Find Wasted Code (static friction-footgun audit)

The active, static inverse of a papercuts log. A papercuts tool is passive — an agent must notice friction and file it. This audit scans the repo for the same north-star footguns and writes them down, so nobody has to notice first:

Agents hit friction constantly — dead-end tool calls, broken links, missing helpers, footgun configs — and silently push through without telling anyone.

It is the third audit sibling, on a distinct axis:

How to run

1. python hooks/local/find-wasted-code.py (if python is missing, python3 …). Optional: --print (summary only, no write), --date YYYY-MM-DD (deterministic in-report date), --selftest (per-rule golden fixtures), --root PATH.
2. Read the report it wrote (docs/wasted-code/report.md): the per-rule findings + the Coverage section.
3. Interpret every row as a review candidate, never a verdict or an auto-fix. broken/confirmed = the target is provably absent from repo state; candidate/inconclusive = needs human judgment. Ambiguous references are listed under Coverage → Unresolved, never as defects.
4. Output a Mode-A chat summary: totals by rule/severity, the W5 baseline counts, and the top confirmed broken items. Do not paste raw file contents.
5. The report is tracked (papercuts philosophy — it shows up in diffs). It is redaction-safe by construction and will not trip the pre-commit secret scanner. Fixes are the operator/PO's call — this audit surfaces, it does not mutate.

The rules (one per north-star category)

Per-rule signatures, the conservative confirmed-vs-coverage contract, and the false-positive classes: references/rule-catalog.md and references/false-positive-examples.md. Cite them; don't restate.

Conservative by construction (load-bearing)

A finding is confirmed ONLY when provable from repository state. Everything ambiguous is Coverage, never a defect — so the audit never blocks or annoys the operator with a false positive. This is the same "don't block the human operator" floor the framework holds everywhere. W5 is deliberately a measured baseline, not a finding list: this repo intentionally fails open/closed in many places (best-effort cleanup, fail-closed scanners), so generic swallow-detection cannot be low-false-positive — the operator reviews the baseline and annotates intentional swallows with find-wasted-code: ignore W5 — <reason>.

Trigger isolation

• Claude Code: hard guarantee via disable-model-invocation: true — the model cannot auto-load this skill; only /find-wasted-code runs it.
• Other adapters (Codex/Cursor/Gemini): advisory — the narrow description above will not match ordinary work, and no hook wires the analyzer. Invoke by name or /prompts:find-wasted-code.
• No lifecycle hook (SessionStart/Stop/PostToolUse/UserPromptSubmit) references this skill or its analyzer. It is inert until the operator triggers it.

Read-only-to-the-project

The analyzer writes ONLY docs/wasted-code/report.md, containment-checked to inside the repo, symlink/hardlink-refusing, atomic, and sentinel-guarded (it will not overwrite a hand-authored file lacking its generated-file sentinel). No memory, overlay, spec, policy, or source edits. Fixes are proposed, never applied.

Non-Claude surfaces

Invoke the find-wasted-code skill by name and run the analyzer directly (python hooks/local/find-wasted-code.py); it is stdlib-only and surface-independent. Metrics come from repo state, so there is no transcript dependency.

Clean-room note

Original Fusebase Flow content. Designed after reviewing public agent-friction and self-improvement projects (papercuts, agentlogs, friction-logging-toolkit, self_improving_coding_agent, opencode-autolearn) for concepts only — the north-star framing and severity vocabulary. No third-party code, prompts, skill files, or hook scripts are copied; the analyzer, rule contract, and report format are clean-room original. See docs/source-map.md.