/curate — Canonical Miner Skill

Role: miner. Input = trinity slice (transcripts, .agents/, bd, git). Output = skill diffs (proposed), bd updates, rare wiki entries. Never mutates code.

Status (2026-05-08): introduced ADDITIVE in Phase 1 (m6v5.D.1 / soc-78s2v). Existing miners (dream, harvest, forge, compile, retro, post-mortem-mining, flywheel, trace, provenance, defrag) stay until Phase 3 shim conversion (m6v5.D.3). Fix-C smoke (soc-wb2aa) gates Phase 3.

Modes (≤8 per Fix-F mode-flag budget)

Mode-budget assertion: 8 modes. Adding a 9th requires demoting an existing one OR refusing the addition (per Fix-F § continuous CI gate).

Anti-goals (hard constraints from architecture):

• NEVER mutates source code.
• NEVER invokes /rpi or any code-mutating flow.
• NEVER performs git operations (no commits, branches, push, rebase, checkout).
• NEVER creates symlinks anywhere.

Quick Start

/curate --mode=harvest                  # cross-rig promotion sweep
/curate --mode=forge                    # mine the most recent session's transcript
/curate --mode=dream --duration=8h      # overnight bounded run
/curate --mode=compile                  # rebuild .agents/ corpus
/curate --mode=retro                    # capture this session's learning
/curate --mode=defrag                   # knowledge defragmentation
/curate --mode=watch                    # in-session drift check
/curate --mode=provenance --bead=soc-X  # walk decision trace for a bead

Execution

Step 1: Resolve mode + scope

Parse --mode. Each mode has its own scope semantics:

Step 2: Acquire lock (when applicable)

For --mode=dream: acquire .agents/overnight/run.lock exclusively. If held, exit with "another curator holds the lock" message; do NOT block.

For --mode=harvest: check the dream lock. If held, defer harvest to next loop fire.

For --mode=forge: per-session lock at .agents/forge/session-<id>.lock. Auto-released at end of run.

Other modes: no lock; concurrent invocation is safe.

Step 3: Run the mode-specific body

Each mode delegates to a body section in this skill (see § per-mode bodies below for outline; full content moves to references/<mode>.md at Phase 2 startup).

Step 4: Produce artifacts

Output is one of (priority order, per architecture knowledge-flywheel rule):

1. Skill diffs — proposed changes to existing skill bodies, written to .agents/skill-diffs/<date>-<skill>.diff. Operator approves before applying. NEVER writes to skills/ directly.
2. bd updates — bd note, bd remember, or new bd create for surfaced issues. Direct, no approval queue.
3. Wiki entries — .agents/research/, .agents/learnings/, ~/.agents/learnings/ (rare; only when knowledge is generally reusable).

Step 5: Append to LOG.md

Every mode appends one line to .agents/LOG.md:

2026-MM-DDTHH:MM:SSZ [curate --mode=<mode>] <run-id> — <short-summary> — outputs: <artifact-paths>

Step 6: Report

1. Mode + scope
2. Output path(s)
3. Surfaced bd issues (if any)
4. Loop continuation hint (next-fire cadence per architecture catalog)

Per-mode bodies (outline)

--mode=dream

Overnight INGEST → REDUCE → MEASURE iterations until halt:

• INGEST: scan .agents/ for new artifacts since last run
• REDUCE: dedupe + score + cluster
• MEASURE: compute knowledge-corpus health metrics
• Halt when: wall-clock budget exhausted, plateau (K sub-epsilon deltas), regression beyond per-metric floor, metadata integrity failure
• Knowledge-only; never code mutation
• Output: .agents/overnight/<run-id>/summary.{json,md} + per-iteration archive

Detailed body remains inline until Phase 2 extraction.

--mode=harvest

Cross-rig promotion sweep:

• Walk .agents/ across all rigs (paths from ~/.agents/rigs.yaml or fleet config)
• Extract learnings/patterns/research artifacts
• Dedupe by content hash
• Promote high-confidence items to ~/.agents/learnings/ global hub
• Roll up flywheel-health metrics as byproduct
• Output: .agents/harvest/latest.json + promoted files in global hub

Detailed body remains inline until Phase 2 extraction.

--mode=forge

Per-session transcript mining:

• Locate latest transcript (~/.claude/projects/<project>/<session>/*.jsonl)
• Extract knowledge candidates (decisions, patterns, anti-patterns, bug fixes)
• Validate candidates against finding-registry contract
• Queue to .agents/knowledge/pending/ for curator review
• Output: pending markdown files; bd notes for high-confidence findings

Detailed body remains inline until Phase 2 extraction.

--mode=compile

Corpus pipeline:

• Mine: extract candidate knowledge from .agents/
• Grow: merge into existing taxonomy
• Defrag: collapse redundant entries
• Lint: check for orphans, contradictions, staleness
• Output: wiki/INDEX.generated.md (rebuilt), .agents/compile/<date>.md (lint report)

Detailed body remains inline until Phase 2 extraction.

--mode=retro

Single-session learning capture:

• Read last N turns + diff summary
• Identify one durable insight (or none — exit clean)
• Append to .agents/retro/index.jsonl
• Optional: surface to bd as note

Detailed body remains inline until Phase 2 extraction.

--mode=defrag

Knowledge defragmentation:

• Find duplicate entries across .agents/ (content-hash + semantic-similarity)
• Find broken backlinks
• Find stale entries (last_cited > TTL, low hit_count)
• Output: .agents/defrag/<date>.md with proposed retirements (NEVER auto-deletes; operator approves)

Detailed body remains inline until Phase 2 extraction.

--mode=watch

In-session drift detection:

• Read last 100 transcript turns
• Detect: research loops without code change, repeated grep-without-read, oscillating decisions
• Write advisory to .agents/watch/<date>.md; surface high-severity to bd note
• Cheap; designed for 15-min cadence

Detailed body remains inline until Phase 2 extraction.

--mode=provenance

Decision-trace walk:

• Given anchor (bead ID, file path, or decision marker)
• Walk: bd graph (parents, blockers, refs) + git log (commits touching anchor) + .agents/ mentio