>

Ask the user using AskUserQuestion for each RELAX-<N>:

• Question: "RELAX-<N>: <finding ID> — <summary>\n Confirmed bug: <bug description>\n Blocking test: <test method> in <test class>\n Relaxation request: <what the fix needs + which test pins>\n\n How should this be resolved?"
• Options:
  1. "Relax the blocking test" — the test pinned an implementation detail. Update it to match the new (correct) behavior, re-run prove-fix to confirm the fix now lands.
  2. "Accept as wontfix" — the old behavior is authoritative (backward compat, external contract). Record the finding in the report as a known-accepted-risk with rationale, and log a non-blocking wontfix: obligation on the most-relevant spec so /curate analysis 19 can resurface it if the rationale ever stales.
  3. "Escalate to /spec-author" — the pin test encodes an implicit spec requirement that was never written down. Author the spec, then the decision becomes a normal spec-conflict.
  4. "Defer to obligation" — log on the relevant spec as an open obligation; the spec becomes DRAFT if APPROVED.

For option 1: read the blocking test, update it to assert the new behavior, re-invoke the prove-fix subagent for this finding (now unblocked). Update the report with the new outcome (CONFIRMED_AND_FIXED). Append a pre-existing-test-modified entry to the report's Pre-existing Test Modifications section with the diff and proof of safety.

For option 2: append to the report's Removed Tests (Not Fixed) section classified DESIGN-CHANGE with the rationale (e.g. "external API contract pins this behavior; cannot change without versioned migration"). Update the prove-fix output's fix_detail to record the user's rationale. The finding stays FIX_IMPOSSIBLE in the audit report but is now explicitly accepted, not silent drag.

Then, if .spec/ exists and a relevant spec can be identified (same heuristic used for option 4 — domain matches the construct under test), append a non-blocking obligation to that spec's open_obligations array:

"wontfix: <finding ID> — <brief rationale>"

The wontfix: prefix lets humans reading /curate output distinguish accepted-risk entries from deferred work. Unlike option 4, the spec stays APPROVED — wontfix is a landed design decision, not a gap. The obligation is the resurface hook: /curate analysis 19 ages it by last-commit time and flags it once it exceeds --obligation-age-days (default 30). A human reviewing the curate report can then decide whether the rationale still holds or the wontfix is now fixable.

If .spec/ is absent or no relevant spec exists, skip the obligation write — the audit report entry is the only durable record. Note this in the prove-fix output so reviewers know the resurface hook is not wired.

For option 3: invoke /spec-author with a brief describing the pinned behavior the test encodes. After spec-author returns, re-evaluate the finding under the new spec — it usually becomes a spec-conflict (handled above) or wontfix.

For option 4: add an open_obligations entry on the most-relevant spec (the one whose domain matches the construct under test). The obligation text references the finding ID and the relaxation request so a future /curate analysis 19 (aging obligations) can resurface it. The spec becomes DRAFT if it was APPROVED.

If cross-cluster unresolved findings exist, ask the user using AskUserQuestion:

• Question: "<n> cross-cluster findings could not be analyzed in this round. Recommend another round with co-clustered constructs."
• Options:
  1. "Another round" — run an additional audit round with co-clustered constructs
  2. "Done" — finish the audit as-is

Job 5: Reconciliation (optional)

This job only runs when .spec/ exists. Check for the directory before launching.

# Check if .spec/ exists — mechanical check, not LLM judgment
test -d .spec/

If .spec/ does not exist, skip this job. Mark it as "Reconciliation (skipped — no specs)".

If .spec/ exists, launch a single subagent:

You are the Reconcile Findings subagent. Read .claude/prompts/audit/reconcile-findings.md for your complete instructions.

Feature slug: <slug> Run directory: <RUN_DIR>

Read all prove-fix outputs and the audit report from the run directory. Check .spec/ and .kb/ for existing entries.

Write spec-updates.md and kb-suggestions.md to .feature/<slug>/.

Return a summary line when done.

Expected return format: a structured RECONCILIATION_SUMMARY_START block containing spec update summaries, KB pattern summaries, and spec coverage (see reconcile-findings.md for the exact format).

