/crap

Score code-risk hotspots with a CRAP-style formula across Rust, Python, and TypeScript.

This skill is for analysis and planning, not auto-fixing. It finds risky functions, ranks them, and turns the output into candidate /describe remediation units.

Modes

• Analysis mode: baseline score, hotspot ranking, next-action suggestions.
• One-shot remediation mode: bootstrap coverage, split work, implement test- first slices, rerun CRAP, commit progress, and iterate until the scoped score is below threshold or a blocker stops the loop.

Enter one-shot remediation mode only when the user explicitly asks for execution, for example:

• do it
• take it under 30
• /crap 25
• /crap --threshold 25
• fix the hotspots
• keep going

Supported v1 Languages

• rust
• python
• typescript

If the user explicitly requests a language outside that set, stop and say so. Do not fabricate scores, findings, or follow-on suggestions for unsupported languages.

Runtime Contract

Resolve and run scripts/analyze_crap.py relative to this SKILL.md. Use references/coverage-targets.md when the repo needs coverage artifact targets. Use references/testing-bootstrap.md when the repo lacks a trustworthy test or coverage baseline. Use references/one-shot-loop.md for autonomous reduction loops. Use scripts/delta_audit.py to validate that score improvements are legitimate during remediation loops.

Examples:

python3 scripts/inspect_test_stack.py
python3 scripts/inspect_test_stack.py /path/to/repo
python3 scripts/inspect_test_stack.py /path/to/repo --json
python3 scripts/analyze_crap.py
python3 scripts/analyze_crap.py /path/to/repo
python3 scripts/analyze_crap.py /path/to/repo --languages rust,python
python3 scripts/analyze_crap.py /path/to/repo --languages python --top 20
python3 scripts/analyze_crap.py /path/to/repo --languages python --threshold 25 --top 20
python3 scripts/delta_audit.py snapshot /path/to/repo -o /tmp/crap-baseline.json
python3 scripts/delta_audit.py audit /tmp/crap-baseline.json /path/to/repo
python3 scripts/delta_audit.py audit /tmp/crap-baseline.json /path/to/repo --json

Threshold Resolution

Treat the target threshold as explicit input when the user provides one.

Resolve it in this order:

1. --threshold N in the skill request or analyzer invocation
2. A bare numeric argument after /crap, for example /crap 25
3. Natural-language phrasing such as take it under 25 or get this below 20
4. Default target: FINAL_SCORE < 30

Guardrails:

• Thresholds must be positive numbers.
• Keep the threshold scoped. A package-scoped run against 25 is not a repo-wide 25 claim.
• In one-shot remediation mode, restate the resolved threshold before starting the loop.

Mutation-Testing Hand-Off

crap and mutate solve different problems.

• crap ranks change risk from complexity plus coverage.
• mutate checks whether the tests around a scoped hotspot are strong enough to kill realistic defects.
• Mutation testing does not directly change CRAP. Only coverage and complexity changes move the CRAP score.

Default hand-off rule:

1. Prioritize coverage bootstrap, characterization tests, and simplification while the scoped FINAL_SCORE is still >= 30.
2. Start mentioning or running the sibling mutate skill when the scoped FINAL_SCORE is below 30, the baseline test path is green, coverage is numeric, the hotspot language is still within CRAP's supported set (rust, python, typescript), and the hotspot scope is narrow enough to mutate economically.
3. Treat FINAL_SCORE < 8 as a strong end-state target for stabilized hotspots, not as the prerequisite for mentioning mutate.
4. If the user explicitly wants mutation testing while the scoped score is still >= 30, keep the scope tight and say that the mutation signal may be noisy until the hotspot stabilizes.

Analyzer Behavior

The bundled script must be the source of truth for score calculation and report formatting.

1. Default target is the current working directory.
2. Default language mode is autodetect across supported files.
3. Coverage sources may come from lcov.info, coverage.xml, or cobertura.xml.
4. Missing coverage stays N/A. Never infer or invent coverage.
5. Sort numeric CRAP rows first, descending by score. Sort N/A rows after numeric rows.
6. The analysis block must end with exactly one machine-readable line: FINAL_SCORE: <value>
7. FINAL_SCORE is the maximum numeric CRAP value in the run. If there are no numeric CRAP values, emit FINAL_SCORE: 0.00.

Prerequisite Bootstrap

If the scope does not have a trustworthy baseline test path yet, fix that before doing hotspot remediation.

