[romamo]

What

• investigation_events() is the loop as a generator of Turn (the model's words, how many tools it called) and Called (the step and the tool's result); investigate() drains it, unchanged for callers
• agent investigate is a treaty streaming command: one line per turn, one per call (effect: created when it wrote a decision), the saved record last (kind: investigation), then the end line with meta.effects. --no-stream gives every event in one envelope; failures still end the stream with exit 92 or 93, after the saved record

Why

The hosted hackathon demo shows the agent working as it happens. It also makes the CLI watchable.

Contract change

agent investigate is unreleased, so its output changes before anyone depends on it: the record moved from data.investigation to the kind: investigation event.

Checks

• uv run pytest: 396 passed (CLI tests read the stream; a new test checks the order turn, call, turn, investigation and the end line)
• uv run ruff check ., uv run mypy: clean
• Live: Super against the replayed disk-full incident streamed 12 events over about 7 seconds

🤖 Generated with Claude Code

https://claude.ai/code/session_01APTdZZw8Jw9PQF4TD55wMv

[shipmill-romamo]

@romamo Decision needed: the stream can lose the investigation record. Should #39 merge before that's fixed?

Review result: MERGE AFTER SMALL FIX, with one regression left open. Pushed to the branch as fast-forwards, with no rewrite:

• 9347341: the exit 93 description and suggestion said "read data.investigation.spec.steps". The error line has no data now, because the record is on the line before it (kind: investigation). This commit fixes that and adds the first CLI test of the exit 93 path
• ea0c0a3: merges main after Require a person at a terminal to approve a decision #37 (treaty rc38). The tests read rc38's bare lines with _seq and the _summary end line, and the CHANGELOG entry describes that shape. The handler needed no change. 405 tests pass locally

The regression. Start cloudfall agent investigate ... and close its stdout mid-run, for example piping into a reader that exits early. The decision already proposed stays on disk, but no investigation record is saved, and the command exits 0. On main the same run keeps the record. This was reproduced on the branch and on main. A proper fix has two parts: the generator returns its record when it's closed, and the record gets a new status (say stopped) in the schema. The status is a stored-state change.

1. Merge now and file the regression as its own issue (recommended): agent investigate hasn't been released, the demo needs the stream, and the record status gets designed on the issue
2. Hold Stream agent investigate turn by turn #39 until it's fixed in this PR, with a stopped status in the investigation record schema
3. Hold Stream agent investigate turn by turn #39 and fix it without a schema change: when the reader goes away, save the record with its current status and exit non-zero

Reply here with a number or your own answer; shipmill takes this up on the tick after your reply.

[romamo]

Decision by @romamo, given in chat; posted by Claude on his behalf.

Merge #39 now, before the regression is fixed. The lost investigation record when stdout closes mid-run is fixed in a follow-up.