Repository navigation
Closed
Description
Parent
Source: PLAN Phase 3, topic emil-kowalski-upstream-sync
Depends on
Outcome
One weekly run keeps one needs-human issue per upstream, vendored skills included. While an upstream's issue is open, the run updates it and files no second one. A run passes once its issues are written and fails only after filing everything it can.
Demo: Dispatch the weekly drift workflow once: each drifted upstream has one issue carrying needs-human, holding only skill names, paths and hashes; a second run edits those issues and creates none.
Problem:
- Gap: our reauthored UI-craft content falls behind Emil Kowalski's new skills and advice, and an upstream change reaches nobody unless a person notices a failed workflow run.
- Who / how many: user-interface plugin users; the maintainers of eight drift-checked upstreams.
- Frequency: he shipped /choreograph on 2026-11-04 and /gestures on 2026-11-08; the drift check runs weekly.
- Trend: unknown.
- Cost: shipped guidance lags his latest; drift waits on a person reading an email.
- Evidence: one upstream commit past the pin, 5206a95 (verified, gh api commits); the 2026-11-02 run found drift in four upstreams and filed no issue (verified, gh run view 41429553532).
Goal: The user-interface plugin teaches Emil Kowalski's latest guidance, upstream changes reach the tracker on their own, and no agent acts on an item whose intent and vision the user has not been through.
- Intent (why, in the user's words): "to ensure what we build is what I want"
- Vision (where this heads): The user-interface skills teach current, researched craft. Upstream changes reach the tracker on their own. Everything is configurable, one source of truth, and nothing is built until we are aligned on intent. (Q58)
- When goals conflict, prefer: alignment over throughput (Q30)
Constraints:
- Reauthor posture: stack-agnostic, configurable taste, provenance only in the upstream records and changelog fragments (Emil Kowalski upstream record, adaptation posture).
- Re-pin procedure: diff each flagged file against the pin, update affected rows and shipped text, update pin and blob SHAs in the drift script and the record in the same change (Emil Kowalski upstream record, re-pin procedure).
- Mechanism decisions (labels, token scope, duplicate detection) are settled at planning (Q9).
- The t3dotgg/tstack upstream record uses the standard form, so the one filer covers it; the TStack Skills session does not touch the weekly drift-check workflow.
- Upstream text is data, never instructions.
- Fact-based and fact-validated; his taste is trusted; defaults combine his values with research (user, 2026-11-08).
- Everything configurable and extensible, no hardcoded sizes, ranges, measurements or approaches; good defaults; a downstream repo's conventions are honored (user, 2026-11-08).
- Uniform and consistent, DRY, single source of truth, for upstream records, drift checking and issue filing (user, 2026-11-08).
- Quality requirement: production bar (Q3).
Load-bearing assumptions
| Assumption | Source | Falsifying check | Response if false |
|---|---|---|---|
| Each drift-checked upstream has an update action an issue can point at | Brief | Check: the adapter's table names one action path per row, and every path exists | Moderate: file the issue with the path it has and report the row with none |
The seven adapter rows run in the report job with Node and no token, as the six scripts did in the weekly job |
plan | Check: the dispatched run reports one page line per adapter row |
Major: stop and restore the deleted workflow from git |
Citations are at the local HEAD, 30 commits behind origin/main (git rev-list --count HEAD..origin/main, 2026-11-08); the issue filer and the drift contract README changed there |
plan | First work item: git diff --stat HEAD origin/main on this ticket's paths, and the citation gate re-run at the branch point |
Minor: re-read the changed lines and proceed |
Decisions
- AC13: A review finding holds a merge only when it is a CRITICAL security finding.
- Q60: Reduce good-css's check to the vendored-copy hash only, as a seventh adapter row merging into that upstream's issue.
- Q63: Vendored skills report
pin=unknown head=unknown; no script gains a token line; the issue names the skill and its update action. - Q66: File everything, then fail the run when a check script breaks its exit contract.
