checkpoint

Checkpoint

Write CHECKPOINT-{stream}-llm.md: a self-contained session snapshot with two layers.

• Summary — a synthesized, reworded view of state (goal, next steps, decisions, hot files). Cheap to regenerate; good enough for orientation.
• Meat — the verbatim signal that reconstruction would lose: the exact phrasing of constraints, rejected paths + why, subtle corrections, live reasoning threads, learnings, pivots, open questions. Copied byte-for-byte from the transcript, never reworded.

Core contract (two paths, both keep model text out of the meat):

1. Author trailers (primary) — each turn self-tags its own signal in an <!-- ckpt … --> comment as it's written; the skill just harvests them (0 LLM, 0 cost, highest fidelity — the author distilled live). Requires the global CLAUDE.md rule; see reference.md § Trailer harvest.
2. Extractive triage (fallback / user-side) — the LLM DESIGNATES which turns matter (returns ids); a script COPIES their text verbatim. Covers the user's exact words (untaggable) and any session predating the rule. See reference.md § Why.

Meat text NEVER passes through a model in either path — no hallucination, no output cost.

Architecture — fork the pipeline, keep the offer in main

The pipeline (index → harvest → triage → collect → write) is mechanical and noisy: Bash output, a triage round-trip, assembler byte-counts. None of it belongs in the main conversation. So run Phases 1–3 in a fork; keep Phase 4 (report relay + /clear offer) in main.

MAIN (this skill, context: main)
 └─ spawn ONE fork (Agent, subagent_type: "fork" — inherits full session context)
     └─ fork runs Phases 1–3: index, harvest, triage (may spawn its own Task sub-agent),
        collect, synthesize summary, WRITE CHECKPOINT-{stream}-llm.md, clean up $TMP
     └─ fork returns ONLY a compact report string (never the meat, never tool output):
        path · #meat turns · histogram(type+weight) · status(full|lean|parse-failed)
 └─ MAIN receives that string → Phase 4: relay it + AskUserQuestion /clear offer