CLI output and trace limitations

Lines printed after a run

Line	Meaning	What to do
M3 run	The run ID created by this invocation.	Use it with --baseline and m3 upload.
M3 feedback	Path of this run's feedback.json.	Read selected fields from this path.
M3 verdicts	Count of test cases per verdict.	Informational.
M3 observations	Tool-error results and completed executions in this run.	A tool error is a server result with is_error=True, not a test failure by itself.
M3: no tests executed; skipped-only runs fail	Every selected test was skipped or deselected, and the run fails.	Unskip or select a test.
M3 execution timeout	One execution reached a deadline. It is printed once per timed-out execution: id, stage, elapsed, feedback.	Informational. The run's result is the pytest outcome, so a test that expects a timeout can still pass. Inspect the execution's diagnostics when the timeout was not expected.
M3 test manifest persistence was incomplete	Pytest results were not fully saved.	Saved history for this run is incomplete; do not use it as a baseline.
M3 required evaluations blocked finalization	A required=True evaluation did not pass.	The run is not successful even if test code caught the exception.
M3 CI excluded	m3 ci test only: the number of tests excluded by markers with ci=False.	Informational.
M3 feedback export failed	feedback.json was not written.	The run's database records remain; rerun to get a feedback file.

Trace limitations

A trace's limitations list says which evidence M3 could not fully observe or save. Check it before treating a missing value as evidence of absence.

Code	Where it appears	Routine?
capture_incomplete	Direct traces: always present. Every finalized direct trace records it, with completeness partial (sdk/src/m3/direct_trace.py:604-605). Agent traces: added when harness output was invalid, truncated, timed out, or duplicated.	Routine on direct traces. On agent traces, read the trace diagnostics before concluding that a call did not happen.
capture_disabled	Provider message and reasoning observations are dropped when provider-message capture is disabled (harness/observation_sink.py:85-89). Raw evidence is also not stored when raw capture is disabled (:123-126).	Routine when the corresponding capture is off. Other normalized harness observations are still recorded.
partial_trace	An ACP turn failed after recording only part of its evidence (harness/acp.py:1730).	No. The turn's evidence is incomplete.
cleanup_failed	A server process or harness could not be shut down cleanly.	No. Results stand, but check for leftover processes.
persistence_failed	Some observations could not be saved (harness/observation_sink.py:120,403).	No. The saved trace lacks evidence; do not treat a missing value as absence.