audit-review

SKILL.md

---
name: audit-review
description: Perform deep feature audits with transition-matrix and logical fault-injection validation. Use when reviewing complex changes, regressions, state-machine behavior, config interactions, API/protocol flows, and concurrency-sensitive logic.
---

# Audit Review

## Purpose

Run a repeatable deep audit for any feature and report confirmed defects with severity.
Default mode is static reasoning unless runtime execution is explicitly performed.

## Workflow

1. If PR scope is large, partition by functionality/workstream first:
   - define partitions and boundaries,
   - review each partition independently with the full workflow below,
   - track per-partition findings and coverage,
   - deduplicate cross-partition findings by root cause,
   - finish with cross-partition interaction risks.
2. Build call graph first:
   - user/system entrypoints (CLI, GitHub webhook handlers, Hetzner API calls)
   - dispatch and validation layers
   - state/storage/cache interactions (server state, runner state, metrics)
   - downstream integrations (GitHub API, Hetzner Cloud API, filesystem)
   - exception and error-propagation paths
3. Build transition matrix:
   - request/event entry -> processing stages -> state changes -> outputs/side effects
   - define key invariants and annotate where each transition must preserve them
4. Perform logical testing of all code paths:
   - enumerate all reachable branches in changed logic,
   - record expected branch outcomes (success, handled failure, fail-open/fail-closed, exception),
   - include happy path, malformed input, integration timeout/failure, and concurrency/timing branches.
5. Define logical fault categories from the code under review:
   - derive categories from actual components, transitions, and dependencies in scope,
   - document category boundary and affected states/transitions,
   - prioritize categories by risk and blast radius.
6. Run logical fault injection category-by-category:
   - execute one category at a time,
   - for each category cover success/failure/edge/concurrency paths as applicable,
   - record pass/fail-open/fail-closed/exception behavior per injected fault.
   - maintain a category completion matrix with status:
     - Executed / Not Applicable / Deferred,
     - outcome,
     - defects found,
     - justification for Not Applicable or Deferred.
7. Confirm each finding with code-path evidence.
8. Produce coverage accounting:
   - reviewed vs unreviewed call-graph nodes,
   - reviewed vs unreviewed transitions,
   - executed vs skipped fault categories (with reasons).
   - mark coverage complete only when every in-scope node/transition/category is reviewed or explicitly skipped with justification.
9. For async/concurrent paths, perform interleaving analysis:
   - write several plausible execution interleavings per critical transition,
   - identify race conditions and state corruption hazards.
10. For mutation-heavy paths, perform rollback/partial-update analysis:
   - reason about exception/cancellation at intermediate points,
   - verify state invariants still hold.

## Python Bug-Type Coverage (Required for Python audits)

- exception handling gaps (bare except, swallowed exceptions, missing cleanup)
- resource leaks (unclosed files, connections, subprocess handles)
- type errors and None handling (AttributeError on None, type mismatches)
- mutable default arguments and shared state mutations
- async/threading race conditions and GIL-related issues
- API contract violations (GitHub API, Hetzner Cloud API response handling)
- configuration parsing errors and missing validation
- credential/secret exposure in logs or error messages
- YAML/JSON parsing edge cases and injection risks

## Cloud Infrastructure Emphasis

For this autoscaling GitHub runners system, prioritize these checks before lower-risk issues:

1. Server lifecycle management (creation, deletion, orphaned resources, cost leaks).
2. GitHub API rate limiting and error handling (token exhaustion, API failures).
3. Hetzner Cloud API error handling (quota limits, network failures, stale state).
4. Runner registration/deregistration race conditions.
5. Scale-up/scale-down logic correctness under concurrent requests.

## Output Contract

- Start with confirmed defects only.
- Group by severity: High, Medium, Low.
- For each defect include:
  - title,
  - impact,
  - file/function anchor,
  - fault-injection trigger,
  - transition mapping,
  - why it is a defect (not a design preference),
  - smallest logical repro steps,
  - likely fix direction (short, concrete: 2-4 bullets or sentences),
  - regression test direction (short, concrete: 2-4 bullets or sentences),
  - affected subsystem and blast radius,
  - at least one code snippet proving the defect.
- Separate “not confirmed” or “needs runtime proof” from confirmed defects.
- Include an **Assumptions & Limits** section for static reasoning.
- Include an overall **confidence rating** and what additional evidence would raise confidence.
- If no defects are found, include residual risks and untested paths.
- For large PRs, include per-partition findings/coverage and final cross-partition risk summary.
- Include a fault-category completion matrix for every deep audit.

### Canonical report order

1. Scope and partitions (if large PR)
2. Call graph
3. Transition matrix
4. Logical code-path testing summary
5. Fault categories and category-by-category injection results
6. Confirmed defects (High/Medium/Low)
7. Coverage accounting + stop-condition status
8. Assumptions & Limits
9. Confidence rating and confidence-raising evidence
10. Residual risks and untested paths

## Standard Audit Report Template (Default: Pointed PR Style)

Default report style should match concise PR review comments:
- fail-first and action-oriented,
- only confirmed defects (no pass-by-pass narrative),
- one short summary line when there are no confirmed defects.

Use the compact template below by default. Use the full 10-section canonical format only when explicitly requested.

```markdown
Audit update for PR #<id> (<short title/scope>):

Confirmed defects:

- **<Severity>: <short defect title>**
  - Impact: <concrete user/system impact>
  - Anchor: `<file>` / `<function or code path>`
  - Trigger: <smallest condition that triggers defect>
  - Why defect: <1-2 lines, behavior not preference>
  - Fix direction (short): <2-4 bullets or sentences>
  - Regression test direction (short): <2-4 bullets or sentences including positive and edge/failure cases>
  - Evidence:
    ```start:end:path
    // minimal proving snippet
    ```

<repeat per defect, sorted High -> Medium -> Low>

Coverage summary:
- Scope reviewed: <partitions or key areas, one line>
- Categories failed: <count/list>
- Categories passed: <count only>
- Assumptions/limits: <one line>
```

## Severity Rubric

- High: realistic trigger can cause crash/UB/data corruption/auth bypass/deadlock.
- Medium: correctness/reliability issue with narrower trigger conditions.
- Low: diagnostics/consistency issues without direct correctness break.

## Checklist

- Verify call graph is explicitly documented before defect analysis.
- Verify invariants are explicitly listed and checked against transitions.
- Verify fail-open vs fail-closed behavior where security-sensitive.
- Verify logical branch coverage for all changed code paths.
- Verify fault categories are explicitly defined from the reviewed code before injection starts.
- Verify category-by-category execution and reporting completeness.
- Verify full fault-category completion matrix is present and complete.
- Verify concurrency and cache/state transition paths.
- Verify multithreaded interleavings are explicitly analyzed for critical shared-state paths.
- Verify rollback/partial-update safety under exception/cancellation points.
- Verify major Python bug classes are explicitly covered (or marked not applicable).
- Verify race/deadlock/crash class defects are prioritized and explicitly reported.
- Verify error-contract consistency across equivalent fault paths.
- Verify performance/resource failure classes were considered.
- Verify findings are deduplicated by root cause.
- Verify coverage accounting is present (covered vs skipped with reason).
- Verify stop-condition criteria for coverage completion are explicitly satisfied.
- Verify every confirmed defect includes code evidence snippets.
- Verify parser/config/runtime consistency.
- Verify protocol/API parity across entrypoints.
- Verify no sensitive-data leakage in logs/errors.