Repository navigation
Gate integrity: one prompt, a status for every verdict, the range on the fact - #102
Merged
Merged
Conversation
The reviewer prompt now exists once, in coordinator/review.py. The bash
copy (_recovery_review_prompt in batch/run-queue.sh) and the test that
kept the two byte-identical are gone.
review_prompt and dispatch take a base parameter and tell the reviewer
the PR's merge-base range (git diff origin/{base}...{co}), not just the
pinned head commit -- a reviewer once read "EXACTLY commit X" as "only
the head commit's own diff" and passed a 38-file PR after reading one
file. wiring.py's do_review resolves the PR's base via gh and passes it
through.
Deleting 80 lines from run-queue.sh moved every line-numbered citation
below it, so docs/design/effect-site-inventory.md and
docs/design/scar-effect-matrix.md are repointed with
tools/repoint-citations.py (baseline HEAD) to keep test_routing.py and
test_scar_equivalence.py green.
…llow the reviewers The status key was `status:<head>:<round>`, and the round only advances on a recorded review_verdict fact -- which an `error` or a terminal FAIL does not write. A later PASS on the same head therefore replayed the earlier key and the kernel returned the prior external id without executing: the gate kept showing the superseded verdict. The key is now content-addressed over the state and the description, as the PR comment's key already was. blocking_count missed the heading shapes live reviewers actually write: `## Blocking findings` and `### **Blocking findings**` counted nothing, a `##` closing heading did not close, a prose trailer on a closer did not close, a parenthetical section count was counted as a finding, and the phrase continuing as prose opened a section and counted the sentence. Fixtures for every one of those, measured. Also: the reviewer's setup fetches the base branch it is told to diff against, so `origin/<base>` exists in its clone on a non-default base; the merge-base gets the same 40-hex validation the head gets before it reaches the tuple, the description and the runner's hand-built JSON; `delta_digest` catches OSError so a missing `gh` binary is an unprovable delta rather than a traceback; the status log line says "recorded", since an idempotent replay posts nothing; and `_round_number` has a test proving it ever returns more than one.
…ict carries it _post_cross_review_status's default description is legacy text that names nothing, and it fires in exactly one case: the derivation's own post did not land, or could not be read back. There it OVERWROTE the merge head's range-naming description -- the status ceasing to attest to anything at the one moment it is the only record. run_item holds the merge-base and the observed head at its merge_ready_pr call site, so it composes the fallback description there and merge_ready_pr threads it through; every other caller passes none and keeps the legacy text, which is all that can honestly be said when the range is unknown. recover_pr_cmd parsed the tuple's two range fields into `_ _` and recorded a verdict with no range at all, though it drives the same derivation run_item does. It now reads them and passes them as record_review's arguments 9-11. Minors: the composed re-stamp description is truncated to GitHub's 140 (it is longer than that by construction, and the overflow cut the suffix saying it was a re-stamp); the two delta CLI calls no longer discard stderr, and the escalation says the CLI failing is one of the causes; the status POST quotes its state argument like every other. Tests: the self-test drives merge_ready_pr both ways and reads the description off the POST (verified by mutation: dropping the argument fails it); the recovery's reader width is asserted against Derived.FIELDS beside run_item's.
…ange ARCHITECTURE.md still described the derivation's boundary as an eight-field tuple and listed "the eight-field boundary's fail-closed width check" among the things that are done. It is ten, and the two new fields are the point of this branch: the merge-base and the three-dot delta digest of what the reviewer read. Both are named, with what they carry and why. skills/cross-review/SKILL.md carries a SECOND reviewer template -- the lead session's, pinned to FETCH_HEAD -- and it said nothing about a range, so the prompt fixed in the coordinator was still wrong in the place a human copies from. It now carries the same sentence, adapted to its placeholders, and fetches the base branch it is told to diff against.
3 tasks done
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
1 participant
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
Gate integrity, the first of the five PRs in
docs/superpowers/specs/2026-09-19-back-half-closed-loop-design.md(§4).origin/<base>...<head>; the bash copy and the byte-identity test are gone.bircher/cross-reviewfor every verdict:success/failure(with the blocking count) /error, described as<vendor> round <n> <word> on <merge-base7>..<head7>, through the journalled effect path.head_sha,merge_base_shaanddelta_digest(schema version 2); the runner passes them through a ten-field derivation tuple.Why
On muesli PR #769 a cross-review of one commit posted a green gate on a 38-file PR, and a later FAIL posted nothing, leaving that green standing. The record is
docs/design/front-half-live-log.md. Supersedes #101.Verification
cd v2 && python -m pytest -q: all green.bash batch/run-queue.sh --self-test: rc 0 as a non-root user (the runner's root-only self-test failure is pre-existing and unrelated).488bee5,ecbb7ef,56eef8e; three Minor residuals on the blocking-finding count's heading detection are recorded for a follow-up issue.