Loop Triage (L1, cache-first)

Cache is king. This loop does not explore the codebase. It reads cached knowledge and minimal host snapshots.

Mandatory load order (do not skip)

1. .github/project-manifest.yaml or .claude/project-manifest.yaml — paths, token_policy, loop_policy
2. LOOP.md — confirm level L1 for this run
3. loop-budget.md — respect max cache files and zero source files at L1
4. STATE.md — resume open queue; do not duplicate done items
5. docs/codebase/README.md + docs/codebase/.codebase-scan.txt — staleness only
6. ≤3 targeted sections from docs/codebase/CONCERNS.md, ARCHITECTURE.md, or CONVENTIONS.md (grep headings; no full-file read unless small)
7. Latest TODO/*.md — open items in ≤5 bullets
8. .copilot/memories/INDEX.md — ≤2 memories if relevant

Optional host snapshot (after cache only)

Only if needed for the report; cap at 3 commands:

• git branch --show-current + git status --short
• gh pr list --limit 5 (if gh available)
• gh run list --limit 3 (if gh available)

Do not run grep/find across source trees at L1.

Produce (L1)

1. Write reports/loops/YYYY-MM-DD-triage.md with:
   • Cache cited: list every file/section used
   • Priorities: ≤5 bullets from cache + TODO
   • Risks: numbered CONCERNS items touched
   • Host snapshot: branch, open PRs (if fetched)
   • Next: 1–3 human actions (no auto-fix)
2. Update STATE.md — Cache used, Open/Done, Last session
3. Append one row to loop-run-log.md
4. Stay ≤120 words in the executive summary at top of report

Chain verifier

Invoke .github/skills/loop-verifier/SKILL.md on the report artifact before marking run complete.

Anti-patterns

• Reading app/, src/, or large trees before cache load
• Auto-commits, auto-PRs, or fixes at L1
• Pasting large cache paragraphs into the report (cite by section only)