You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Canary: schedule ci.yml and publish.yml ahead of the consumers' Monday builds (PLAN P0 0c) #302
PLAN's P0 item 0c, split out of #138 when that tracker closed. The release gate (#135, #136) tests a candidate before v0 moves, but only when a releaser dispatches it. Between releases, a change outside this repository, such as a deploy service's API, a CLI or the :latest image, reaches the canary's preview and publish paths no sooner than it reaches the consumers.
What the canary, QuantEcon/test-actions-lecture-intro, runs today (checked 2026-11-10):
cache.yml runs weekly on Sunday at 03:00 UTC and by hand, a day ahead of the consumers' Monday builds. It is the only canary path with a lead.
ci.yml runs on pull_request and workflow_dispatch, with no schedule. Its preview-netlify and preview-cloudflare steps pass no dispatch-alias, so a dispatched run deploys nothing and still goes green.
publish.yml runs only on a publish* tag push. It has no workflow_dispatch, and its job-level if: accepts only tag pushes, so a scheduled or dispatched run would skip its only job and never go red.
The change PLAN describes:
schedule ci.yml and publish.yml ahead of Monday, as cache.yml is;
have ci.yml pass dispatch-alias and fail unless each preview reports deployed as 'true';
give publish.yml an if: that accepts a scheduled run.
To settle when this is picked up:
publish.yml and the gate's publish-deploy deploy the same GitHub Pages site, in the pages concurrency group. A scheduled publish replaces the gate's fixture site with the canary's, and the next gate run swaps them back.
PLAN defers this until the gate has replaced the hand-staged legs that these two workflows still carry. v0.18.0 runs the gate and the hand legs side by side once, as agreed on #135.
Carried here from #100 when it closed: detect a canary that stops running, not only one that goes red.#100 found that the canary's weekly cache.yml build ran every Sunday until 2026-04-12, then missed 16 Sundays in a row and gave no signal, because a run that never happens looks the same as one that passes. That is the pattern of GitHub disabling a schedule after 60 days without activity in the repository. Scheduling ci.yml and publish.yml here adds two more paths that could go quiet the same way. So this issue's question of how a failed scheduled run reaches a person should cover a run that doesn't happen at all, for example a check that the newest scheduled run is less than a week old. Low priority, with the rest of this issue.
Issue check (2026-11-15, against main at c21e481 and the canary's main at afd76d2) and the maintainer's decisions (2026-11-16). The change holds in substance, revised in four ways.
The deferral goes. The release gate can't replace the hand-staged legs by construction: a preview's pull-request half needs a pull_request event, release assets need a tag-triggered run, and the canary's lectures read intersphinx from nine live sites. A schedule on main doesn't touch those legs, which run on staging branches and publish-* tags.
What a schedule catches is weekly drift. No consumer schedules its preview or publish path, so a scheduled run catches what moves between releases without a commit here: the floating majors these actions call, the deploy services' APIs, the organisation's shared NETLIFY_AUTH_TOKEN, and each release's first by-tag run through those paths.
The alert has to live in this repository. GitHub disables a repository's schedules after 60 days without activity. The canary's cache.yml stopped after 2026-04-12 and ran again only after a manual re-enable on 2026-08-05, while the eight red Sunday runs before it (2026-02-22 to 2026-04-12) reached no one. The canary's last push was on 2026-11-12, so without a watcher its schedules would stop again around 2027-01-11.
Question
Decision
Should the release gate also build the canary's lectures/?
No: the gate stays on its frozen fixture. A staged release that changes what a build reports also dispatches the lecture matrix at the candidate, before v0 moves
How does a failed or missing scheduled run reach a person?
A scheduled canary-watch.yml here reads the canary's workflows and runs through the API. It goes red when a workflow isn't active (disabled_inactivity included), when the newest scheduled run is missing or too old, or when that run failed. It keeps one issue here, assigned from a repository variable, and the issue gives the gh workflow enable command
A scheduled publish deploys over the gate's fixture on the canary's one Pages site
Accepted and documented. Checks read what the site serves, and the publish cron stays clear of the release window
After v0 moves
The release procedure dispatches the canary's ci.yml, cache.yml and publish.yml once, from the first release after this issue's canary PR lands
Proposed order: canary-watch.yml first, then one canary PR for this issue and #310, then the drift check and the docs (TESTING.md, CONTRIBUTING's staged-release steps).
Activity
docs(plan): v0 is on v0.17.0; release gating's open item is #302 (#303)
mmcky commentedon Nov 10, 2026
Carried here from #100 when it closed: detect a canary that stops running, not only one that goes red. #100 found that the canary's weekly
cache.ymlbuild ran every Sunday until 2026-04-12, then missed 16 Sundays in a row and gave no signal, because a run that never happens looks the same as one that passes. That is the pattern of GitHub disabling a schedule after 60 days without activity in the repository. Schedulingci.ymlandpublish.ymlhere adds two more paths that could go quiet the same way. So this issue's question of how a failed scheduled run reaches a person should cover a run that doesn't happen at all, for example a check that the newest scheduled run is less than a week old. Low priority, with the rest of this issue.Claude Code (on behalf of mmcky)
mmcky commentedon Nov 15, 2026
Issue check (2026-11-15, against
mainatc21e481and the canary'smainatafd76d2) and the maintainer's decisions (2026-11-16). The change holds in substance, revised in four ways.pull_requestevent, release assets need a tag-triggered run, and the canary's lectures read intersphinx from nine live sites. A schedule onmaindoesn't touch those legs, which run on staging branches andpublish-*tags.NETLIFY_AUTH_TOKEN, and each release's first by-tag run through those paths.ci.ymlcan't be scheduled as it is. It restores withsave-cache: 'true', which onmainwould save a build cache that every pull request then restores. It is rebuilt fromtemplates/ci.ymltogether with Canary: bring itsci.yml,cache.ymlandpublish.ymlback in line withtemplates/#310; the template saves no cache, and keeping the build cache is a row in Canary: bring itsci.yml,cache.ymlandpublish.ymlback in line withtemplates/#310's keep-or-adopt table.cache.ymlstopped after 2026-04-12 and ran again only after a manual re-enable on 2026-08-05, while the eight red Sunday runs before it (2026-02-22 to 2026-04-12) reached no one. The canary's last push was on 2026-11-12, so without a watcher its schedules would stop again around 2027-01-11.lectures/?v0movescanary-watch.ymlhere reads the canary's workflows and runs through the API. It goes red when a workflow isn'tactive(disabled_inactivityincluded), when the newest scheduled run is missing or too old, or when that run failed. It keeps one issue here, assigned from a repository variable, and the issue gives thegh workflow enablecommandv0movesci.yml,cache.ymlandpublish.ymlonce, from the first release after this issue's canary PR landsProposed order:
canary-watch.ymlfirst, then one canary PR for this issue and #310, then the drift check and the docs (TESTING.md, CONTRIBUTING's staged-release steps).Claude Code (on behalf of mmcky)
fix(canary-watch): the same red after a green run is reported; only t…
ci(canary-watch): watch the canary's scheduled workflows from here (#302
ci: ci.yml, cache.yml and publish.yml as templates/ at v0, scheduled;…
ci(canary-watch): check the canary for drift from templates/ at v0 (#310
docs: point versioning.md at the release gate; PLAN's issue rows (#310)
ci(canary-watch): close the drift check's gaps from review (#310)
ci(canary-watch): check the canary for drift from templates/ at v0 (#310
release: v0.21.0 (#334)