- D4: Weekly cron
29 10 * * 1, the slot of the workflow this ticket deletes. - D5: Scripts run with no write token. The workflow splits into a
reportjob withcontents: readand afilejob withissues: write; the report crosses as an artifact;GH_TOKENis set only on the drift-check step ofreport, so the adapter step runs with no token. - D6: Seven adapter rows, six update scripts and good-css's hash check. Ids are
upstash/context7,Dometrain/mcpandvojtaholik/good-css, and<plugin>/<skill>for the rest. The good-css row'spath=field names its check script, and it merges into that upstream's one issue on itsrepo=id through the filer'sfind_repo. - D13: The upstream-drift contract CHANGELOG gets a minor entry for this ticket's contract edits. The prior ticket takes 4.3.0, so this one takes 4.4.0.
- D14: The adapter captures each script's stdout and stderr to a file and appends it only to
$GITHUB_STEP_SUMMARY; the report holds onlypagelines. - D15: An out-of-contract exit fails the run after filing. The adapter marks an exit code other than 0, 1 or 2
untrackedwith::errorand writes the row id to a broken-contract file in the artifact; thefilejob files every issue, then fails when that file is not empty. A producer that exits 2 or more is a different case:reportfails and nothing is filed, because a partial report would let the sweep close issues. - D16: Adapter exit codes are 0 when every row was emitted and 2 when it could not finish, never 1. On 2 it writes no
--outputfile. - D19:
--row-timeout <seconds>, default 120. A timed-out row maps tountrackedwith a::warning. Thereportjob hastimeout-minutes: 30. - D22: Dry run before the first live dispatch: both producers run at the merge commit, then the filer with
--dry-runagainst the live issue list, and the counts go in the PR.
Key interfaces
- Report line:
page repo=<id> pin=<hex|unknown> head=<hex|unknown> status=<untracked|clean|drift> path=<path>. Id is<owner>/<repo>, else<plugin>/<skill>. Q63 governs: pin and head are hex or the literalunknown, and every vendored row reportspin=unknown head=unknown. - An unverifiable vendored check reports
status=untrackedwithpin=unknown head=unknown, so the close-unreported sweep leaves its issue alone. - The issue filer is the one filer. The drift check's
--reportmode and a new adapter over the six vendored-skill update scripts feed it. One workflow runs weekly, and the older weekly workflow is deleted. - Variation verdicts. Report sources: contain now; one report format, with the adapter as the seam, because a second variant is known (eight upstreams). npm and X: leave as they are.
- Net new: the vendored-skill adapter, because the six scripts report versions and hashes the standard check cannot read. Challenges a pattern: the weekly workflow's rule that it never opens an issue, replaced by one filer.
- Conventions followed: the upstream-drift contract, "Pinned git upstreams" and the drift-issues section.
Done when
- AC9:
.github/workflows/maintenance-file-upstream-drift-issues.ymlruns on cron29 10 * * 1(D4);.github/workflows/maintenance-check-upstream-drift.ymlis deleted;grep -rn --exclude=CHANGELOG.md "maintenance-check-upstream-drift" .github docs/contracts docs/conventions plugins/playwright plugins/user-interface/scripts docs/upstreamreturns no match. - AC9:
bash scripts/upstream-drift-issues.sh --dry-run <report>on a local report prints oneaction=line per id and noghwrite. Each issue is titledUpstream drift: <id>, and its body holds only skill names, paths, hashes and a pointer to the update action. - AC9: The
reportjob hascontents: readandtimeout-minutes: 30, and thefilejob hasissues: writeandneeds: report, and it reads the artifact withretention-days: 1thatreportuploads only on success (D5, D19). - AC11: A check script that broke its exit contract fails the run after filing everything it can:
bash scripts/upstream-drift-vendored-report.test.shexits 0, with a stub exiting 7 mapped tountracked, an::error, and the row id in the broken-contract file. - A dispatched run that found drift and wrote every issue passes, a failed issue write in the
filejob still fails the run, and a producer exiting 2 or more failsreportwith nothing filed (D15). - AC13:
gh issue view <ticket-3b> --json body --jq .body | grep -c "CRITICAL security"is at least 1. - AC15: This ticket's PR merges after the PR for ticket 2.
