scanning-quality

name	scanning-quality
description	Scans the codebase for bugs, logic errors, cache races, workflow problems, insecure defaults, security regressions in the diff, and variant analysis on prior findings. Runs a Workflow that fans out one finder per scan type in parallel, runs variant-analysis as a dependent stage, adversarially verifies High/Critical findings, deduplicates, and produces an A-F prioritized report. Use when preparing a release, investigating quality issues, running pre-merge checks, or whenever a recent diff touches security-sensitive code.
user-invocable	true
allowed-tools	Workflow, Task, Read, Grep, Glob, Write, AskUserQuestion, Bash(pnpm run check:*), Bash(pnpm run test:*), Bash(pnpm test:*), Bash(git status:*), Bash(git diff:*), Bash(git log:*), Bash(rg:*), Bash(grep:*), Bash(find:*), Bash(ls:*)
model	claude-opus-4-8
context	fork

scanning-quality

Quality analysis across the codebase via a Workflow. Cleans up junk files, runs structural validation, then fans out one finder agent per scan type in parallel (variant-analysis as a dependent stage, adversarial verify on High/Critical), deduplicates, and produces an A-F prioritized report.

Modes

• Default (interactive): AskUserQuestion is used to confirm cleanup deletions and to pick scan scope.
• Non-interactive: /scanning-quality non-interactive (or any of the aliases below) skips every AskUserQuestion and applies safe defaults: scan scope = all types, cleanup = leave junk files in place (don't delete without confirmation), report-save = yes (reports/scanning-quality-YYYY-MM-DD.md). Use this when running headlessly (CI cron, programmatic Claude, any non-TTY driver). The four-flag programmatic-Claude lockdown rule already strips AskUserQuestion, so headless runs default to non-interactive automatically. Call it out explicitly so future readers understand the contract.

Detect non-interactive mode via any of: --non-interactive argument, non-interactive argument, SCANNING_QUALITY_NONINTERACTIVE=1 env var, or absence of AskUserQuestion in the available tool surface.

Scan Types

Legacy scan types (agent prompts in reference.md):

1. critical - Crashes, security vulnerabilities, resource leaks, data corruption
2. logic - Algorithm errors, edge cases, type guards, off-by-one errors
3. cache - Cache staleness, race conditions, invalidation bugs
4. workflow - Build scripts, CI issues, cross-platform compatibility
5. workflow-optimization - CI optimization (build-required conditions on cached builds)
6. security - GitHub Actions workflow security (zizmor scanner)
7. documentation - README accuracy, outdated docs, missing documentation
8. patch-format - Patch file format validation

Modular scan types (one file per type under scans/, easier to extend than the monolithic reference.md):

9. variant-analysis: for each High/Critical finding from above, search the rest of the repo for the same shape. See scans/variant-analysis.md.
10. insecure-defaults: fail-open defaults, hardcoded credentials, lazy fallbacks. See scans/insecure-defaults.md.
11. differential: security-focused diff against a base ref. See scans/differential.md.
12. bundle-trim: for repos that ship a built bundle (today: rolldown), identify unused module paths the bundler statically pulled in but the runtime never reaches. Reports candidates; the trim loop itself lives in the trimming-bundle skill. See scans/bundle-trim.md.
13. deadcode-removal: surface dead source files, test-only helpers, stale // eslint-disable / // oxlint-disable directives, and dead string-literal constants. Captures the fleet rule that socket/export-top-level-functions REQUIRES export on helpers (exports exist for tests), so the scan never recommends dropping export to colocate. See scans/deadcode-removal.md.

Adding a new scan type: drop a file under scans/<name>.md describing mission, method, output shape, when-to-skip; same shape as the three above. The orchestrator picks them up by directory listing; no edits to this SKILL.md needed beyond appending to the list.

The split exists because adding a 12th, 15th, 20th scan type into reference.md produces exactly the "this and also that and also the other thing" file CLAUDE.md's File-size rule warns about. Per-type files keep each scan reviewable in isolation.

Process

Phase 1: Validate Environment

git status

Warn about uncommitted changes but continue (scanning is read-only).

Phase 2: Update Dependencies

pnpm run update

Only update the current repository. Continue even if update fails.

Phase 3: Install zizmor

Install zizmor for GitHub Actions security scanning, respecting the soak time (pnpm-workspace.yaml minimumReleaseAge in minutes, default 10080 = 7 days). Query GitHub releases, find the latest stable release older than the threshold, and install via pipx/uvx. Skip the security scan if no release meets the soak requirement.

Phase 4: Repository Cleanup

Find junk files (interactive mode confirms each batch via AskUserQuestion; non-interactive mode lists what was found in the report and leaves them in place; don't delete files without explicit confirmation, even on a clean dirty-tree):

• SCREAMING_TEXT.md files outside .claude/ and docs/
• Test files in wrong locations
• Temp files (.tmp, .DS_Store, *~, *.swp, *.bak)
• Log files in root/package directories

Phase 5: Structural Validation

node scripts/fleet/check/paths-are-canonical.mts

Report errors as Critical findings. Warnings are Low findings. (The fleet's structural validator is paths-are-canonical.mts, the path-hygiene gate. If a repo has a richer structural validator under a different name, run that instead. Every fleet repo ships paths-are-canonical.mts.)

