Repository navigation
feat(brief): lead with how the PR was decided (0.24.0) - #60
Merged
Merged
Conversation
- The decisions section becomes "How this PR was decided": a table of what happened, the decision and the commit as evidence, first in the brief. Rows are fixed labels, outcomes and the change's own paths; no transcript text. - New kinds: a question the agent asked (which option was picked, by position only) and an independent review (reviewer type and a verdict in fixed words, from the call's result, a later task notification, or the reviewer's last message in Entire's stored subagent transcript). - A checkout and its worktrees count as one repo, so work carried from one worktree to another keeps its decisions; leaves/returns read a variable the command assigns. - A test run against code swapped in before it (git show ref:path >, stash, apply -R) or put back after it in a loop is a probe. Replay on 100 checkpointed commits: briefs with a decision 21 -> 24, every changed line checked by hand. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PR brief: 3 agent commits · 8 of 25 test files reach this diffRun at least these. Selection is a safe superset; it never marks a test safe to skip.
What each commit did
Claims checked
What each changed file reaches
Paths above are under Tests to run (8)
Blast radius by changed file: 22 dependents
blastline 0.24.0. Selection is advisory unless your workflow gates on it. How selection works. Graph |
Review of #60: - A sibling worktree counts only where the change was carried from it as a patch (git diff > file there, git apply file here), and only from after its own last commit; the folder-name family rule leaked a parallel PR's tests into this brief and merged ~/.worktrees repos. - The swap-in probe fires only when git show writes over the same path, on a bare git stash, or git apply -R; saving a copy elsewhere, stash pop and 2> no longer drop real runs. - verdictOf reads a Verdict line, a line opening with a fixed word, or a check lane's result line; blocking words win, prose "not APPROVED" is nothing; READY TO MERGE and BLOCKED are mapped. - Reviewer match needs a word (not "critical"/"preview"); multi-select answers report each position; alike labels claim none; questions and reviews dedupe by tool call; table cells escape "|". - Tests for each, plus the stored-subagent verdict path. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… the brief The spliced notification is written across lines like a real one, so its verdict word opens a line; "not APPROVED" on a Verdict line is no verdict. Stored-transcript tests for an approving and a blocking review. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
taylorg009
marked this pull request as ready for review
October 22, 2026 01:06
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.
Why
Reviewers of agent-written PRs want to know why the change looks the way it does. The brief's decisions section rarely said, for two reasons:
What changes
AskUserQuestioncall shows the option picked, by position only (e.g. "the recommended option 1 of 3", or "answered in their own words"). The question and labels are never quoted.tasks/transcript; otherwise "verdict not recorded".git diff > <file>there,git apply <file>where the commit was made), and only from after its own last commit. A parallel PR in another worktree never leaks in. Variables can build on earlier ones (NEW="$REPO/x"), and the "left the repo" and "back at the root" checks read them.git show <ref>:<path> > <path>, a baregit stash,git apply -R), or when the fix is copied back after it, including from a loop'sdo cp. Saving a copy elsewhere,git stash popand2>are ordinary steps.No transcript text is quoted. Every cell is a fixed label, an outcome, a runner name, one of the change's own paths, or a commit link.
Evidence
The new lead section, rendered from a real replayed checkpoint: a failing-then-passing test and three check-lane reviews, whose verdicts were read from the stored subagent transcripts.
The real web-app PR: its fix was made in one worktree and carried to another as a patch. Its decision window now covers that hand-off (7 lines instead of 3), and the brief shows
vitestfailed → fixed, then it passed. That's the view-mode contract test that made the agent defer a fetch; its commit message confirms it.Replay on 100 checkpointed web-app commits (
checkpointFor, each commit's own files): briefs with a decision went from 21 to 22, plus review rows on three more, and every new or changed line was checked by hand against its transcript.Tests: 446 pass. Each review finding has a case that fails on the previous commit:
git show … > /tmp,stash popand2>still count;The stored-subagent verdict path is now tested too.
An independent review requested changes (three blocking false-row bugs; privacy held). The second commit fixes all of them.
🤖 Generated with Claude Code