pre-pr-audit

Files (3)

• SKILL.md
• output-template.md · references/
• scoring-model.md · references/

SKILL.md (first lines)

# 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

```mermaid
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 |

From sd0xdev/sd0x-harness, MIT.

More from sd0xdev/sd0x-harness