Phase 6: Determine Scan Scope

In interactive mode, ask the user which scans to run via AskUserQuestion (multiSelect). Default: all scans.

In non-interactive mode, run all scan types; no prompt.

Phase 7: Execute Scans

Run the enabled scans as a Workflow (not ad-hoc Task spawns). The scan set is independent fan-out + a dependent variant-analysis stage + a dedup/synthesize barrier — exactly what Workflow models, and the structured-output schema makes each finder return validated data instead of free text the orchestrator re-parses. The skill invoking Workflow is a sanctioned opt-in; pass the enabled-scan list as args.

Author the script inline (don't pre-Write it). Shape:

1. phase('Scan') — parallel independent finders. One agent() per enabled scan type whose prompt is the scan's reference.md section (legacy 1–8) or scans/<type>.md (modular). Each uses agentType: 'Explore' (read-only sweep), a FINDINGS_SCHEMA ({ scanType, findings: [{ file, line, issue, severity: critical|high|medium|low, pattern, trigger, fix, impact }] }), and runs under parallel(...) — variant-analysis is NOT in this batch (it depends on the others).
2. Barrier → dedup. Collect all finder results, .filter(Boolean), flatten findings, then dedupeFindings(...) from scripts/fleet/scanning-quality/lib/findings.mts (dedup by file:line:issue with a normalized issue key — genuinely needs all findings at once, so the barrier is justified). The dedupe key, the refute threshold, and the grade rubric all live in that one tested module.
3. phase('Variant') — dependent stage. For each High/Critical deduped finding, one agent() (the scans/variant-analysis.md prompt) searching the repo for the same shape; fold new variants in with mergeVariants(base, variants) (dedups across the combined set).
4. phase('Verify') — adversarial pass (thorough/release runs only): per High/Critical finding, spawn a skeptic that tries to REFUTE it ({ isReal, why } schema); dropRefuted(findings, votesByIndex) removes the ones a majority refuted (a tie keeps — the conservative direction). Skip for a quick scan — log() that it was skipped so the report doesn't read as fully verified. Skeptic prompts must demand INDEPENDENT re-derivation, not agreement: a vote whose why merely restates the finding's own reasoning counts as no evidence — the prompt requires the skeptic to (a) name the exact code path/input that makes the bug reachable, (b) for concurrency claims, walk the await boundaries explicitly (JS continuations between awaits are atomic — two synchronous statements after the same await can NEVER interleave with another caller), and (c) check whether the "bug" is documented intentional behavior before confirming. A run where skeptics refute ZERO findings is a credulity signal — the orchestrator must hand-verify the top findings itself before reporting (a real scan of this shape once confirmed 46/46 including two false positives).
5. Synthesize — a final agent() takes the deduped+verified JSON and writes the A-F prioritized markdown report (sections by severity, file:line refs, fixes, coverage metrics). The narrative is the agent's; the grade itself is gradeOf(findings) from the lib (the same A-F rubric scanning-security uses), so the two scanners can't disagree on a count→letter.

Return { report, findingCount, bySeverity } from the script (bySeverity = countBySeverity(findings) from the lib). Each finder's FINDINGS_SCHEMA replaces the old free-text "File / Issue / Severity / Pattern / Trigger / Fix / Impact" shape — same fields, now validated.

Phase 8: Save the report

The Workflow returns the synthesized A-F markdown. Save it:

• Interactive: offer to save to reports/scanning-quality-YYYY-MM-DD.md via AskUserQuestion.
• Non-interactive: save unconditionally to reports/scanning-quality-YYYY-MM-DD.md (create the dir if missing). If Write isn't in the allow list, emit the full markdown to stdout with a leading === REPORT MARKDOWN === marker so the runner can capture it.

Phase 9: Summary

Report final metrics: dependency updates, structural validation results, cleanup stats, scan counts, and total findings by severity.

Commit cadence

This skill is read-only. It scans and reports, it doesn't fix. Cadence rules apply to handing the report off, not to fixes:

• Save the report before acting on it. If the user opts to save (reports/scanning-quality-YYYY-MM-DD.md), commit the report file in its own commit (docs(reports): scanning-quality YYYY-MM-DD). That snapshot is referenceable later when fixes land.
• Don't fix in-skill. If findings need fixes, hand off to the appropriate skill (/fleet:guarding-paths for path drift, refactor-cleaner agent via /fleet:looping-quality for code-quality findings) and commit those fixes per that skill's own cadence rules. Don't bundle scan + fixes in one commit.
• One report per scan run. Re-running the skill produces a new report; don't overwrite the previous one's git history. Commit each fresh report so the trend line is visible.