* feat(ci): weekly Codex sweep for date-carrying to-dos
Mondays 06:23 UTC (plus default-branch-only manual dispatch with
dry_run): a deterministic prefilter collects candidate lines where
to-do markers co-occur with date tokens, plus every deprecated compat
record from the plugin registry; a pinned openai/codex-action step
judges each candidate in context (genuine dated commitment vs
historical date or fixture) and writes an OVERDUE / DUE-30-days /
FUTURE report; a separate privileged job validates the report (tracked
file:line locations, date shape, inert text) and upserts one
marker-tagged tracking issue via the Barnacle app token, commenting
only when items newly become due.
The app token is minted on a fresh runner and checkout — never beside
Codex or its child processes; only the validated report artifact
crosses jobs. No permission-* subsets on token minting (installations
reject explicit subsets; see pr-ci-sweeper).
* chore(ci): localize upsert helpers with no external consumers
install_node now calls validate_downloaded_script after the NodeSource
download step. The Debian pacman-ignore harness did not mock those
helpers, so empty mktemp files failed closed and status was 1.
Signed-off-by: Sebastien Tardif <sebtardif@ncf.ca>
The rebase onto main introduced validate_downloaded_script between
download_file and run_required_step. Since the test mocks
run_quiet_step (which swallows the actual download_file call), the
temp file stays empty and validate_downloaded_script fails early,
preventing the test from reaching the intended run_required_step
failure path.
* feat(models): add Claude Opus 5 support
* test(models): align Opus 5 setup fixtures
* test(models): update ambient Opus 5 detection
* fix(models): reconcile Opus 5 support with main
Run-identity supersession (#113230) resolves the scenario this heuristic
softened: superseded runs and rerun-attempt leftovers are filtered before
classification, so aggregate-FAILURE-with-green-duplicates now classifies
as GREEN or PENDING on evidence. The only remaining reachable case —
identity-less app check runs posting duplicate same-name results — is
genuinely ambiguous and now honestly exits FAILING (15) instead of the
special verify-manually exit 17.
* fix(scripts): ignore superseded workflow runs in watch-pr-ci rollup classification
Draft->ready re-triggers leave cancelled superseded runs on the head SHA
forever, and GitHub's aggregate rollup state counts them, so the watcher
emitted terminal FAILURE while the replacement run was still in progress
and could never reach GREEN (observed on PR #113150, head bd1b9a0e).
- fetch run identity per check and resolve same-name checks to the newest
run/check id (GitHub latest-name-wins); drop cancelled checks from
replaced runs; keep older runs' unique jobs visible
- paginate statusCheckRollup contexts (bounded, 10 pages) so >100-context
rollups are not truncation-blind; changed or lost snapshots throw into
the bounded retry
- classify GREEN when aggregate FAILURE is explained solely by superseded
contexts with full visibility; truncation still fails conservatively
* ci: scope GitHub App token permissions in stale, labeler, auto-response, and pr-ci-sweeper
Fixes red main: #112963 bumped zizmor v1.22.0 -> v1.28.0, whose github-app
audit flags create-github-app-token mints without permission-* inputs (14
high findings, Workflow Sanity red on main since 3b7b2a2a1f). Most
workflows already migrated to scoped tokens; these four were stragglers.
Scopes follow each consumer's actual API surface: stale needs issues/PR
write plus actions read for its state-cache check; labeler needs label
CRUD (issues write), PR write, members read for maintainer gates, and
contents read where actions/labeler reads its config; auto-response needs
issues/PR write plus members read; pr-ci-sweeper needs actions write to
re-fire runs, checks read, and PR write. Verified locally with the exact
CI invocation (zizmor 1.28.0, repo config, regular persona, medium
severity/confidence): no findings, ignore/suppress counts match CI.
