fix(review): say where GitHub shows a review's verdict, so the body opens with a finding - #1362
Merged
Merged
Conversation
…th a finding Self-review COMMENTs across consumers open with "Reviewing my own PR, so this is a COMMENT rather than a verdict" (earlier, "...an approval"). No instruction contains the phrase. The review skill frames every review as a verdict, and on a self-authored PR, where APPROVE is off the table, a session sometimes spends the opening line explaining the missing one. Once posted, the line spreads: each later round reads the earlier bot reviews in full for dedup and mirrors their form. Replaying the three PRQL/prql#6352 review sessions, cut just before each wrote its body (4 runs per arm), the opener appeared 0/4, 4/4 and 4/4 in rounds 1-3. With a sentence noting that GitHub labels the review approved or commented above the body, so the body opens with the first finding, round 2 kept the line in 1 of 8 runs, moved down from the opening; round 3, which had two earlier reviews to copy, kept it in 7/7. Naming the self-authored situation in the prompt raised the rate in a synthetic A/B, so the sentence sits in Submit and applies to every review. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…e them The review states (`APPROVED`, `CHANGES_REQUESTED`, `DISMISSED`) were already code-formatted in skill prose; the event names a session submits (`COMMENT`, `APPROVE`) were bare caps in 19 places. Both are API values, so they now read the same. This doesn't change the self-review opener: replaying the PRQL/prql#6352 review sessions, round 2 kept "so this is a COMMENT rather than a verdict" in 4/4 runs with backticks alone, and backticks on top of the Submit sentence matched the sentence alone (1/4), with the one surviving line now reading `COMMENT`. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
tend-agent
approved these changes
Sep 21, 2026
Merged
max-sixty
added a commit
that referenced
this pull request
Sep 22, 2026
Bumps the generator to 0.2.15, syncs `uv.lock`, and adds the 0.2.15 changelog section. The release ships the restrict-updates ruleset requirement and the `AGENTS.override.md` pin (#1367), the follow-up review that focuses on the pushed commits (#1368), a rising bar for further fix rounds on PRs with no human author (#1366), #1365, #1362, and the setup-uv v10.2.0 bump (#1359, #1360). `uv run pytest` (1216 passed), the worker's vitest suite and typecheck, and every pre-commit hook pass on this commit. > _This was written by Claude Code on behalf of max-sixty_ 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
max-sixty
pushed a commit
to max-sixty/leaf
that referenced
this pull request
Sep 22, 2026
Regenerates the nine `tend-*` workflows on tend 0.2.15, up from 0.2.14. Every job now pins `max-sixty/tend/claude@0.2.15`, so merging adopts that release's agent behaviour as well as the YAML below. ## What changes for this repo **`tend check` holds branch protection to a ruleset the bot cannot bypass.** Required reviews alone no longer pass the `branch-protection:main` check: the bot holds write, so its own approval satisfies a review requirement on a pull request someone else opened, and it can then merge. A read it cannot decode is now a failed read rather than a pass — previously an unreadable protection response reported the check green without having verified anything. Leaf already carries the `Merge access` restrict-updates ruleset, so this check stays `PASS`; `tend check --fix` now offers to create that ruleset for any failing branch, not only for the specific "bot can still merge" message. **`tend-mention` measures its queue delay before checkout and `setup:`.** The delay step moved ahead of both, so the number it reports is the wait for the job to start rather than that wait plus however long this repo's `setup:` takes. Leaf's mention runs will report a smaller delay than before for the same actual queue time. **Reviews open with a finding.** A review body no longer spends its first lines restating where GitHub shows the verdict, a follow-up review focuses on what was pushed while keeping the whole PR in view, and each further fix round on a PR with no human author has to clear a higher bar before the bot pushes again. **Action pins.