Help us improve
Share bugs, ideas, or general feedback.
From subagent-system
Audits pipeline output in 4 levels: structural checks, content sampling, severity gating, semantic review, and regression tests. Requires a test command as argument or Level 4 is skipped.
npx claudepluginhub minhthang1009/dotclaude --plugin subagent-systemHow this skill is triggered — by the user, by Claude, or both
Slash command
/subagent-system:audit-outputThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
**Required argument:** `$ARGUMENTS` = the project's test command (e.g., `npm test`, `pytest -q`, `cargo test`). If not provided, Level 4 regression tests will be SKIPPED and the audit report will be marked incomplete.
Creates p5.js generative art with seeded randomness, noise fields, and interactive parameter exploration. Use for algorithmic art, flow fields, or particle systems.
Share bugs, ideas, or general feedback.
Required argument: $ARGUMENTS = the project's test command (e.g., npm test, pytest -q, cargo test). If not provided, Level 4 regression tests will be SKIPPED and the audit report will be marked incomplete.
Run the following checks in order. Each level must complete before the next begins.
Level 1 — Structural checks (automated):
0. Resolve PROJECT_ROOT: Bash("git rev-parse --show-toplevel 2>/dev/null || echo '.'") — use this for all path references below.
[start-commit] from [PROJECT_ROOT]/.claude/checkpoints/chain-start-commit:
Bash("cat [PROJECT_ROOT]/.claude/checkpoints/chain-start-commit 2>/dev/null || echo MISSING")
If MISSING: skip git diff and note "chain-start-commit not found — scope violation check skipped."Bash("git -C [PROJECT_ROOT] diff --name-only [start-commit]"): compare files changed against files assigned to each agentLevel 2 — Content sampling:
5. Use the coverage-verifier skill: spot-check 2–3 randomly selected files per agent for actual coverage
6. Use the fact-checker skill: verify the top 5 findings by severity against source code
7. Run git diff HEAD -- [file] for every file claimed as edited: count real edits vs claimed edits
Level 2.5 — Severity gate: First verify FINDINGS_REPORT.md exists:
Bash("test -f \"[PROJECT_ROOT]/.claude/FINDINGS_REPORT.md\" && echo EXISTS || echo MISSING")
If MISSING: set PIPELINE_STATUS: NEEDS_RERUN and note "FINDINGS_REPORT.md not found — consolidate-findings must run before audit-output." Skip the severity-gate and proceed directly to Level 3.
If EXISTS: run the severity-gate skill — it reads .claude/FINDINGS_REPORT.md directly (the consolidated report from consolidate-findings, NOT the fact-checker output). Do NOT pass fact-checker output as injected text to severity-gate. If any CRITICAL findings are verified:
PIPELINE_STATUS: NEEDS_RERUN immediatelyACTION_REQUIREDLevel 3 — Semantic review: 8. Assess goal alignment: does the pipeline output address the original objectives? 9. Note any findings from the pipeline-reviewer if it ran as part of the pipeline
Level 4 — Regression: 10. Run the project's test command if available 11. Run the project's lint and type-check commands if available
Write the results to [PROJECT_ROOT]/.claude/AUDIT_REPORT.md (using the PROJECT_ROOT resolved in Step 0):
# AUDIT_REPORT — [timestamp]
## Level 1 — Structural
- Tasks: [PROCESSED]/[TOTAL]
- Files modified: [N] (expected [M])
- Scope violations: [list or NONE]
## Level 2 — Content
- Coverage: [THOROUGH/SUPERFICIAL/FABRICATED] for [N] agents sampled
- Fact check: VERIFIED [N] / CONTRADICTED [N] / UNVERIFIED [N]
- Git diff verify: [N] real edits / [N] claimed edits
## Level 3 — Semantic
- Goal alignment: YES | PARTIAL | NO
## Level 4 — Regression
- Tests: PASS | FAIL | SKIPPED
- Lint: CLEAN | ERRORS | SKIPPED
## Verdict
PIPELINE_STATUS: CLEAN | ISSUES_FOUND | NEEDS_RERUN
ACTION_REQUIRED:
- [specific action if ISSUES_FOUND, or NONE]