kenoxa/run-curate

Session: Generate per SPINE.md. When auto-triggered from build-finalize, inherit calling session's ID.

Phase Trace: per phase-audit.md table format. Log at gather, dispatch, synthesize, present, apply. Include action counts per type.

Phases

Subagent references (backticked): dispatch to subagent -- do NOT Read into mainthread.

| # | Phase | Agent | Reference | |---|-------|-------|-----------| | 1 | Gather | -- | -- | | 2 | Dispatch | @curator + @envoy | references/curate-envoy.md | | 3 | Synthesize | @synthesizer | references/curate-synthesis.md | | 4 | Present | -- | -- | | 5 | Apply (user-gated) | -- | -- |

| Phase | Base | Envoy | Cap | |-------|------|-------|-----| | Dispatch | 1 | 1 | 2 |

Envoy always dispatches (frontier multi-provider) — knowledge correctness warrants the cost.

1. Gather

Read AGENTS.md ## Project Knowledge section and each referenced knowledge file.

Auto mode (build-finalize passes knowledge_candidate: yes items): collect candidates alongside existing entries. Standalone mode (user invokes /run-curate): full review of existing entries. Also scan .scratch/ for session logs that contain entries with knowledge_candidate: yes — collect these as additional candidates using the same shape (what_was_attempted, result, assumption_corrected, knowledge_candidate). Other workflow skills may emit candidates in their finalize phases.

Staleness detection (incremental): for each knowledge file declaring paths: frontmatter, run git log --since=<updated_date> --name-only -- <paths>; skip files with zero changes. Build staleness matrix (knowledge_file, last_updated, changed_neighbors, stale) and pass to @curator.

Memo-drift audit (standalone mode only): for each .claude/.../memory/project_*.md file whose description references mutable state (model picks, tier mappings, dependency versions, infrastructure defaults), run git log --since=<memo_updated> --oneline -- <referenced_code_paths> to check whether code moved on. Surface drifted memos as follow_up items in the curate plan — do NOT auto-update memory files; the user owns memory content.

Glossary discovery (optional): check for UBIQUITOUS_LANGUAGE.md at project root. If present, extract a lightweight term list (term + aliases if available). If absent, record [GLOSSARY_SKIP: not found] in session trace and continue — do not suggest creating one.

2. Dispatch

Dispatch @curator and @envoy in parallel:

@curator — receives:

• candidates -- learnings with evidence anchors (auto mode) or collected from session scans (standalone)
• existing_entries -- parsed entries with file contents
• staleness_matrix -- from Gather phase
• glossary_terms -- lightweight term extract from Gather (omit when glossary absent)

Output: .scratch/<session>/curate-plan.md

@envoy — coverage gap discovery per references/curate-envoy.md. Base output path: .scratch/<session>/curate-envoy.md (run.sh writes per-provider files curate-envoy.<provider>.md). Advisory only — curator retains sole promote/update/prune authority.

Await both before proceeding.

3. Synthesize

Dispatch @synthesizer via references/curate-synthesis.md to merge curator plan + envoy per-provider outputs. Output: .scratch/<session>/curate-synthesis.md.

4. Present

Read .scratch/<session>/curate-synthesis.md.

Present grouped by action type (promote / update / prune). Show token budget impact. When staleness matrix has stale entries, highlight them. Include envoy GAP/WATCH items labeled [ADVISORY: envoy]. Include glossary findings labeled [ADVISORY: glossary] — these carry E1 ceiling and cannot gate promotion.

Require explicit user approval before proceeding -- never auto-apply.

5. Apply

Gate: user approval received. On approval, mainthread applies changes:

• Promote: create/update docs/<name>.md with telegraphic content + updated: frontmatter. Requires E2+ evidence anchor.
• Update: edit existing knowledge files in-place
• Prune: remove files + AGENTS.md entries
• Index: update AGENTS.md ## Project Knowledge with backticked paths + glosses

Skip actions the user explicitly declines. Report final state.

Anti-Patterns

• Promoting E0/E1-only learnings without evidence anchor
• Creating knowledge files for content derivable from code or git history
• Exceeding AGENTS.md token budget
• Promoting skill/agent operational details to docs/ knowledge files (belongs in skill/agent files)
• Treating envoy GAP/WATCH items as promotable without E2+ evidence
• Running staleness detection against all knowledge files regardless of changed neighbors (full-corpus sweep)
• Gating promotion on glossary alignment — glossary findings are advisory, not blockers
• Suggesting glossary creation when none exists — with-terminology owns that advisory