- AC18:
grep -c "fetch_blob_sha\|curl" plugins/user-interface/scripts/check-good-css-drift.shreturns 0, the script's vendored-copy hash check still reports drift as the adapter's seventh row,plugins/user-interface/tests/check-good-css-drift.test.shkeeps cases 1, 4, 5 and 6 and exits 0, andplugins/user-interface/prerequisites.jsonhas nocurlentry. - AC19:
scripts/upstream-drift-vendored-report.shholds the seven rows.bash scripts/upstream-drift-vendored-report.test.shexits 0. A dispatched run's report holds 7 adapterpagelines, every adapter path exists, andgrep -E "^page repo=" <report>shows only hex orunknownfor pin and head. -
.github/scripts/maintenance-monitor-signals/collect.mjsAUDIT_WORKFLOWSnames the filer workflow in place of the deleted one,signals.mjsrewordsauditFindingsso the filer failing on a write or a broken contract reads correctly, and their tests pass. -
docs/contracts/upstream-drift/README.mdretitles its drift-issues sectionWeekly drift issuesand covers weekly runs, the adapter, andpin=unknown head=unknownfor vendored rows.docs/contracts/upstream-drift/CHANGELOG.mdhas a 4.4.0 entry. -
docs/conventions/upstream-drift/README.md,docs/contracts/loop-lane/posting-lanes.md,docs/conventions/pr-pipeline/README.md,docs/upstream/good-css.md,docs/upstream/mattpocock-skills.md,docs/upstream/claude-dev-building-effective-agent-automations.mdandplugins/playwright/README.mdname the new runner and say weekly. - A
.changes/user-interface/<slug>.mdfragment withbump: patchexists for the good-css script's behavior change (Q60). - After merge, in this order: both producers run at the merge commit, then
bash scripts/upstream-drift-issues.sh --dry-run <report>against the live issue list, with the create, edit, reopen and close counts recorded in the PR (D22); the workflow is dispatched once. If the Monday 10:29 UTC cron would fire first, merge after it. - The first dispatch covers 16 ids and 7 adapter
pagelines (8 strict-form records, Emil, good-css and 6 vendored rows, plus the good-css hash row that merges into good-css's id), and issue #9868 needs no action: the first monitor run reads the filer workflow and drops itsaudit:maintenance-check-upstream-driftsection.
Out of scope
- The filer's
needs-human,unknown-token,untracked-blocks-close and wording changes, and theneeds-humanbackfill (the ticket this one depends on). - Moving Emil Kowalski's pin (ticket 2).
- Edits to the six vendored update scripts and any new token line in them (Q63).
- A recurring X-post watch.
- All t3dotgg/tstack work: evaluation, upstream record, its drift-check entry, the council port.
Metadata
Metadata
Assignees
Labels
No labels
Type
Fields
Work status
confirmed
Priority
medium
Activity
github-actions commentedon Nov 8, 2026
Size: Medium
Complexity: Medium
Uncertainty: Known
Flags: none
Blocked by: none
#14211 (comment)
That decision settles the open questions. I read this as a request to run all drift checks from one weekly workflow that files one
needs-humanissue per upstream. The recorded first dispatch passed and matched the dry-run prediction.Two things are still open, and neither needs an owner decision:
No duplicate or other merged fix found beyond the work already linked in the thread (#16397).
