classify

Classify — propose work items from conversation

Flag-gated (spec §6): meaningful only when classifier.enabled: true in .work/config.yml. Propose, don't dispose — suggestions go to .work/suggestions.jsonl (gitignored, per-clone, append-only); nothing enters the event log until the main loop or the human promotes.

1. Spawn ONE background subagent

Spawn a single background subagent (never more than one) with this prompt:

Analyze the recent exchange for trackable work that has no work item yet. For each plausible item, append ONE JSON line to .work/suggestions.jsonl:

{"suggestion_id": "<ULID>", "source_span": "<turn or transcript ref>",
 "proposed": {"level": "task", "kind": "feature", "parent": null,
              "milestone": null, "title": "..."},
 "confidence": 0.82,
 "open_questions": [], "dedupe_against": ["<item ULIDs checked>"]}

• Mint each suggestion_id fresh: python3 -c "import sys; sys.path.insert(0,'bin'); import ulid; print(ulid.new())"
• Dedupe FIRST: check the current bin/worklog fold titles AND prior unconsumed suggestions already in .work/suggestions.jsonl. Never re-propose an item that exists or was already suggested; record what you checked in dedupe_against.
• Read classifier.min_confidence from .work/config.yml (default 0.7). If confidence < min_confidence, proposed.kind MUST be "triage" and the doubt goes in open_questions — never a confident guess on a load-bearing field.
• NEVER append to .work/todo.jsonl, .work/done.jsonl, or any event log. Suggestions only.
• NEVER ask the user anything — you cannot block. Write suggestions and exit.

2. Dispose next turn (main loop)

Next turn, read .work/suggestions.jsonl, skipping records already marked consumed:

• High-confidence suggestions: offer them for promotion — bin/worklog promote <suggestion-id> creates the real item.
• Low-confidence (kind:triage) suggestions: surface their open_questions to the user as questions, not items.
• Rejected suggestions: mark consumed by appending {"consumed": true, "suggestion_id": "..."} to .work/suggestions.jsonl.

The subagent gathers and proposes; the main loop, which can talk to the human, disposes.