Repository navigation
Investigation record: answer and raw tool-call arguments are still unbounded #57
Description
What happens
Found in the review of #52 (#50). #52 caps each turn's reasoning and text in spec.trace at 8000 characters (D-4), but two fields of the investigation record still have no length limit:
spec.answer(config/schemas/v1/agent-investigation.schema.json, set insdk/src/cloudfall/investigation.pyfrom the model's final text) holds the whole final answer, so the last turn's text is stored uncut in the record even though its trace copy is now cappedtrace[].toolCalls[].argumentskeeps the model's raw argument string unbounded
Both can quote host output, the same reason D-4 capped reasoning.
Question
Should these get the same 8000-character cap with the cut marked? answer is what the finding is read from, so cutting it may need a different rule (keep the start, or refuse instead of cut).
Activity
shipmill-romamo commentedon Oct 9, 2026
Triage: implement, once the rule is settled. Waits on your decision in the next comment. Confirmed on
main: neitherspec.answernortrace[].toolCalls[].argumentshas amaxLength, andinvestigation.pystores both as the model returned them. The change bounds stored state and a published schema, so it goes through the design gate (D-4 covers only the trace's reasoning and text).shipmill-romamo commentedon Oct 9, 2026
@romamo Decision needed: how should the investigation record bound
spec.answerand the rawtoolCalls[].arguments?answerfrom the end (keep its start, where the finding is) and the arguments from the end too, with the cut marked (recommended)answerwhole, since it's the finding and the model writes it, not the hostReply here with a number or your own answer; shipmill takes this up on the tick after your reply.
romamo commentedon Oct 9, 2026
1 — cap both
spec.answerandtoolCalls[].argumentsat 8000 characters, keeping the start and marking the cut(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-6 in #59.
Design:
spec.answerand eachtrace[].toolCalls[].argumentskeep at most 8000 characters, keeping the start (unlike D-4's reasoning cap), and mark the cut the same way D-4 doesconfig/schemas/v1/agent-investigation.schema.jsonbounds both fields with amaxLengththat allows the cut markerA PR follows.
shipmill-romamo commentedon Oct 9, 2026
PR: #62.
spec.answerandtrace[].toolCalls[].argumentskeep at most 8000 characters, keeping the start and ending with[last N characters cut]. The schema bounds both withmaxLength: 8000. Only the stored copy is cut: the finding is still parsed from the whole answer and the tool runs on the whole arguments. So a record can hold a finding whose lines are past the cut, and capping never turns a long answer intounstructured.Still unbounded and out of scope here:
spec.steps[].arguments,finding.rootCauseandfinding.why. That is filed separately.Merge pull request #59 from cloudfall-dev/docs/decisions-56-57-58
Cap the investigation's answer and raw tool-call arguments at 8000 ch…
shipmill-romamo commentedon Oct 9, 2026
Released in v0.6.0rc10.