1. Run python3 scripts/inspect_test_stack.py {target} before improvising shell discovery when any of these are true:
   • the analyzer is all N/A
   • the repo has tests but no machine-readable coverage artifact
   • the repo does not appear to have tests at all
2. Interpret the inspector result strictly:
   • ready: reuse the existing baseline and coverage artifact
   • add-coverage-target: keep the fast path intact and add an additive coverage target
   • bootstrap-tests: add the smallest repo-native harness that makes CRAP measurement trustworthy
3. Prefer narrow characterization tests around the hottest scope instead of broad suite architecture when bootstrapping.
4. In mixed or monorepo scopes, narrow to the package or crate that owns the hotspot if root automation is too thin to provide a trustworthy baseline.
5. Do not claim a repo-wide numeric result after narrowing to one package.
6. Do not use divide-and-conquer for hotspot slices until the baseline test path and coverage lane are real. Treat bootstrap work as an upstream node.

Coverage Bootstrap

When the analyzer returns FINAL_SCORE: 0.00 because every finding is coverage N/A / CRAP N/A, treat that as a coverage-artifact gap first.

1. Inspect repo automation before answering:
   • start with python3 scripts/inspect_test_stack.py {target}
   • then inspect Makefile, package.json, pyproject.toml, or Cargo.toml only as needed to implement the chosen bootstrap lane
2. If a machine-readable coverage target already exists, point to that exact target or script.
3. If coverage exists only as a terminal report, prefer adding an additive target instead of mutating the canonical fast path.
4. If the repo has a Makefile and the user wants a real numeric CRAP score, prefer creating Makefile targets instead of handing back a long one-off command.
5. Reuse the example names in references/coverage-targets.md:
   • pytest-cov-xml
   • vitest-cov-lcov
   • cargo-cov-lcov
   • optional aggregate crap
6. After adding or identifying the right target, run it if feasible, then rerun the analyzer.
7. If you only bootstrap coverage for a subset of the original target, label the rerun explicitly as package-scoped or path-scoped and name that path.
8. If the inspector reports bootstrap-tests, establish the smallest viable test lane first, then add the coverage target. Do not skip straight to coverage flags on an otherwise nonexistent test harness.

Guardrails:

• Keep fast-path developer targets like make test or make pytest intact.
• Prefer sibling targets with explicit artifact names over silently changing existing targets.
• If the repo has no Makefile, mirror the same behavior in package-manager scripts, but still prefer Makefile in repos that already use it.
• For mixed-language repos, create one target per tested ecosystem, then an optional aggregate make crap wrapper.
• If the user invoked /crap from repo root, do not silently redefine the result as repo-wide after narrowing to one package or one language. Either produce a true repo-wide rerun or state that the numeric result is partial.
• If the repo has no tests, do not stop at "coverage missing." Bootstrap the smallest repo-native harness needed for a trustworthy score.

Scope Discipline

Treat scope as part of correctness.

1. If the analyzer runs against the user’s original target unchanged, call the result repo-wide or target-wide.
2. If you rerun against a narrower path such as packages/python-server-quickstart, call the result package-scoped or path-scoped and name the exact path.
3. If the user started at repo root and only one ecosystem has usable coverage, say that the repo-wide numeric score is still unknown.
4. Prefer adding per-language coverage targets plus an aggregate make crap wrapper when the user wants a true repo-wide score in a mixed-language repo.

Delta Integrity Audit

In one-shot remediation mode, every iteration must pass a delta integrity check before committing. This prevents agents from gaming the score through structural tricks that lower the metric without improving actual code quality.

When to run

Run the audit after every re-measure step in the one-shot loop (between step 4 and step 5 in references/one-shot-loop.md).

How to run

1. At the start of the remediation loop, capture a baseline snapshot:

python3 scripts/delta_audit.py snapshot {target} --languages {languages} -o /tmp/crap-baseline.json

2. After each iteration's re-measure, audit the delta:

python3 scripts/delta_audit.py audit /tmp/crap-baseline.json {target} --languages {languages}

3. Read the DELTA_INTEGRITY line:

   • clean: proceed to commit.
   • warning: review the flags, proceed if justified.
   • suspicious: stop the loop. Show the flags to the user. Do not commit until the user acknowledges or the suspicious changes are reverted.

4. After a clean commit, take a fresh snapshot for the next iteration.

What it detects

Category	Signal	Verdict
split-without-reduction	Function disappeared, 2+ replacements in same file, sum(CC) >= original	suspicious
scope-escape	File disappeared without git deletion or rename into scope	suspicious
hollow-coverage	New test files with zero assertions	suspicious
hollow-coverage	Existing test file assertions dropped to zero	warning
scope-narrowing	Target path changed between snapshot and audit	suspicious

Guardrails

• Never commit with DELTA_INTEGRITY: suspicious unless the user explicitly acknowledges the flags.
• When a split is flagged, explain to the user whether the split genuinely reduced coupling or just distributed the same complexity.
• When scope escape is flagged, verify whether code was intentionally deleted or just moved out of the analyzed target.
• A clean delta audit does not guarantee the work is good — it guarantees the score movement is honest. Test quality still needs separate validation (mutation testing, code review).

