Remove Stale Code

Use the deterministic report as a lead, never as authorization by itself. This workflow handles only deletion_ready findings and revalidates every guard against the current checkout before editing.

Workflow

1. Start or resume a sessions.py maintenance session before edits.
2. Read logs/nightly-reports/stale-code.json. Stop if it is missing, malformed, has status != "ok", or contains no deletion_ready findings.
3. Compare subject_commit with git rev-parse HEAD. If they differ, run python3 scripts/stale_code_daily.py --dry-run-notify and reread the report.
4. Select only findings whose classification is exactly deletion_ready. Apply a user-provided fingerprint/path filter when present. Otherwise process at most five findings from one category in one run.
5. For every selected finding, recompute its SHA-256 fingerprint from category, subcategory, repository-relative file, numeric line (or zero), and code, joined with NUL bytes. Skip it if the fingerprint differs.
6. Reread the target and search the repository for the symbol and its dynamic forms. Skip it if references, metadata registration, glob loading, decorators, reflection, public exports, generated paths, migrations, fixtures, route conventions, or compatibility markers make usage ambiguous.
7. Apply only the smallest finding-specific change. For Ruff F401 imports, reread and remove only the selected import statement or selected imported name from a multi-name statement with apply_patch; never run a file-wide autofix. Then run ruff check --select F401 <file> and inspect the diff. For an unsupported finding kind, stop rather than inventing a deletion strategy.
8. Rerun python3 scripts/find_dead_code.py --json --all and confirm each selected fingerprint disappeared without creating detector errors.
9. Run the smallest targeted lint/test command for every changed file. Do not treat the detector rerun as a substitute for product tests.
10. Review the final diff. Use python3 scripts/sessions.py deploy only after all selected candidates and targeted checks pass. Never commit or deploy skipped candidates as if they were removed.

Refusal Rules

• Never edit review_only or suppressed findings.
• Never trust a report from another commit.
• Never bulk-remove more than five findings or mix categories in one run.
• Never delete a Svelte component or CSS selector solely from zero literal references.
• Never delete compatibility, public API, framework registration, generated, migration, fixture, or route code without a separately approved contract.
• Never continue when a required analyzer, reference search, or targeted test fails.

Output

Report selected fingerprints, removed findings, skipped findings with reasons, changed files, detector rerun result, targeted verification commands, and deploy commit when applicable.