gc

/gc

Scan recent work for drift / entropy / AI slop — code that compiles and passes tests but has rotted away from the project's conventions: duplicated helpers, dead code, inconsistent error handling, stale docs, golden-rule violations. This is the garbage-collection sensor (Concept ⑥) — not git gc.

Like /verify and /test-plan, it uses a fresh-context subagent so the scan isn't biased by the agent that wrote the code. It is report-only — it never refactors in bulk.

Steps

1. Dispatch drift-analyst via the Task tool:

   Task tool with:
     subagent_type: drift-analyst
     prompt: Scan the changed/active code for drift against golden-rules.md plus the generic drift
             heuristics. Bound the scope to changed files (git diff vs HEAD) or the active feature —
             plus the one deliberate doc-drift exception: an unchanged *.md file still enters scope
             when it mentions a symbol this change touched. Follow the fixed report format.
             Read-only — recommend, do not refactor.
   

   If the user passed a path or feature id as an argument, name it in the prompt.

2. Surface the report verbatim — ### Golden-rule violations, ### Generic drift findings, ### Recommended actions, ### Verdict, ### Uncertainties. The point of fresh-context analysis is that the user sees the analyst's actual words.

3. If verdict is DRIFT FOUND, offer (do not auto-apply):

   • Fix specific findings within the current feature's scope — transparently (show each change), highest-severity first. Do NOT bulk-refactor or wander outside the active feature (that's a scope-jump — confirm first).
   • Capture a recurring finding as a golden rule via the capturing-golden-rules skill, so /gc catches it automatically next time.
   • Leave as-is and note it in session-handoff.md (e.g. deliberate debt to pay later).

4. If verdict is CLEAN, report it with the .harness-anchor/drift-<ts>.md evidence path. Note that a clean scan is a floor, not a guarantee (the heuristics are partial).