Repository navigation
chore(agents): stacks merge forward, the ADR PR is reviewed first, and CI failures are triaged before a fix run - #798
Merged
Conversation
…d CI failures are triaged before a fix run stack-sync.sh merges main and each parent branch forward through a feature's PR stack in temporary worktrees, so no rebase and no force push; the deliver skill links the stack at build time, syncs the ADR branch and reviews the ADR PR before the first build, and sends a CI failure to an implementer only when ci-triage.mjs says the PR reaches it (reruns for flaky and unknown, a sync for base-inherited). The ADR index merges by union and src/adr-index.test.ts keeps it complete; ADR 0022 gets its missing row. PROCESS-REVIEW-2026-10-21.md items 9 to 12.
V3RON
marked this pull request as ready for review
October 21, 2026 08:54
V3RON
marked this pull request as draft
October 21, 2026 08:57
V3RON
marked this pull request as ready for review
October 21, 2026 09:04
V3RON
added a commit
that referenced
this pull request
Oct 21, 2026
…andidates are probed, mechanical fix-checks spawn no lens, the audit proves coverage (#799) ## What this is Layer four of `docs/internal/PROCESS-REVIEW-2026-10-21.md`: the review changes, items 13, 14, 15 and 23. Stacked on #798. - Item 13: round 1 reports every site of a confirmed defect class. The real #775 round 1 found one STRICT INTEGER column and round 2 its sibling; a replay found both in round 1, so the reach is there and the brief now asks for it. A sibling found in round 2 is recorded as round 1's miss. - Item 23: a candidate in a safety class is never downgraded to a note or dismissed at synthesis on reasoning alone; the coordinator probes it or it blocks. Both #752 replays lost the safety finding that way; one coordinator saw it and wrote "name-only matching is narrow". - Item 15: a fix-check whose every hunk is mechanical spawns no lens; the coordinator reads the diff and records it. One semantic hunk sends only the lens it reaches. - Item 14: the audit lists every changed function and proves a test goes red when its body is deleted (`covered <c>/<f>` on the audit line). About 20 of 45 review findings were test-coverage classes the audit should have caught. ## Verification Replayed with these rules through `review-replay.sh <pr> 1 a580434`, once each, against the baseline runs of the same snapshots (write-up, sections 4.5 to 4.8): | Snapshot | Baseline recall | With the changes | USD | Minutes | | --- | --- | --- | --- | --- | | #752 round 1 | 4/6 | 6/6 | 1.41 → 1.67 | 12 → 17 | | #775 round 1 | 6/10 | 8/10 | 6.47 → 8.07 | 34 → 37 | #752: the safety finding came back blocking with a reproduction, where both baseline runs had downgraded it to a note; EPERM-counted-dead stayed blocking for want of a probe (item 23). #775: all three STRICT INTEGER columns in one round-1 finding, with probes, where the real review needed round 2 (item 13); the socket probe found in round 1 again. Still missed on #775: the fake table accepting extra keys. Cost per round 1 is up about 20 percent; the rounds it should remove are measured on the next deliveries. Rows are in `.agents/replay/results.jsonl`. *Written by an agent.*
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this is
Layer three of
docs/internal/PROCESS-REVIEW-2026-10-21.md: the delivery and stack changes, items 9 to 12. Stacked on #796.stack-sync.sh <top pr>reads the chain from GitHub and merges forward, bottom to top, in temporary worktrees. A merge is a fast-forward push, so no rebase, no force push and no human in the loop;gh stack rebasefailed 11 of 14 times in the last runs, the one run that merged forward needed nobody. The deliver skill links each PR into the stack as soon as its draft exists. The ADR index merges by union (.gitattributes) andsrc/adr-index.test.tskeeps it complete and ordered; ADR 0022 had no row.main, so tasks stop inheriting red CI (the docs-manuals test was 8 of 28 CI lines).ci-triage.mjs <pr>sorts each failed test as flaky, base-inherited, in the PR's reach, or unknown; only reach gets an implementer, the rest a rerun or a sync.ci-failures.mjs --branch <name>reads a branch's newest run for the base comparison.The fix brief's
stack:line now means "merge the named parent into your branch"; delivery rule 15 says a stack moves by merging forward.Judged by
delivery-counters.mjs --since <date>on the next feature runs: CI lines sent to fix runs (26 over the last 20 merged PRs), stack syncs (12), empty builds (2), no-change fix runs (6); and gate refusals per run (two of four runs).Written by an agent.