Repository navigation
🌱 ci: read the crux file-size ratchet baseline from crux main instead of a stale pin - #17222
Conversation
|
Hive status — updated in place; edits do not notify. Full history is in the Activity Monitor. hive/awaiting-ci: shared-ci incident #17224 is closed and FOSSA is green on sibling PRs, so the branch was rebased onto current v5 (no content change, 71f0a88 → 58afd77) to re-run the stale red 🐝 Hive Agent: — hive: agent=scanner backend=copilot model=claude-fable-5.1 copilot=1.0.88 Review findings from Conclusion: the change does what it says and is safe as written, but it conflicts with #17152 (human-authored, same line, opposite answer). One of the two must be closed; that is a maintainer's call, not mine. What I verified (read the workflow at head
Not duplicates — conflicting. Same file, same line, different content: Hive-authored review; verified against the tree at the reviewed commit. Review findings from Conclusion: the mechanics are correct, but this PR conflicts with #17152 (not a duplicate — both edit What I verified
Why this needs a person rather than a hive call The two issues were each "decided by hive" in opposite directions: #17149 → keep the pin, bump it (PR #17152); #17221 → track
Since Review findings from HUMAN DECISION NEEDED — accept running a shell script fetched from a floating Delta re-review (security), Prior findings at the new head:
|
Open PR overlap checkBase branch: Conflicts are checked with every PR merged onto the current tip of
|
71f0a88
ℹ️ Sentinel notice — informational; author is trusted (hive agent), merge is not blockedHive flagged this PR (author @hivecommons-hive[bot], head
Hive did not add the |
There was a problem hiding this comment.
docs-currency — reviewed head 71f0a88.
In-scope: the operator-facing text moved with the behaviour. The header comment (.github/workflows/crux-file-size-ratchet.yml:6-8) now explains why the baseline is read live and cites crux#124, and the step name (:34) no longer says "pinned". This is exactly the replacement #17221 prescribed; no other file in hive references the old pin SHA. No in-scope blocker.
Out-of-scope (adjacent, for backlog): hive still carries a second second-copy of the baseline — src/internal/testutil/crux_file_size_ratchet.txt, consumed by TestCruxFileSizeRatchet (src/internal/testutil/crux_file_size_ratchet_test.go:29-33 calls it "a verbatim copy of crux's .github/go-file-size-ratchet.txt" refreshed "when crux's copy is next refreshed"). It is more stale than the pin this PR removes: agent/manager.go 1825 vs crux main 1657, agent/manager_kick.go 1613 vs 1348, and it lists four files crux has since dropped. The new workflow catches what the test misses, so nothing regresses here, but the test's own doc comment now overstates its currency and the drift class #17221 describes lives on in that file.
— hive: agent=adjudicator-5 backend=copilot model=claude-fable-5.1 copilot=1.0.88
Review follow-ups (below
|
|
CI triage (scanner), head 71f0a88 unchanged: the red Deferring; no code change needed. The sweep merges once the gate re-runs green. 🐝 Hive Agent: — hive: agent=scanner backend=copilot model=claude-fable-5.1 copilot=1.0.88 |
|
The red check here ( 🐝 Hive Agent: — hive: agent=scanner backend=copilot model=claude-fable-5.1 copilot=1.0.88 |
|
Shared CI incident: the red check(s) on this head are GitHub release-download 504s, tracked in #17224 — deferred until the baseline recovers; no changes needed here. 🐝 Hive Agent: — hive: agent=scanner backend=copilot model=claude-fable-5.1 copilot=1.0.88 |
|
CI note: the only red check on 71f0a88 is FOSSA license scan, and it failed before scanning anything — the action could not fetch the FOSSA CLI ( 🐝 Hive Agent: — hive: agent=scanner backend=copilot model=claude-fable-5.1 copilot=1.0.88 |
There was a problem hiding this comment.
Decision needed: @clubanderson should decide whether hive CI may execute crux's ratchet script from an unpinned main ref (this PR) or must keep a pinned SHA that is bumped on every crux baseline change (#17152).
Conclusion: the diff does what #17221 asks and is mechanically correct, but it reverses a deliberate design choice in a .github/workflows file, and an open PR (#17152) makes the opposite choice on the same line. A maintainer has to pick one; both cannot merge.
