D0015 — Workers that finish into nothing (W1–W4)

[obotclaw[bot]]

Four calls on the worker closeout contract you set today — the rule that every worker agent has to finish into something you can act on, and the check that enforces it.

Full artifact (recommended read, ~6 min): https://jwildfire.github.io/obot.roadmap/reports/decisions/2026-08-15-worker-closeout-check/

The situation. You said tracking a day of agent work now feels close to critical, and that keeping the roadmap current is the Navigator's job. Asked whether to gate at spawn or check at closeout, you chose closeout: every worker finishes into a pull request for a planned release, a question for you, or a request for config work only your hands can do. Measured over the last two days: 41 agent jobs ended, 9 wrote no close-out line, 19 had nothing captured automatically, and 3 real workers left no recoverable trace of what they did at all. Nobody noticed at the time.

The four calls:

W1 (D0015.1) — Is the three-outcome rule the definition of done? A worker with none of the three has not finished, it has evaporated, and today that is silent. Recommendation: adopt it — a worker with no outcome produces a named drift line, and "I looked and there's nothing to do" counts as a valid answer provided it is written down.

W2 (D0015.2) — How does the Navigator learn a worker closed? Recommendation: read the harness's own job ledger on the existing five-minute sweep. It is free, needs no new plumbing, and — unlike the scratchpad — sees workers that died as well as workers that finished. Subagents have no such record, so their work is attributed to the parent that spawned them; measured usage of that lane here is zero, so the cost today is nil.

W3 (D0015.3) — How do we tell which agent made a change? This is the hard one, and it is what your Navigator tab needs. Every agent write carries the same bot identity, so GitHub can never tell forty agents from one. Recommendation: two lanes, visibly distinguished — record changes as they happen from now on (the only complete, non-self-reported option), and reconstruct the past from what survives while marking every reconstructed row as such. The gaps get shown, not smoothed over.

W4 (D0015.4) — Does this live inside the existing roadmap audit, or beside it? Recommendation: inside. Same twenty rules, scoped to the closing worker and attributed to it, plus exactly one new rule the board audit cannot express. No second engine.

Requirement: #184 (milestone 2026q3, under the autonomy goal). Nothing is built — the work waits on these answers.

"Adopt all" is a complete answer.

---

This discussion was drafted by Claude Code using Opus 5 in an unattended sibling session and reviewed by @jwildfire