[kentwelcome]

Changes

Use spacedock-dev/behavior-diff-action@v0.1 in prepare, run and publish. Set core-ref to github.sha during prepare; downstream jobs inherit the frozen revision.

PR events select the merge revision; issue comments select the default-branch revision; manual dispatch selects the chosen workflow ref. Comparison head approval remains separate from executable-core approval. Credential and publication boundaries remain unchanged.

Verification

Actionlint 1.7.7 passed. Two independent read-only AI reviewers approved with no material findings.

Live verification

A separate synthetic test PR will be prepared. User explicitly selected PR creation only: no model calls or workflow dispatch are authorized. A future manual run must select this workflow branch and the test PR exact head, with fast mode enabled. Do not use an issue comment to test this unmerged workflow.

[kentwelcome]

Independent AI review — round 1

Base: a5dacb0. Approved patch committed as 068abb5.

WorkflowCorrectness: APPROVE. WorkflowTrust: APPROVE. Both reviewed the entire workflow and Action consumers. No material findings or decisions required; no unauthorized scope. Existing phases and authorization remain; only the Action reference and frozen core selection change.

Optional execution precautions from WorkflowTrust: review the complete executable core before releasing credentials; use synthetic public-safe data; manually select the updated workflow branch for an unmerged test, exact test PR head, and fast mode. An issue comment would exercise default-branch code instead.

Primary check: Actionlint 1.7.7 passed. Reviews were read-only; no live model-backed verification. User chose PRs only, so test execution remains unapproved and pending.

[kentwelcome]

Synthetic verification PR prepared: #46 (draft, disposable, do not merge). Its e2e-only change is excluded from automatic dogfood triggers. No live run was started, per the user’s PRs-only decision. Future approved manual verification must select workflow ref ci/dogfood-action-core-ref and the exact current #46 head, with fast=true.

[kentwelcome]

Correction: #46 is closed. Replacement test PR #47 is branched from and targets ci/dogfood-action-core-ref, as requested. Corrected-base integration independently approved. No live run started.

[kentwelcome]

Replacement automatic-trigger test: #48. #47 is closed. #48 branches from and targets ci/dogfood-action-core-ref and changes plugin/skills/behavior-diff/SKILL.md, matching the opened-PR path filter. User explicitly approved the automatic paid run with existing three-trials-per-side settings. No manual retries authorized.

[kentwelcome]

Post-review live verification

Reviewed head remains 068abb5; both independent approvals still apply. Format and Unit CI passed.

Automatic test PR #48 run https://github.com/spacedock-dev/behavior-diff/actions/runs/38412036227 completed successfully: prepare, run, and publish all succeeded. The published result reports Before 3/3 and After 3/3 trials complete, extraction complete, report available, runner exit 0. Its selected core SHA is 2b0de34, matching the core_revision visible in the run matrix. Result: #48 (comment)

Disposition: integration verification supports the workflow change; no new material defect observed. This proves this invocation completed and published its selected core, not general model correctness. The disposable test instruction remains confined to #48 and is not included in this merge.