What I verified
.github/workflows/crux-file-size-ratchet.yml:38—ref: 13372ca…→ref: main. Withfetch-depth: 1and the non-cone sparse checkout of the two files,actions/checkoutresolves a branch name fine; nothing else in the job depends on the ref being a SHA.- The job already runs
persist-credentials: falseandpermissions: contents: read(lines 19–20, 31, 40), so the script it executes (crux-ratchet/.github/scripts/check-file-size-ratchet.sh, read at cruxmain) has no write credential. I read that script: it isfind/wc/awkoversrc/crux, no network, exit codes 0/1/2. - The header comment at lines 6–7 on
v5says "Pin the script and baseline together to a reviewed crux commit" — the pin was the #15331 design, not an accident. This PR rewrites that comment, which is the right thing to do if the design changes.
Why this needs a person, not the hive
- #17221 (this PR's issue) says "Decision taken by hive: use
ref: main"; #17149 (the issue behind #17152) says "Decision taken by hive: keep the pinned-SHA design". Two hive lanes decided opposite ways on the same question within 9 hours. That is exactly the case where the hive should not pick. - The tradeoff is real:
ref: mainremoves the drift class (crux lowered baselines twice in a week and broke its ownmainfor 4h on 11-07), but it means hive CI runs whatever bash is on cruxmainat check time. Same org, read-only token — low blast radius — but it is still an unpinned third-repo script in a workflow, and the repo's existing convention (everyuses:here is SHA-pinned) points the other way. - A middle path a maintainer may prefer: pin the script SHA, read only the 12-line baseline from
main. Neither PR does that.
Relation to #17152: not a duplicate — it bumps the pin to 7f93191…; this PR removes the pin. They conflict on line 38. If this PR is chosen, #17152 should be closed; if #17152 is chosen, this one should. I have not closed anything.
Correctness, style and docs otherwise look fine; the comment rewrite keeps the #15331 citation and adds crux#124.
— hive: agent=reviewer backend=copilot model=claude-fable-5.1 copilot=1.0.88
|
/lgtm Posted by @clubanderson from the Hive dashboard; the merge gate decides from here. |
|
@hivecommons-hive[bot]: you cannot LGTM your own PR. DetailsIn response to this:
Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository. |
|
/approve Posted by @clubanderson from the Hive dashboard; the merge gate decides from here. |
|
[APPROVALNOTIFIER] This PR is NOT APPROVED This pull-request has been approved by: hivecommons-hive[bot] The full list of commands accepted by this bot can be found here. DetailsNeeds approval from an approver in each of these files:Approvers can indicate their approval by writing |
There was a problem hiding this comment.
Decision needed: @kellyaa should decide whether hive's crux-file-size-ratchet check should execute crux's ratchet script from a floating ref: main (this PR, #17222) or from a bumped pinned SHA (#17152) — the two PRs change the same line and cannot both merge.
Conclusion: the diff does what #17221 asks and is mechanically correct, but it conflicts with #17152 on the same line and it moves a CI step from a pinned to a floating ref for a script that is executed, which is a maintainer call rather than a code defect.
What I verified (workflow read at v5 and at head f0243e0):
- The only functional change is
.github/workflows/crux-file-size-ratchet.yml:38:ref: 13372ca7…→ref: main. The script path, baseline path, sparse-checkout,persist-credentials: false,permissions: contents: read(line 20) and therun:invocation (line 50) are unchanged, so the job still has no write credential and the fetched tree cannot exfiltrate a token. - The fetched file
.github/scripts/check-file-size-ratchet.shis run withbashat line 50. Withref: main, whoever can push to cruxmaincan change what hive's PR CI executes without a hive-side review. The job is read-only and 5-minute-bounded, so the blast radius is small, but it is non-zero and it is the exact property the original pin (#15331) bought. - #17221 itself anticipates this: "If maintainers prefer to keep a pin for supply-chain reasons, the bump is the fallback." So the question is open in the issue, not answered.
Conflict, not duplicate, with #17152. Both PRs edit line 38 of the same file against v5. #17152 bumps the pin to 7f931914 (crux main 2026-11-08); this PR drops the pin. They resolve #17221/#17149 by opposite strategies: #17152 fixes today's drift and will drift again on crux's next baseline lowering (crux#124 pattern); #17222 removes the drift class but adopts a floating ref. Whichever lands, the other should be closed — I have not closed anything.
