Full quality scan

Skill silvio-l/skills/skills/full-quality-scan

My personal Claude Code skills. Engineering process, context hygiene, and AI-agent maintenance — straight from my .claude directory.

Install

npx -y skills add silvio-l/skills --skill full-quality-scan

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Runs all configured linters and security scanners (cppcheck, ESLint, Semgrep, osv-scanner, dart analyze) repo-wide, then fixes every finding. Use for a repo-wide quality or security scan, e.g. "scan everything" or "fix all lint issues".

SKILL.md

4.2 KB, as published. Nobody here has run it

Full Quality Scan & Fix

Overview

Threshold: ≤ 10 findings → fix directly. > 10 → plan first, then fix with subagents.

Phase 1 — Scan

Run the bundled wrapper. It is the single source of truth for tool flags and skip-rules — never duplicate the tool invocations into the agent's prompt.

bash skills/full-quality-scan/scripts/scan-all.sh
# or, when invoked as the installed skill:
bash ~/.claude/skills/full-quality-scan/scripts/scan-all.sh

Output contract — one finding per line:

BUCKET|FILE:LINE|MESSAGE
…
---
TOTAL_FINDINGS=<n>

Exit code: 0 = clean, 1 = findings present. The wrapper skips silently for tools that are not installed; the summary line is always emitted.

Bucket names emitted by the wrapper: dart, cpp, js_ts, sast, deps. Anything else is a wrapper bug, not an unknown tool — fix the wrapper.

Phase 2 — Triage

Parse each BUCKET|FILE:LINE|MESSAGE row and group by bucket. The bucket names map directly to fix strategies:

Bucket	Source tool	Fix strategy
`dart`	flutter analyze	`dart fix --apply`, then manual
`cpp`	cppcheck	manual C++ edits
`js_ts`	eslint	`eslint --fix`, then manual
`sast`	semgrep	manual, case-by-case
`deps`	osv-scanner	`npm audit fix` / `flutter pub upgrade`

TOTAL_FINDINGS from the wrapper's summary line is the threshold input:

≤ 10 total → summarize the findings, get the user's OK, then fix inline (Phase 4). Don't start editing files before the user confirms the fix list. > 10 total → build a plan in Phase 3.

Phase 3 — Plan (> 10 findings only)

Present a numbered fix plan grouped by bucket and severity:

PLAN
════
[deps]   1. Upgrade X (CVE-2025-…) — flutter pub upgrade / npm audit fix
[cpp]    2. floating_button_window.cpp:468 — unreadVariable tProg
[js_ts]  3. hero-carousel.ts:9 — prefer-const
...

Get user confirmation, then spawn one subagent per bucket with ≥ 3 findings. Set model explicitly on every spawn — fixing lint/security findings is Sonnet-tier work; never let the subagent inherit the orchestrator's model (subagent spend is the biggest cost driver):

Agent(description="Fix cppcheck findings in windows/runner/",
      model="claude-sonnet-4-6",
      prompt="Fix these cppcheck findings in windows/runner/: [list]. 
              Fix only the listed findings — do not touch or 'improve' surrounding code. 
              Run cppcheck after each logical group to verify. 
              Do not commit.")

Buckets with < 3 findings are fixed inline by the orchestrating agent.

Phase 4 — Fix

For each bucket (inline or via subagent):

Apply auto-fixable changes first (dart fix --apply, eslint --fix, npm audit fix).
Fix remaining findings manually — read the exact file/line before editing.
Re-run the bucket's tool after every logical group to confirm green.

Never suppress a finding without a written reason.

Phase 5 — Verify

Re-run bash scripts/scan-all.sh. The wrapper must exit 0 with TOTAL_FINDINGS=0 before proceeding.

If any finding remains: return to Phase 4 for the affected bucket only — do not loop through clean buckets again.

Phase 6 — Commit

Committing is irreversible — show the user the changed files and the commit message, get an explicit OK, then commit. Never auto-commit without confirmation.

bash scripts/scan-all.sh        # must exit 0
git add <changed files>
git commit -m "fix(quality): resolve <tool> findings\n\n<summary>"

The wrapper is the only quality gate at this stage; project-specific test/typecheck commands are out of scope for this skill and live in the project's own commit hooks or CLAUDE.md. One commit per major bucket is fine; one combined commit is also acceptable.

Tool availability

The wrapper skips silently for any tool that is not installed — its summary line is the authoritative record of what ran. Never install tools on behalf of the user without asking.