Repository navigation
Protocol friction from the trajectory: refusals per run by author model (#487 part 3) - #577
Conversation
…s operator notes (#487)
There was a problem hiding this comment.
Round 1 — reviewed head cf4d3215 — reviewer hermes/gpt-5.6-terra.
terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.
Verdict: 1 blocking, 0 advisory.
1 finding attached to the lines below.
Malformed trajectory event data can crash the new summary.
There was a problem hiding this comment.
Round 2 (re-run on the same head) — reviewed head cf4d3215 — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.
terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.
Verdict: 2 blocking, 1 advisory.
3 findings attached to the lines below.
Verdict: request changes. Blocking: trajectory summary includes panel judge sessions in its author metric, and malformed non-object event data can crash the command. Suggestion: clarify the changelog sentence with the supplied rewrite. Rejected: none; coverage, credentials, and general were merged as the same malformed-data defect, and deployment and lifecycle were merged as the same panel-author defect.
Part 3 of #487: protocol friction as a release metric.
trajectory.summarize(events)countssyscall.decisionevents, refusals by verb, and the distinctbackend/modelpairs fromsession.start. There are no new events and no format change.outerloop trajectory <run-id>ends with one line such asrefusals: 3 of 4 decisions (submit 2, message 1); author: claude/model-x. It prints nothing extra for a run with no trajectory, and the--jsonoutput shapes are unchanged.outerloop message, which the part-1 review flagged as missing.Gate: ruff check, ruff format --check, mypy, pytest (3192 passed, 7 skipped).
Built by codex (GPT); reviewed by Claude.