If the floating ref is accepted, the header comment this PR adds (lines 6–8) explains the choice accurately. If the pin is preferred, #17152 is the smaller change and already carries a fresh SHA.
— hive: agent=reviewer backend=copilot model=claude-fable-5.1 copilot=1.0.88
There was a problem hiding this comment.
HUMAN DECISION NEEDED — pin vs. track main for the crux ratchet checkout: accept this PR's ref: main or land #17152's re-pin (both edit the same line).
Delta re-review (f0243e0 → ee2f104): the delta is a v5 merge only (#17151's automerge files); the PR's own diff vs v5 is unchanged and still limited to .github/workflows/crux-file-size-ratchet.yml.
- Still present (in-scope):
.github/workflows/crux-file-size-ratchet.yml:38ref: main— the checked-outcheck-file-size-ratchet.shis executed in CI from an unpinned ref. #17221 explicitly defers this to maintainers ("If maintainers prefer to keep a pin for supply-chain reasons, the bump is the fallback"), so it is not the hive's call. - Still present (in-scope): #17152 (re-pin to
7f931914) is still OPEN and changes the same line; whichever lands second conflicts. Pick one and close the other.
Otherwise the diff matches #17221's exact proposed replacement (comment text, step rename, ref: main) with no scope creep.
— hive: agent=ci-maintainer backend=copilot model=claude-fable-5.1 copilot=1.0.88
There was a problem hiding this comment.
HUMAN DECISION NEEDED — accept executing a shell script from floating hivecommons/crux@main in hive CI (supply-chain trade-off the linked issue explicitly defers to maintainers), or fall back to a bumped pin.
Security delta re-review at b09656ad (previous: ee2f1042). The only change since the prior head is Merge branch 'v5' bringing in #17234 (crux gates service); .github/workflows/crux-file-size-ratchet.yml is byte-identical to the prior head, so nothing new to assess in the delta.
Prior findings:
- [security] still present —
.github/workflows/crux-file-size-ratchet.yml:38checks outref: mainand:49runsbash .../check-file-size-ratchet.shfrom it. Blast radius is deliberately small (permissions: contents: readat :20-21,persist-credentials: falseat :31/:40,pull_requestnotpull_request_targetat :10, same-org source), so this is not a live exploit — but it moves hive from "reviewed SHA" to "whoever can push crux main can run code in hive PR CI". #17221 names this as the maintainer call ("If maintainers prefer to keep a pin for supply-chain reasons, the bump is the fallback"); the answer is not stated there, so it needs a person. - [intent-alignment] still present — #17152 (re-pin to current crux main) is still OPEN and edits the same line; one of the two must yield.
If the floating ref is accepted, no further change is needed from this perspective.
— hive: agent=ci-maintainer backend=copilot model=claude-fable-5.1 copilot=1.0.88
There was a problem hiding this comment.
HUMAN DECISION NEEDED — accept a floating ref: main cross-repo checkout (executes crux's shell script unpinned) vs. re-pinning as #17152 does; the two PRs edit the same line.
Delta re-review (b09656a → 415d3dc): the only new commit is Merge branch 'v5' bringing in #17242 (dashboard "board pressure" rename — 12 files, none touched by this PR). The PR's own diff to .github/workflows/crux-file-size-ratchet.yml is unchanged.
Prior findings at this head:
- Still present —
.github/workflows/crux-file-size-ratchet.yml:38ref: mainconflicts with #17152 (still OPEN, same file, same line, re-pins to7f931914). One of the two must be closed; #17221 itself names the pin bump as the fallback "if maintainers prefer to keep a pin for supply-chain reasons", so this is the decision #17221 left to maintainers. - Still present —
.github/workflows/crux-file-size-ratchet.yml:38moves from a pinned SHA to floatingmainfor a checkout whose.github/scripts/check-file-size-ratchet.shis then executed (:48). Same org,persist-credentials: false,contents: readonly — but it is still an unpinned script execution a human should own.
Intent alignment otherwise holds: the diff is exactly the replacement #17221's "Exact replacement" block specifies, with no scope creep.
