replicatedhq/kURL #6211

Verified directly from supplied sources: the head-side fetch.py filter at line 111 adds and not i.get("cancelledAt") # cancelled instances are not failures (with finishedAt now on line 112), while the base-side file at lines 106-112 filters only on isSuccess/isUnsupported/isSkipped/finishedAt. This is an executable behavior change to which instances are reported as failures, introduced by this patch, and it sits in candidate.changedRegions (fetch.py line 111). The head SKILL.md (line 13 area) was updated to promise the new 'not cancelled' semantics, confirming the doc and code change together. The scope-integrity concern — a behavioral helper-script change shipping alongside docs/skill edits under a no-code framing — is therefore grounded in verified changed lines; reviewers relying on a docs-only premise would not scrutinize this one behavioral condition. The PR-body wording itself is not in supplied sources, but the decisive causal fact (the undisclosed-type behavioral filter change) is fully established; the declaration claim is not contradicted by anything supplied.

The changed line in fetch.py (head, and not i.get("cancelledAt")) makes the failure filter depend on a JSON response field that the skill's own documentation never establishes. The head SKILL.md lifecycle section (line 77) enumerates only dequeuedAt/finishedAt/isSkipped/isUnsupported as instance fields, and the new cancel section (line 89) describes cancelled_at solely as a DB column set by the cancel endpoint, documenting the endpoint response as {pendingCancelled, runningCancelled} — never cancelledAt as a serialized instance field. The serialization code (tgapi) lives in kURL-testgrid, a separate repository per AGENTS.md, so no source in this repo confirms the field name. The defect is change-introduced (the base filter did not reference cancelledAt), the trigger (analyzing a run containing cancelled-while-running instances) is realistic, and the hedged consequence (filter may be a no-op, cancelled instances reported as failures) and remediation (verify a live run.json and document the field, or add a defensive failureReason=='cancelled' condition) follow. While the camelCase convention documented for other DB columns (dequeued_at→dequeuedAt, finished_at→finishedAt) makes cancelledAt plausible, plausibility is not an established contract, and the candidate correctly scopes this as a low-severity doc-contract gap requiring verification. The mismatch between the code's new dependency and the skill's documented API surface is directly verifiable in the supplied head sources.