** `astral-sh/setup-uv` moves from v10.1.0 to v10.2.0 in the two workflows that set it up directly (`tend-mention`, `tend-notifications`). [Compare 0.2.14...0.2.15](max-sixty/tend@0.2.14...0.2.15) <details><summary>Upstream commits in this range</summary> - `chore: regenerate workflows with tend 0.2.14` (max-sixty/tend#1358) - `fix(review): say where GitHub shows a review's verdict, so the body opens with a finding` (max-sixty/tend#1362) - `fix(mention): compute the queue delay before checkout and setup` (max-sixty/tend#1365) - `fix(review): raise the bar for each further fix round on a PR with no human author` (max-sixty/tend#1366) - `Require a restrict-updates ruleset, pin AGENTS.override.md, and rewrite the security docs` (max-sixty/tend#1367) - `chore: bump setup-uv to v10.2.0` (max-sixty/tend#1359) - `chore: bump repo-local setup-uv and wrangler pins` (max-sixty/tend#1360) - `chore(token-usage): drop the stream-json artifact copy its own comment marked for removal` (max-sixty/tend#1361) - `fix(review): focus a follow-up review on what was pushed, with the whole PR in view` (max-sixty/tend#1368) - `chore: release 0.2.15` (max-sixty/tend#1369) The `AGENTS.override.md` pin in max-sixty/tend#1367 and the dropped stream-json artifact copy in max-sixty/tend#1361 have no effect here — the first applies to the Codex harness and the second to tend's own repository. </details> Co-authored-by: leaf-agent <318509791+leaf-agent@users.noreply.github.com>
max-sixty
pushed a commit
to max-sixty/worktrunk
that referenced
this pull request
Sep 22, 2026
Regenerates the nine `tend-*.yaml` workflows with tend 0.2.15 (from 0.2.14). The version bump and the `astral-sh/setup-uv` pin (v10.1.0 → v10.2.0, in the `tend-mention` and `tend-notifications` preflight jobs) are mechanical; one step actually moves, in `tend-mention.yaml`. **`tend-mention` measures its queue delay before checkout and setup.** The "Compute queue delay" step used to run after `actions/checkout` and `./.github/actions/tend-setup`, so the number it recorded was the wait for the job to start *plus* the several minutes this repo's setup action takes to install the Rust toolchain and sandbox tools. It now runs first, and the delay reports the queue wait alone ([max-sixty/tend#1365](max-sixty/tend#1365)). Behavior that arrives inside the action rather than in these files, and that shows up in what the bot posts here: - **PR reviews lead with a finding.** A review body now opens with something actionable instead of restating the verdict GitHub already renders in its own header ([max-sixty/tend#1362](max-sixty/tend#1362)), and a follow-up review scopes itself to what the latest push changed while still reading the whole PR ([max-sixty/tend#1368](max-sixty/tend#1368)). - **Fewer self-directed fix rounds on bot PRs.** On a PR with no human author the review session applies its own findings; each further round now has to clear a higher bar, so a bot PR stops accumulating marginal follow-up commits ([max-sixty/tend#1366](max-sixty/tend#1366)). - **`AGENTS.override.md` joins the pinned instruction paths.** Codex reads it in preference to `AGENTS.md`, so on a PR-derived checkout it is now restored from the base commit alongside `CLAUDE.md`, `AGENTS.md`, `.claude/`, and `.agents/` — a fork can no longer reach the session through that filename. The same release makes a restrict-updates ruleset a `tend check` requirement; this repo already satisfies it (`tend check` passes every line today) ([max-sixty/tend#1367](max-sixty/tend#1367)). Full comparison: [0.2.14...0.2.15](max-sixty/tend@0.2.14...0.2.15). Co-authored-by: worktrunk-bot <254187624+worktrunk-bot@users.noreply.github.com>
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
2 participants
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.
Self-review COMMENTs open with "Reviewing my own PR, so this is a COMMENT rather than a verdict." (PRQL/prql#6352 has it in three rounds). No instruction contains or asks for the line. The first round produced it by chance; in 12 replays of that session it came up in none. Later rounds copy it, because each reads the earlier bot reviews in full for dedup and follows their form.
The Submit step now says GitHub labels the review approved or commented above its body, so the body opens with the first finding. It applies to every review: naming the self-authored case in the prompt raised the rate in a synthetic A/B. Replaying that PR's three review sessions, cut just before each wrote its body:
Nothing tried stops round 3, including an explicit prohibition: once two posted reviews carry a line, the session copies it.
The second commit backticks the event names (
COMMENT,APPROVE) in skill prose, to match the review states (APPROVED,CHANGES_REQUESTED) already backticked beside them. On its own it doesn't change the opener (round 2 stayed at 4/4).🤖 Generated with Claude Code