feat(inbox): a settled review comes back when someone asks for you in words - #45
Merged
Merged
Conversation
… words The reopen rule listened for GitHub's re-request button and nothing else, so a settled row stayed invisible while its author waited — having said so on the PR in plain words. "@you this is ready for another look" is the same ask, made the way people actually make it, and GitHub emits no event for it at all. The poll already reads the conversation for whose-move; it now also reads it for the last comment naming you, and reopens under the same after-settling test the button uses. Your own comments never count, bots never do, `@org/team` never does, and a longer login that starts with yours is not you. Two places, because a fix that only worked while a request was open would half-work: rows GitHub is still asking about read the button first and the words only if the button did not already answer, and rows it has stopped asking about — you reviewed on github.com, which clears the request — read the words alone, off the state-refresh leash that was already spending a call on them. An unsettled row still costs no GitHub call. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
There was a problem hiding this comment.
Pull request overview
This PR updates cerber’s daemon reopen logic so a review you previously settled (reviewed/skipped) can return to the inbox not only when GitHub’s re-request button is used, but also when someone explicitly @-mentions you in a PR comment asking for another look. It extends the existing polling behavior while preserving the “pushes don’t reopen settled work” principle and documents the normative rules in the spec and lifecycle docs.
Changes:
- Add mention-based “asked again” detection by fetching PR conversation comment bodies, matching
@loginrobustly, and selecting the newest qualifying mention by timestamp. - Refactor daemon reopen logic to model “ask” as a discriminated union (
requestedvsmentioned) and centralize reopen wording for logs/history. - Add unit tests for both reopen paths (awaiting-search path and out-of-search state-refresh path), including the precedence short-circuit and non-qualifying mentions (bots/self/team/longer-prefix logins), plus update SPEC/docs accordingly.
Reviewed changes
Copilot reviewed 7 out of 7 changed files in this pull request and generated no comments.
Show a summary per file
| File | Description |
|---|---|
| src/server/daemon.ts | Implements the two-path reopen logic (re-request event vs @-mention), adds Ask union + centralized wording, and integrates a state-refresh reopen check. |
| src/server/daemon.test.ts | Adds coverage for typed @-mention reopening and ensures the button short-circuit and exclusion rules behave as intended. |
| src/core/gh.ts | Extends conversation fetching to include comment bodies and adds mentionsYou/lastMentionOfYou helpers. |
| src/core/gh.test.ts | Adds focused tests for mentionsYou and lastMentionOfYou, and updates existing fixtures for the new body field. |
| SPEC.md | Updates the normative spec sections to describe both ask mechanisms, their costs/leashes, and wording consistency requirements. |
| docs/lifecycle.md | Updates lifecycle documentation to reflect both reopen mechanisms and where they run. |
| CLAUDE.md | Updates architecture notes to reflect the new reopen behavior and evidence sources. |
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
|
🎉 This PR is included in version 0.27.0 🎉 The release is available on: Your semantic-release bot 📦🚀 |
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.
The problem
A review you settled — marked reviewed or skipped — stays out of the inbox no
matter how many times the author pushes. That is deliberate: a row dragged back
every time someone commits is cerber overruling you. The one way back in was
somebody asking for your review again, and cerber only recognised one form of
asking: GitHub's re-request button.
People frequently don't press it. They write "@you this is ready for another
look" instead — the same ask, made the way people actually make it, and one
GitHub emits no timeline event for at all. Cerber read straight past it, so the
row sat in settled, invisible, while its author waited on a review they had
already asked for in plain words on the PR.
The fix
The poll already reads each PR's conversation, for whose-move-is-it. It now
reads the same comments for the last one that names you, and reopens the row
under the same after-settling test the button already used: an ask newer than
settledAtundoes the settle, anything older is the ask you already answered.What doesn't count, and why:
push would otherwise be an ask on every push.
@org/team. A room being addressed is not somebody asking for you —the mirror of the rule the button check already applies to team requests,
and for the same reason: this undoes a decision of yours, so the safe side is
doing nothing unless somebody named you.
@memust not match@me-bot,and an email address must not match at all.
Two places, because one would half-work
the button first, and the conversation only if the button didn't already
answer. A live re-request says everything a comment could, so the second call
is spared.
not an edge: reviewing on github.com clears the request and drops the PR out
of the awaiting search — which is precisely the state a "ready for another
look" comment tends to arrive in. That check hangs off the state-refresh
leash that was already spending a call on those rows.
An unsettled row still costs no GitHub call on either path, a failed read still
changes nothing, and the leash stamp is still spent before the call so an
outage can't turn every settled row into a retry storm.
Technical details
src/core/gh.ts— the conversation read now carries each comment's body.mentionsYou()matches@logincase-insensitively with a not-a-login-charboundary;
lastMentionOfYou()picks the newest qualifying comment by daterather than by position, since this undoes a settle and shouldn't lean on the
API returning things in order.
src/server/daemon.ts— the reopen now takes anAsk, a discriminated unionover the two ways of asking.
reopenIfAskedAgain(awaiting search) andreopenIfAskedInWords(state refresh) both funnel into onereopen(), whichkeeps the existing atomic re-check against the on-disk artifact so a
re-review or send started mid-read isn't overwritten.
askWording()is oneplace, so the terminal line and the row's history note can't tell different
stories about why a decision of yours was undone.
try. A write that fails is a realfault and is no longer swallowed as "the read didn't work out".
Verification
pnpm typecheck,pnpm test(516 passing, 18 new) andpnpm buildall greenlocally.
Beyond the unit tests, the new code was run against the real PR conversation
that surfaced this — the row as it stood would have come back into the inbox on
the next poll, dated by the comment that asked:
New tests cover both paths (in and out of the awaiting search), the
precedence short-circuit, mentions older than the settle, and each thing that
must not count as an ask.
Docs
SPEC.md§9.6 is rewritten for both ways of asking, with normative rules forthe mention and for where each check runs; §8.2's transition diagram, §21.4's
pseudocode and the conformance checklist follow.
docs/lifecycle.md's "Askedagain" section and
CLAUDE.md's daemon paragraph likewise.Worth knowing
A
sentrow is still never reopened by this — sent rows track the headinstead, so a push is what brings them back. That's unchanged and deliberate.
🤖 Generated with Claude Code