Repository navigation
ci(bot): enable distillation and cross-review - #1815
Conversation
Signed-off-by: Vu Anh Phung <vu.phung@databricks.com>
There was a problem hiding this comment.
🔀 Cross-reviewed: two independent reviews (Claude Opus 5.5 · GPT-6.1 Terra), checked by a judge model (Claude Sonnet 5.5).
Verdict: 2 Nit
Looks good, with 2 low concerns about where the seeded lessons say they came from. The reviewer split between trusted workflow code and the PR checkout looks right. On pull_request events, edits to the workflow YAML itself still run from the PR's merge ref, which is expected given the fork==false gate.
Other findings
- ⚪ Nit — Several seeded counts don't match their cited PR lists: L-0050 and L-0068 say 2× but list 3 PRs, L-0057 says 1× but lists 2, and L-0012 says 3× but cites only #1707. The count may measure log entries rather than PRs, but the file doesn't say so. Make the counts match the cited sources, or document what a count measures. (Found by Claude Opus 5.5 · confirmed by the judge)
- ⚪ Nit — L-0072 cites #1466, which is the PR the agent tried to fetch, not where the lesson came from (author run #39734137201). Other lessons (L-0001, L-0027, L-0029, L-0031, L-0070) cite the legacy learning PR #1637 instead of the originating run. Cite the run IDs, or say 'author run', so provenance can be traced. (Found by Claude Opus 5.5 · confirmed by the judge)
There was a problem hiding this comment.
Verdict: 1 Low
Looks good, with one low-severity note: a seeded lesson cites the wrong source PR. The reviewer workflow now runs trusted code from the base commit and checks out the PR head separately with persist-credentials: false, which tightens security. Manual dispatch now also blocks forks and closed PRs. The new distill config keeps the legacy #1637 branch separate.
Signed-off-by: Vu Anh Phung <vu.phung@databricks.com>
There was a problem hiding this comment.
🔀 Cross-reviewed: two independent reviews (Claude Opus 5.5 · GPT-6.1 Terra), checked by a judge model (Claude Sonnet 5.5).
Verdict: 1 Low
Looks good, with 1 low concern. Checking out the trusted base, gating the PR to open non-fork PRs, and checking out the head separately as read-only are all sound. The archive keeps the #1637 entries. The only issues are some lesson counts and sources in lessons.md that don't match the archive.
Signed-off-by: Vu Anh Phung <vu.phung@databricks.com>
Description
Apply kernel #476's onboarding: enable distill mode with learning-PR auto-merge off, and pin the bot engine to 6014478 for two independent reviewers plus a judge. Manual reviews use trusted workflow code and a separate resolved PR checkout.
Preserve all 93 rules from #1637 at d6b0766 as a frozen archive and seed 73 lessons. New learning uses dated journal files and a fresh rolling branch.
Testing
98 focused engine tests passed. Config, archive preservation, lesson routing/source checks, seven reviewer-input cases, Bash syntax and actionlint passed. Live cross-review approved the current head with no findings.
Telemetry Errors
Additional Notes to the Reviewer
After merge, run learning once and inspect its first journal/lessons PR. Check #1637 for newer entries before retiring that legacy PR.
NO_CHANGELOG=true