Output Flow

Phase 1: Run the analyzer

Run the script and show its report.

Phase 1.5: Bootstrap coverage targets when scores are all N/A

If the report is all N/A and the user asks how to get a real score, inspect the current scope with scripts/inspect_test_stack.py, then add the missing test and coverage prerequisites instead of only suggesting an ad hoc raw command. Use references/testing-bootstrap.md for harness bootstrap and use references/coverage-targets.md for examples.

Phase 1.6: Preserve scope in the close-out

If the rerun scope changed from the original target:

1. Say so before presenting the numeric score.
2. Name the narrowed path explicitly.
3. Do not call the resulting FINAL_SCORE repo-wide.

Phase 2: Suggest next work only when numeric hotspots exist

If the analyzer produced one or more numeric CRAP findings, emit the suggested follow-on block and ask exactly:

Do we need to adjust any of these findings/next actions, or does this look good to launch?

If there are no numeric findings, stop after the analysis output. Do not add divide-and-conquer recommendations.

Phase 3: Apply ask-cascade on the response

If the user answers adjust or otherwise asks for changes:

1. Ask exactly one strategic question first about what to adjust.
2. Do not ask detailed follow-ups until that answer is known.
3. Re-evaluate the follow-up questions after the user answers.

Good first question:

What should we adjust first: scoring accuracy, hotspot grouping, follow-on /describe items, or the launch recommendation?

This gate is mandatory. Do not blast the user with all possible follow-up questions at once.

If the user answers launch or approves the next actions:

1. Keep the decomposition read-only until the next explicit execution request.
2. Recommend divide-and-conquer as the follow-on path when there are multiple hotspot groups.
3. Turn each hotspot group into its own /describe candidate plus one cross-language synthesis item.

Phase 4: One-shot remediation mode

If the user explicitly asks for execution, use references/one-shot-loop.md and run the bounded loop directly.

Requirements:

1. Set a threshold first using the resolution order above. Default to FINAL_SCORE < 30.
2. Establish a trustworthy baseline before choosing slices.
   • run scripts/inspect_test_stack.py when the baseline is unclear
   • bootstrap tests before coverage when the scope lacks both
3. Use divide-and-conquer when hotspot groups are independent and the runtime supports it after measurement prerequisites are ready. Otherwise use the same concern split in a single-agent loop.
4. Implement tests first for each slice.
5. After each slice, rerun the canonical test path, the coverage target, and the analyzer.
6. When a slice drops below 30, the baseline is green, and the hotspot language is one of CRAP's supported v1 languages (rust, python, typescript), optionally run the sibling mutate skill on that narrowed scope before the next CRAP rerun.
7. Use scripts/analyze_crap.py ... --top 20 for inner-loop reruns to keep output concise while preserving the true FINAL_SCORE.
8. After each re-measure, run scripts/delta_audit.py audit against the baseline snapshot. If DELTA_INTEGRITY is suspicious, stop the loop and show the flags to the user before committing.
9. Use the commit skill after each stable logical batch. Do not wait until the very end if the loop is making clean progress. After committing, take a fresh delta audit snapshot for the next iteration.
10. Continue iterating until:
    • the scoped score is below threshold, or
    • a real blocker or diminishing-return stop condition from references/one-shot-loop.md is hit.
11. Report the score delta after each loop step:

• baseline score
• current score
• top moved hotspots
• commits made
• blocker or next slice

Reporting Rules

• Show file path, symbol, CC, coverage, and CRAP for each finding.
• Use N/A for coverage and CRAP when no matching coverage data exists.
• Keep the final score line machine-readable and unique.
• Do not hide lower-scoring supported languages just because one language is worse.
• Do not include unsupported files in the ranking.
• When coverage artifacts are missing, explain the gap and prefer creating a repo-native target over returning only a one-off command.
• Treat the analyzer output as the formatting source of truth. Do not replace its Suggested /describe follow-on block or its final launch gate question with a handwritten alternative.
• If the scoped FINAL_SCORE is below 30 and the reported language is still within CRAP's supported v1 set (rust, python, typescript), mention the sibling mutate skill as a follow-on path. Make it explicit that mutation results are a test-quality signal, not a direct CRAP input.
• If you add a human summary, place it after the raw analyzer output and keep scope labels exact.
• If a threshold was explicitly provided or the user is in one-shot mode, state the resolved target alongside the score comparison.
• In one-shot remediation mode, keep inner-loop analyzer reruns concise with --top, but always report the true current FINAL_SCORE.