Open Loops

Carry the thread's unfinished business across the wall between surfaces.

Overview

A long-lived agent chats in a thread — "I'll charge the car later", "let's go with plan B", "I'll clean up the README" — then a scheduled job fires with zero memory of that thread and either repeats a finished task or silently drops a pending one. The conversation and the cron/trigger surfaces don't share a brain.

open_loops.py reads a transcript and extracts only what is still open:

• → commitment — "I'll do X", "할게요" (closed when a later turn says it's done)
• ⏳ deferral — "later", "나중에", "follow up" (closed on completion)
• ◆ decision — "let's go with X", "하기로" (recorded; never auto-closed — it persists)
• ? open_question — a turn ending in ? (closed by the first substantive reply)

It emits a small, stable JSON ledger the scheduled surface loads before it acts — the missing handshake between a thread and everything that fires later. Bilingual (EN + KO), pure standard library, deterministic. A ruler for handoff hygiene, not an LLM.

When to use

• Cross-surface handoff: a cron job or trigger loads the ledger for the relevant thread before running, so it knows what the owner is still waiting on.
• Session handoff hygiene: before ending a session or spawning a fresh one, gate on open-loop count — refuse a "clean" handoff that leaves 8 promises open.
• Self-audit: an agent checks whether it left commitments dangling in a thread.

Not for: summarizing a conversation (this tracks only the unresolved), sentiment, or verifying that a closed loop was actually done (it trusts the "done" signal).

The method

1. Get the transcript. JSONL ({"role","content"} per line or a JSON array), or plain text with [speaker] message / speaker: message lines.
2. Extract the ledger.

   python scripts/open_loops.py thread.jsonl --json
   # or:  cat thread.txt | python scripts/open_loops.py -

3. Feed open_loops to the scheduled surface. In a cron/trigger prep step, read the JSON and inject the open commitments/deferrals into that run's context so it acts with the thread, not blind to it.
4. Gate a handoff. Refuse to call a handoff clean when too much is dangling:

   python scripts/open_loops.py thread.jsonl --max-open 5   # exit 1 if >5 open

5. Re-run after closing loops. Say "done"/"끝났어요" in the thread and watch the commitment move from open to closed on the next extract. Open count is the score.

Anti-patterns

• Treating it as a summarizer. It deliberately ignores everything that isn't an open loop. If you want a recap, use a recap tool.
• Trusting a closed loop as verified. A loop closes on a "done" signal, not on proof the work happened — pair with real verification for anything that matters.
• Over-tuning the regexes into false positives. A noisy ledger a cron job trips over is worse than a conservative one that misses an edge case. Keep it tight.
• Auto-closing decisions. Decisions persist by design — they describe a standing choice, not a task; don't "resolve" them just to zero the count.

Example

$ printf '[민서] charge the car tonight\n[bob] set the charge later\n[민서] and what do you think of plan B?\n' \
    | python scripts/open_loops.py -
OPEN LOOPS: 2  (commitments 0, deferrals 1, decisions 0, questions 1)  | closed 0
  ⏳ [deferral] (bob) set the charge later
  ? [open_question] (민서) and what do you think of plan B?

The cron job that fires at 9pm can now read that ledger and know the car charge is still owed — instead of firing blind.