Repository navigation
OUTERLOOP_AUTO_UPDATE: deploy update policy — off by default, release, or main - #331
Conversation
…lt, release, or main The deploy step used to pull main whenever a PAT file was present. That is the right behaviour for the kernel's own developers and the wrong one for an adopter, who installs a release and upgrades when they choose. The policy is now explicit, read from .env before the fetch so a change needs no chain restart: - off (default): the checkout stays as it is. - release: fetch the public repo's release tags and move to the newest by version order (dev/rc pre-releases count). No credential needed. - main: follow every merge. The lab deployment sets this. The environment rollback logic is unchanged: whatever moved the checkout, the venv is synced to the commit that is checked out or the deploy goes back to the pair that matched. init writes the default for Slurm deployments; the .env trust check now also works with BSD stat. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
# Conflicts: # CHANGELOG.md
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
Round 1 — reviewed head 0f2d6cb7 — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.
terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.
Verdict: 1 blocking, 1 advisory.
1 finding attached to the lines below.
A failed sync runs manually upgraded code on the old environment. [deployment+prose] With off, DEPLOY_PREV stays empty, so a manual checkout update followed by a failed uv sync takes this branch and leaves the new checkout running on the old environment. (scripts/tick_deploy.sh:142; high confidence)
Verdict: one blocking off-policy sync-failure issue and one advisory release-tag selection issue. Merged duplicates: deployment and prose agree on scripts/tick_deploy.sh:142; credentials, general, and lifecycle agree on scripts/tick_deploy.sh:107–108. Rejected findings: none; coverage reported no findings.
…ced commit; release tags come from the public remote - A successful sync records the installed commit beside the environment (.venv/.synced-head), and a failed sync rolls back to that commit rather than to "HEAD before the fetch". A checkout the operator moved by hand under the off policy now gets the same code-and-environment guarantee. Without a record the checkout stays, as before. - The release policy asks the public repo for its tags (ls-remote), picks the newest, fetches only that tag with a forced update and resets to it; the checkout's own tags are never consulted, so a stale or local v-tag cannot be chosen. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
Round 1's blocking finding (tick_deploy.sh:142, no thread was attached): fixed in af75951. A successful sync now records the installed commit beside the environment (.venv/.synced-head) and a failed sync rolls back to that commit, not to "HEAD before the fetch". A checkout the operator moved by hand under the off policy gets the same code-and-environment guarantee; without a record the checkout stays, as before. Tests: hand-moved checkout with a differing lockfile rolls back to the recorded commit with no fetch; a successful sync writes the record; no record keeps the checkout. |
There was a problem hiding this comment.
Round 1 — reviewed head af75951b — reviewer hermes/gpt-5.6-terra.
terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.
Verdict: nothing blocking — 1 advisory note.
1 finding attached to the lines below.
A first deploy without the new sync record does not preserve the claimed code-and-environment pairing after a manual checkout change.
1 participant
What
OUTERLOOP_AUTO_UPDATEis the deploy step's update policy, read from.envbefore the fetch so a change needs no chain restart:off(default): the checkout stays as it is. An adopter installs a release and upgrades when they choose.release: fetch the public repo'sv*tags and move to the newest by version order (at one X.Y.Z a final release beats rc beats alpha beats dev). No credential needed.main: follow every merge, as before. This is for the kernel's own developers; the lab deployment sets it.Why
The deploy step pulled
mainwhenever a PAT file was present. That couples "has a bot PAT" to "runs unreleased code", which is right for us and wrong for an adopter. Following every merge stays an internal thing; external auto-update, when opted into, stays at release level.Notes
initwritesOUTERLOOP_AUTO_UPDATE=offinto a Slurm deployment's.envso the knob is visible where it is set. The local loop runs the installed package and is upgraded by hand (docs say how)..envtrust check (owner, not group/world-writable) is factored into one place and now also works with BSDstat, so the deploy tests run on a Mac.releasepicksv0.2.0a1overv0.2.0.dev1,v0.1.1,v0.1.0,v0.1.0rc1,v0.1.0.dev10and ignoresvlatest/v1; at-tag stays put; no tag runs previous code;mainauthenticates through GIT_ASKPASS and the PAT never appears in argv; the.envline wins over the environment in either spelling; an unknown value isoffand says so.🤖 Generated with Claude Code