[shipmill-romamo]

Summary

The investigation record kept spec.answer and spec.trace[].toolCalls[].arguments unbounded after #52 capped the trace's reasoning and text (D-4). Both are the model's text and can quote host output. This caps them as D-6 (#59) settles.

• sdk/src/cloudfall/investigation.py: new ANSWER_LIMIT and ARGUMENTS_LIMIT (both TOOL_TEXT_LIMIT, 8000) and _head_text, the mirror of D-4's _trace_text. It keeps the start and ends with a \n[last N characters cut] mark, D-4's [first N characters cut]\n turned round. As in D-4, the mark counts against the limit, so a stored value is never longer than 8000 characters, mark included (a cut value can be one to three under). Characters are counted, not bytes
• Only the stored copy is cut:
  • the finding (and so status: answered) is still read from the model's whole answer, so a finding past the cut still counts
  • the raw arguments are cut in _call_text, which builds only the trace's ToolCallText. _run_call still parses the endpoint's own argument string, so the tool runs on the whole arguments and nothing parses the cut copy as JSON
• The stream's turn events carry the same Turn, and the last investigation event is the saved record's as_document(), so what agent investigate prints matches the record
• config/schemas/v1/agent-investigation.schema.json: maxLength: 8000 on spec.answer and trace[].toolCalls[].arguments, with descriptions, the same bound D-4 put on text and reasoning

Records already on disk: handled as #52 did. Stored investigation records are not re-validated on read, still true on main: InvestigationStore validates only on save, nothing in the SDK or engine loads investigations/, and evaluations/nemotron-incident/run.py reads them as plain JSON and uses only finding.proposed. A record written before with a longer answer or arguments stays readable everywhere it is read today; only revalidating it against the new schema would fail.

No schema lock or snapshot test covers this schema; none needed updating. One existing assertion changed: test_a_long_reasoning_is_cut_to_the_limit_and_marked asserted the record's answer stayed whole (that was #52's design); it now asserts the start is kept with the mark.

Out of scope, unchanged: spec.steps[].arguments (the parsed object that ran) and finding.rootCause/why stay unbounded.

Test plan

• test_a_long_answer_keeps_its_start_and_its_finding[8000|8001|50000]: exactly 8000 is kept whole; 8001 and 50,000 (Cyrillic) keep the start and the mark, the finding is still read, and the saved record passes the schema
• test_a_finding_past_the_cut_is_still_read_from_the_whole_answer
• test_the_cli_streams_the_answer_as_the_record_keeps_it: the stream's record event and the file on disk have the same cut answer
• test_long_raw_arguments_keep_their_start_in_the_trace_only[8000|8001|50000]: the tool receives the whole arguments, steps[].arguments is whole, the trace's raw string is cut with the mark
• test_the_schema_refuses_an_answer_or_arguments_past_the_limit[answer|arguments]: 8001 characters fail the schema
• uv run ruff check ., uv run mypy, uv run pytest (full suite, 476 passed), uv run cloudfall config validate --project config/examples: all pass

Fixes #57

🤖 Generated with Claude Code