Parse the summary block from the return. Do NOT read spec-updates.md or kb-suggestions.md — the subagent's return contains everything needed to present the feedback menus. The files exist for "review" mode (when the user wants full details) and for /curate to pick up later if skipped.

5a. Present and process spec updates

Extract the SPEC_UPDATES: section from the reconciliation return. If updates were suggested (count > 0), present inline:

Display the spec updates to the user:

── Feedback loop: spec updates ────────────────

Spec updates (<n> suggested):

  1. <requirement text — one line, observable behavior>
  2. <requirement text>
  ...

Then use AskUserQuestion to get the user's choice:

• "Apply all" — add all requirements to their specs now
• "Review" — show full details before deciding
• "Skip" — defer (resurfaces via /curate)

Do not present the KB menu or proceed until the user has responded.

If apply: for each suggested requirement, read the target spec file, append the requirement to the appropriate category section, and run spec-validate.sh to verify the spec is still valid. Then run spec-resolve.sh "<feature slug>" to check whether the new requirements conflict with or invalidate requirements in other specs. If conflicts are found, display them and ask the user to resolve before continuing.

Annotation follow-up (mandatory). Each newly-minted requirement was created because audit's prove-fix landed code that satisfies it. That code exists right now and must carry an @spec annotation to its new R-id — otherwise the requirement is born already-uncovered. For each applied requirement:

1. Identify the fix code prove-fix landed for the originating finding (the prove_fix.applied_files list in the per-finding artifact, or the diff against the audit's starting commit).
2. Open the smallest enforcing site in that fix and add a comment matching rules/spec-annotation-protocol.md:

   <comment-marker> @spec <spec-id>.R<n>

   Use the language's normal comment syntax. Place the annotation above the method, class, or guard clause whose behavior satisfies the new requirement.
3. Re-run spec-trace.sh <spec-id> and confirm the new R-id has at least one annotation reported.

If the fix code does not have an obvious enforcing site (rare — usually implies the requirement was added at the wrong granularity), surface this to the user and ask whether to (a) revise the requirement to match the actual code, (b) waive the annotation expectation with a documented reason, or (c) add the annotation to the corresponding test if one was written by prove-fix.

Report what was added and where each annotation landed. After all applies and annotations complete, rename the file: mv spec-updates.md spec-updates.applied.md

If review: read .feature/<slug>/spec-updates.md NOW (this is the one case where reading the file is justified — the user explicitly asked for full details). Display the full requirement blocks with notes, then re-offer apply/skip per item. Run the same validation after any applies. If all items were applied or skipped individually, rename the file.

If skip: note in the audit report that suggestions were deferred. They will resurface when /curate detects the deferred feedback file.

5b. Present and process KB patterns

Only present this menu after the user has responded to 5a (or if no spec updates exist). Extract the KB_PATTERNS: section from the reconciliation return. If patterns were found (count > 0), present inline:

Display the KB patterns to the user:

── Feedback loop: KB patterns ─────────────────

KB patterns (<n> updates + <n> new):

  1. <Update/New> <pattern name> — <one-line description>
  2. <Update/New> <pattern name> — <one-line description>
  ...

Then use AskUserQuestion to get the user's choice:

• "Create all" — create/update all KB entries now
• "Select" — choose which to create
• "Skip" — defer (patterns stay in kb-suggestions.md for later)

Do not proceed until the user has responded.

If create: for each pattern, invoke /research "<subject>" context: "audit adversarial pattern from <target>. Suggested: patterns/<concern>" as a sub-agent with the pattern's description and affected constructs as context. The <concern> is the pattern's lens (validation, concurrency, resource-management, testing, transactions), NOT the discovery domain. /research will infer type: adversarial-finding from the hint and use the adversarial-finding template at .kb/_refs/adversarial-finding-template.md.

After all creates complete, rename: mv kb-suggestions.md kb-suggestions.applied.md

If select: present each pattern individually for create/skip. After all items are processed, rename the file.

If skip: note as deferred. The file remains for /curate to pick up.

5c. Show spec coverage

After the user has responded to both menus (or if no updates/patterns were found), show spec coverage:

Spec coverage: <exercised>/<total> <spec-ID> requirements exercised
  (<N> unexercised — either well-implemented or not reached by this audit)
──────────────────────────────────────────────

Mark Job 5: Reconciliation complete.