Pre-PR Audit

Pre-deploy ultimate verification. Aggregates test quality, coverage, risk, and AC traceability into a single confidence index (0-100).

When NOT to Use

Need	Use Instead
Code review	/codex-review-fast
Post-deploy runtime verification	/feature-verify
Periodic project health	/project-audit
Just run tests	/verify
Just check risk	/risk-assess

Positioning

auto-loop done → [user invokes] /pre-pr-audit → /smart-commit → /push-ci → /create-pr
                                     ↑
                        Pre-deploy ultimate verification

Prohibited

❌ git add | git commit | git push — per @rules/git-workflow.md

This skill audits and reports. It does not modify code or commit.

Arguments

Arg	Description	Default
--mode fast|deep	Execution depth	fast
--strict	Non-zero exit on ⛔ PR-Blocked	off
--json	Machine-readable JSON output	off
--base <ref>	Compare against ref	HEAD

Workflow

sequenceDiagram
    participant U as User
    participant PA as /pre-pr-audit
    participant RA as /risk-assess
    participant TR as /codex-test-review
    participant CC as /check-coverage
    participant ST as State / Artifacts

    U->>PA: /pre-pr-audit [--mode fast|deep]

    par Phase A: Collect
        PA->>ST: [Background Agent] Read precommit/verify state (freshness)
        PA->>RA: [Foreground — blocking] /risk-assess --mode fast
        PA->>PA: [Background Agent] Detect changed files + map to tests
    end

    Note over PA,RA: A2 (/risk-assess) must complete before Phase B — Module Selection depends on top_affected

    par Phase B: Targeted
        PA->>TR: [Foreground] /codex-test-review (top-N risky modules)
        PA->>CC: [Background Agent: coverage-analyst] /check-coverage (deep mode only)
        PA->>PA: [Background Agent] AC trace (if request doc exists)
    end

    PA->>PA: Phase C: Aggregate + hard-fail + gate
    PA->>U: Phase D: Report + confidence index

Phase A: Collect (parallel with dispatch annotations)

Step	Dispatch	Action	Source
A1	Background Agent	Read precommit gate state	node scripts/review-state.js check --format=json（installed copy first: .claude/scripts/review-state.js）— precommit.passed is content-addressed, so no separate freshness check; checker absent → report "no verdict source", never a pass
A2	Foreground (blocking)	Invoke /risk-assess --mode fast	Skill tool — get risk_level, top_affected, overall_score
A3	Background Agent	Detect changed files + map to test files	git diff --name-only HEAD → match against test/ patterns

A2 MUST stay foreground: Phase B Module Selection depends on top_affected from /risk-assess output. A1 and A3 dispatch as background Agents and run in parallel with A2.

Phase B: Targeted (parallel, mode-dependent, with dispatch annotations)

Step	Dispatch	fast	deep	Action
B1	Foreground	top-3	top-10 + supplement	/codex-test-review on selected modules (see Module Selection)
B2	Background Agent (coverage-analyst)	skip	✅	/check-coverage for pyramid balance
B3	Background Agent	skip	✅ (if request doc)	AC trace: parse ## Acceptance Criteria → map to test evidence

B2 uses coverage-analyst agent (agents/coverage-analyst.md) for background execution. B3 dispatches as a background Agent for AC parsing. B1 stays foreground as the primary assessment.

Module Selection (deterministic)

1. Source: /risk-assess JSON top_affected array (each: file + dependent_count)
2. Sort: dependent_count descending, tie-break file ascending (stable sort)
3. Select: N=3 (fast) or N=10 (deep)
4. Deep supplement: git diff --name-only for files beyond top_affected cap
5. Fallback: if /risk-assess unavailable, git diff --stat sorted by lines-changed descending

Phase C: Aggregate

1. Score each dimension per references/scoring-model.md
2. Apply evidence confidence cap
3. Check hard-fail overrides
4. Compute final index

Phase D: Output

Output report per references/output-template.md. Include:

• Confidence index (0-100) + gate sentinel
• Per-dimension score table
• Hard-fail check results
• Findings (prioritized)
• Next actions

5 Dimensions

#	Dimension	Weight	Checks
1	Execution Integrity	25%	All tests pass, precommit passed, HEAD freshness, no flaky indicators
2	Coverage Adequacy	25%	Unit/Integration/E2E pyramid balance, per-module gap analysis
3	Test Quality	20%	AAA compliance, naming, assertion depth, edge cases (per @rules/testing.md)
4	Risk-to-Test Alignment	20%	High-risk files have tests, risk level ↔ test depth proportional
5	Evidence Governance	10%	AC traceability, exception policy compliance, expiry check

Gate

User-Facing (3-tier)

Gate	Score	Sentinel
✅ Ready	>=85	✅ PR-Ready
⚠️ Needs attention	60-84	⚠️ PR-Caution
⛔ Not ready	<60	⛔ PR-Blocked

Hard-Fail Overrides (force ⛔ regardless of score)

Override	Condition	Source
Precommit stale	Not passed after latest edit	Auto-loop state
Prohibited domain exception	Security/data-integrity/regression AC uses manual exception	@rules/testing.md
Exception cap exceeded	Exception count > AC-count-based cap	@rules/testing.md
Expired exception	Past ISO 8601 expiry date	@rules/testing.md
Invalid reason class	Non-enum reason in exception	@rules/testing.md
Unverified exception	No Codex VALID_EXCEPTION verdict	@rules/testing.md
Evidence stale	Artifacts HEAD SHA ≠ current HEAD	Freshness check
Critical untested	/risk-assess HIGH+ on zero-coverage files	Risk alignment

Scoring

See references/scoring-model.md for formulas. Summary:

• Check score: pass=1, partial=0.5, fail=0, N/A=excluded
• Dimension score: applicable_sum / applicable_count × 100
• Raw: weighted average (N/A dimensions excluded, weights renormalized)
• Cap: evidence confidence (1.0 / 0.9 / 0.75 / 0.6)
• Final: round(raw × cap)

Verification

• All 5 dimensions scored (or N/A with reason)
• Hard-fail checks evaluated
• Gate sentinel output matches score
• No git add/commit/push executed
• Findings include file:line references

References

• references/scoring-model.md — Scoring formulas, weights, confidence caps
• references/output-template.md — Report format, JSON schema, sentinel strings
• @rules/testing.md — Evidence model + exception policy (hard-fail source)
• @rules/testing-project.md — Project-specific overrides

Examples

Input: /pre-pr-audit
Action: fast mode → collect state + risk → test-review top-3 → aggregate → 85/100 ✅ PR-Ready

Input: /pre-pr-audit --mode deep
Action: deep mode → collect + risk → test-review top-10 + coverage + AC trace → aggregate → 72/100 ⚠️ PR-Caution

Input: /pre-pr-audit --strict
Action: fast mode → score 55 → ⛔ PR-Blocked → exit code 1