Redaction
gstack's /cso repair runs stop failing on the helper's own file names
<finding id>-before-<plan hash>.jsonREDACTION_FAILEDWhen a gstack /cso evaluation ran a repair, the secret redactor read the finding ID in the helper's own file names as a credential, and failed the whole cell. Those names now pass. Anything else still gets redacted.
Tested claims
/cso counts a tested claim as false only when it could have been proven
In /cso's release qualification, twelve of the thirty-three claims marked falsely tested sat on the evaluator's own fixtures. Those now count neither way. A claim on a real incidental issue needs a passing replay of its signed repair.
Unchanged
Every other /cso bar stays where it was
replayEvidenceHash is rejectedThe rest of the bar stays put. Recall, repair, held-out repair, the version two baseline and every threshold still need the seeded issue. A malformed or reused replay hash is rejected.
Your repo
Which repo would you point /cso at first?
CHANGELOG.md"If you…" notes for 1.91.136.0cso-eval.tswriting an evaluatorIf you write an evaluator for /cso, the upgrade notes cover the new replay field. So, which repo would you point /cso at first?















