Repository navigation
agent investigate exits 0 when saving a stopped investigation fails #56
Description
What happens
Found in the review of #54 (#44). When agent investigate streams and the reader closes stdout mid-run, the run saves the investigation as stopped and exits 94 INVESTIGATION_STOPPED. If that save fails (repro: make investigations/ a file instead of a folder, then close the reader after the first line), treaty prints a traceback on stderr and the command exits 0. The stream_stopped flag is never set because the save raised first.
That breaks D-3 (docs/decisions.md): a run that couldn't finish must exit non-zero. Before #54 the same case also exited 0, so this isn't a regression.
Expected
A failed save of the stopped record exits non-zero with no traceback, probably with the code the repo already uses when the investigation store can't be written. Code: sdk/src/cloudfall/app.py, the GeneratorExit handler in agent_investigate.
Activity
shipmill-romamo commentedon Oct 9, 2026
Triage: implement, once the exit code is settled. Waits on your decision in the next comment.
Confirmed on
main(after #54):InvestigationStore.save(sdk/src/cloudfall/investigation.py) lets anOSErrorfrom the write escape unwrapped, so it isn't only the stopped path: any run whose record can't be written ends in treaty'sHANDLER_CRASHEDtraceback. In the stopped path the crash happens beforestream_stoppedis set, so the exit stays 0, which breaks D-3. A fix wraps the write error in the store and maps it inagent_investigate, which needs an exit code (a contract).shipmill-romamo commentedon Oct 9, 2026
@romamo Decision needed: what should
agent investigateexit with when it can't write the investigation record?PRECONDITION, with the store's error naming the folder (recommended): the store's folder isn't writable, the same kind of refusal the repo already gives for a bad--decisionsfolder; for a stopped run, the record is lost but the exit is non-zero, as D-3 asksRECORD_UNWRITABLE): a script can tell "couldn't save" apart from other preconditions, at the cost of one more code to documentReply here with a number or your own answer; shipmill takes this up on the tick after your reply.
romamo commentedon Oct 9, 2026
1 — exit
PRECONDITION, with the store's error naming the folder(decision by @romamo, given in chat; posted by Claude on his behalf)
shipmill-romamo commentedon Oct 9, 2026
Triage: implement, with the answer above (option 1), recorded as D-5 in #59.
Design:
InvestigationStore.save(sdk/src/cloudfall/investigation.py) wraps anOSErrorfrom the write in the store's own error, naming the investigations folderagent_investigate(sdk/src/cloudfall/app.py) maps that error toPRECONDITION, with no traceback, on every path that saves: the normal end, theGeneratorExit(stopped) path, and any other savePRECONDITIONinstead of 0; its record is lost, as the answer acceptsA PR follows.
shipmill-romamo commentedon Oct 9, 2026
PR: #61.
InvestigationStore.saveraisesInvestigationError(agent_investigations_unwritable, naming the folder) when it can't create the folder or write the record, andagent investigateexits 4PRECONDITIONwith no traceback. On the stopped path, treaty would print a raised error as a traceback and exit 0, so the handler keeps the error andcloudfall.cliwrites it to stderr as one JSON line and exits 4. Trade-off: that stopped run's record is lost, as D-5 accepts.Merge pull request #59 from cloudfall-dev/docs/decisions-56-57-58
shipmill-romamo commentedon Oct 9, 2026
Released in v0.6.0rc10.