Pause implement, issue follow-up coding, and the Codex review loop
so Cursor only classifies new issues. Restore with
CURSOR_AUTOMATION_MODE=full and the commented schedule crons.
Ignore success-case TAP YAML diagnostics and normalize volatile console
noise in baseline comparison so red-but-unchanged candidate runs pass as
baseline_only. Align the session-backed fastPut expectation with live-path
uploads, harden SCP abort teardown races, drop mixed hallucinated research
citations instead of failing classify, and soften verification failure
copy so it no longer claims the patch introduced the failure.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ci): dedupe same-head @codex review requests
Plain maintainer/Cursor-bot @codex review comments did not plant the
cursor-external-codex marker, so own_rerequest on synchronize posted a
second request for the same SHA. That concurrent pair produced a clean
issue comment from one job and a findings review from the other.
Skip re-request when any trusted or Cursor-bot author already requested
this head via external marker, cursor-codex-head pin, or a recent plain
@codex review.
* fix(ci): only dedupe plain @codex against the current head push
Plain unpinned requests no longer skip solely by age — that blocked
re-request after a new synchronize when an older plain @codex still
looked "recent". Callers pass the head commit time as notBefore so
only post-push plain requests suppress a duplicate.
* fix(ci): use PR head-change time for plain @codex dedupe
Commit author/committer dates predate cherry-picks and rebases, so a
plain @codex for the previous head could suppress re-request after a
new synchronize. Pass pull_request.updated_at (moves with the push)
as notBefore instead.
* fix(ci): dedupe @codex only via head-pinned markers
Drop plain-unpinned timestamp heuristics — they cannot safely know which
SHA a bare @codex review targeted. Skip re-request when a trusted author
already pinned this head with cursor-codex-head or cursor-external-codex.
* fix(ci): let maintainer @bot reclassify auto-closed issues
Author reopen disputes still stay on issue_followup when
triage:already-available/unclear remain, but an explicit maintainer
@netcatty-bot mention must re-enter issue_classify so re-triage can
run without an open bot PR.
* fix(ci): prefer maintainer @bot over author follow-up routing
When a trusted maintainer is also the issue author, route explicit
@netcatty-bot mentions as maintainer re-triage so auto-closed
already-available/unclear issues can re-enter classify.
* fix(ci): stop reopen from replaying issue triage
Already-admitted issue reopen events were routed back into full classify,
so follow-up handoffs and maintainer reopens could close again or overwrite
ready-for-human with needs-info. Skip bot reopen, hand auto-closed human
reopens to maintainers, and keep already-available disputes on follow-up.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ci): keep ready-for-human follow-ups off classify
After reopen handoff drops triage:already-available, actionable author
replies could still be refined into classify and auto-close again. Treat
ready-for-human like the dispute signal and keep those comments on
follow-up.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ci): preserve already-available dispute label on handoff
Keep triage:already-available when handing reopened auto-closed issues
to humans, and only use that label (not ready-for-human) to block
reclassify. This stops the close loop without blocking feature_defer or
other ready-for-human follow-ups from re-entering classify.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ci): keep unclear reopen disputes off classify
Match the already-available dispute guard for triage:unclear / unclear so
human reopen handoffs of auto-closed unclear issues cannot re-enter
classify and close again on the next author follow-up.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ci): revalidate auto-close labels before handoff
Queued ready_for_human_handoff must not overwrite a maintainer who already
cleared triage:already-available/unclear and moved the issue (for example
to ready-for-agent). Skip the handoff when those outcome labels are gone.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* test(ci): cover skipped reopen handoff without auto-close labels
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ci): do not reopen issues already closed before handoff
Queued ready_for_human_handoff must respect a maintainer who re-closed the
issue while the job was waiting. Skip when live state is closed instead of
forcing ensureOpen.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix(ci): reclassify non-auto-close issue reopens
After a merged automation fix, cleanup leaves triage:bug-ready/admitted on
the source issue. Human reopen of those labels must re-enter classify so a
failed fix can spawn another implementation; only bot reopen stays skipped.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Only automation bot PRs re-check source-issue pending comments before
mark-ready, so maintainer Fixes-only PRs are not stuck draft. Write
WebSearch/WebFetch denials after the isolated research pass so classify
and follow-up research can still call those tools.
- Format recovery dispatches triage before clearing invalid-format
- Shared CJK-friendly title rules; codex label terminal state machine
- Research steps in classify; agent TITLE + PR body for bot PRs
- Single @codex review via CODEX_REQUEST_GITHUB_TOKEN (binaricat)
Public issue comments no longer show a generated-by line. Grounding stays
in code_findings/reasoning only so reply can use UI language without file
paths, symbols, or dense parentheses. Classify prompt and rewrite templates
match that tone.
* fix(ci): stop Cursor issue triage crashing on Search pagination
Octokit normalizes Search API pages so response.data is already the items
array. Mapping response.data.items returned undefined and then crashed on
candidate.user during daily-limit admission — so every valid new issue
failed before Cursor classify could run.
* fix(ci): classify already-available issues with how-to and auto-close
Add an already_available triage category for requests the product already
ships: after code-grounded inspection, reply with the concrete UI entry
path, then close as completed. Low confidence falls back to other instead
of closing. Keeps the Search pagination fix so new issues can classify.
Classification must open real source files first. The prompt mandates
search→read→classify, JSON now requires code_paths/code_findings, and
normalizeClassification rejects ungrounded replies so agents cannot answer
from issue text alone.
Isolate publish onto fresh runners, harden Codex loop head scoping,
protect fork bootstrapping, and polish classification replies and
mark-ready error handling after local codex review.
Tighten re-triage, Codex single-flight signals, protected-path checks on
commit ranges, trusted helper loading from RUNNER_TEMP, and failure
handling so own/bot loops stay safe and forks only re-@codex.
Drop Cursor CLI review for third-party/fork PRs. Initial Codex review
stays with the existing connector; after contributor synchronize we only
comment @codex review (deduped per head SHA).
Wire GitHub Actions to classify and implement issues via Cursor CLI,
open draft PRs, request @codex review for own/bot PRs, fix findings,
and only post read-only review comments on third-party PRs.