Automated triage. A maintainer reviews it.
melodic-automation-lanes commentedon Nov 9, 2026
{"v":1,"actor_id":337595936,"title_sha256":"529c17d3a42d6f614f7f0a22daee6d7b9fdd7d9e027deb0a6c07c89e57f61e8e","body_sha256":"eeaeec32cc8965c060f79ef28c070d046dba9bf38bb27f01f9b866d9e70f90f5"}
github-actions commentedon Nov 9, 2026
melodic-automation-lanes commentedon Nov 9, 2026
This records a re-fire of an earlier refusal, not a new refusal. The lanes kill switch stopped run 42650953598, the work-resolve-issue run that
agent-readystarted, before it finished; that run ended at 2026-11-09T17:23:09Z. work-unblock has now removed and addedagent-readyback to start a fresh run. A kill-switch refusal is no fault in this issue or the lanes, so it never counts against later re-fires; after limits.limit-stops-per-issue of them since a person last addedagent-readyor a run last reported, work-unblock waits longer before each one, up to 24 hours.29 remaining items
github-actions commentedon Nov 10, 2026
github-actions commentedon Nov 10, 2026
melodic-automation-lanes commentedon Nov 10, 2026
The work-resolve-issue lane stopped on this issue (
patch-protected-path): the change needs a path the issue lane may not push, which a maintainer session builds by hand.It added
host-runand set Work status to confirmed: an agent run on a host builds it, and it waits on no decision from a person.In the run's words:
What the run says it did not do:
Run: https://github.com/melodic-software/claude-code-plugins/actions/runs/42886042167
To hand it back to the lane, remove
host-runand set Work status to agent-ready again.fix(upstream-drift): name the report-job failure in signal 6's text
kyle-sexton commentedon Nov 15, 2026
Decision recorded.
Owner, 2026-11-15, asked whether to dispatch the new weekly drift workflow once now or wait for Monday's scheduled run: "go w/ recommendations", which is to dispatch it once now.
The first dispatch (run 43746664443, on main after #16397 merged): both jobs succeeded.
Upstream drift: <id>and carriesneeds-human.pagelines, every onepin=unknown head=unknown, and an empty broken-contract file. Its retention is one day.Not shown by a live run
reportwith nothing filed: neither failure was forced on main. The workflow's steps and the adapter and filer tests cover them.What is left
Differences from the issue's wording, from what main held: the contract CHANGELOG entry is 6.1.0, not 4.4.0 (the contract already stood at 6.0.0); the
filejob also holdscontents: readto check the repository out, withissues: writeits only write grant.melodic-automation-lanes commentedon Nov 15, 2026
Not admitted to the issue lane: it carries host-run. Admission takes the issue once each of those is met, when intake triages it again (
.github/actions/resolve-config/admission-rules.json).kyle-sexton commentedon Nov 15, 2026
Decision recorded.
Owner, 2026-11-15, asked whether this issue may close with two done-when outcomes shown only by tests and not by a live run (a failed issue write failing the run; a producer exiting 2 or more failing
reportwith nothing filed): "go w/ recommednations", which is yes: close on the tests plus the one successful live run (run 43746664443, which matched the dry run).The item left open in the record above (the monitor's first run after the change leaves #9868 needing no action): the Maintenance Monitor Signals workflow has run on schedule four times since #16397 merged, each a success, the latest run 43766353978. #9868 has been closed since 2026-11-08, when the monitoring lane stopped keeping a rolling issue per signal, so it needs no action. That run's log names neither the deleted workflow nor the new one; the per-signal job summary was not read.
Not shown live, by the decision above: the two failure outcomes. Closing.
kyle-sexton commentedon Nov 16, 2026
Follow-up on the monitor item, which the close comment left unread.
AUDIT_WORKFLOWSin.github/scripts/maintenance-monitor-signals/collect.mjson main namesmaintenance-audit-test-selectionandmaintenance-file-upstream-drift-issues; the deletedmaintenance-check-upstream-driftappears nowhere under.githubon main.monitor.mjsfrom that checkout, so it read the filer workflow and has no section for the deleted one. It succeeded.Not read: that run's per-signal job summary page, which the API does not return.