— hive: agent=ci-maintainer backend=copilot model=claude-fable-5.1 copilot=1.0.88
There was a problem hiding this comment.
Decision needed: @clubanderson should decide whether hive CI may execute crux's ratchet script unpinned from ref: main (this PR) or must keep a reviewed SHA pin that is bumped on drift (#17152) — the two PRs edit the same line and cannot both merge.
Conclusion: the change does what #17221 asks and I found nothing wrong with it mechanically, but it removes a pin on a script this workflow executes from another repository, and #17152 fixes the same drift the opposite way (re-pin to crux 7f93191). Same line of .github/workflows/crux-file-size-ratchet.yml, incompatible strategies — only one can land, and that choice is a supply-chain policy call rather than a code question. I have not closed anything.
What I verified:
crux-file-size-ratchet.yml:48runsbash crux-ratchet/.github/scripts/check-file-size-ratchet.sh <baseline>. After this PR the script and baseline come from whatever is on cruxmainat check time (:38,ref: main). Crux's.github/is not overwritten by the crux-sync publish (the script's own header says it lives there for that reason), so it is maintained by crux committers directly, not derived from hive.- Blast radius is small: the job has
permissions: contents: read(:19-20), both checkouts usepersist-credentials: false(:30,:39), and therunstep references no secret. A bad script on cruxmaincould flip this check's verdict, not reach anything else. - The script's interface at crux
maintoday is unchanged — still[baseline]as the single positional arg, exit 0/1/2 — so the invocation at:48keeps working; the baseline file still uses<lines> <path>lines. - The drift claim in #17221 is real: crux
main's baseline currently listsagent/manager.go 1657where the13372ca7pin allowed 1658, which is exactly the case that broke crux main on 11-07.
The decision: #17221 (filed by the hive's ci-maintainer) chose ref: main and gave its reasoning; #17152 (a hive contributor, from #17149) chose a fresh pin. Unpinning removes the drift class permanently; pinning keeps the "reviewed commit executes in our CI" property that the original #15331 workflow comment at :6-7 asked for. Either is defensible; what should not happen is both landing by merge order. Whichever survives, the other should be closed with a pointer.
If the pin is kept, note that #17152's pin will need bumping every time crux lowers a baseline (crux#118, crux#125 did so within one week).
— hive: agent=reviewer backend=copilot model=claude-fable-5.1 copilot=1.0.88
There was a problem hiding this comment.
Decision needed: @clubanderson should decide whether hive's crux-file-size-ratchet workflow tracks crux main (this PR, #17222) or keeps a pinned SHA that is bumped (#17152) — the two PRs change the same line to opposite designs and cannot both merge.
Conclusion: the change does what it says and CI is green on it, but it is one side of an unsettled A/B. #17152 edits the same line of .github/workflows/crux-file-size-ratchet.yml (ref: at line 37) to a new pin (7f93191…), while this PR sets ref: main. These conflict rather than duplicate — whichever lands first makes the other dirty — and the two linked issues record opposite choices: #17149 says "keep the pinned-SHA design … just move the pin", #17221 says "use ref: main … if maintainers prefer to keep a pin for supply-chain reasons, the bump is the fallback". The issue itself leaves this to maintainers, so I am not picking.
What I verified
- Diff is confined to the one workflow: a comment rewrite, the step rename, and
ref: 13372ca7…→ref: main(lines 34–37 at3992238).permissions: contents: readandpersist-credentials: falseon both checkouts are unchanged (lines 20–21, 31, 39). - The checked-out script
crux/.github/scripts/check-file-size-ratchet.shis executed withbashfromsrc/crux(line 46). Withref: main, whoever can land a commit to cruxmainunder.github/(which crux-sync does not overwrite, per the script header) controls a shell step in hive's PR CI. Blast radius is small — read-only token, no persisted credentials, same org — but it is a workflow executing unpinned cross-repo code, which is why this is a code-owner call rather than mine. - The drift mechanism described in #17221 is real: crux's baseline is lowered whenever a sync shrinks a file, so a pin is stale by construction. Tracking
mainremoves that class; a pin removes it only until the next lowering.
Trade-off in one line: ref: main = never stale, but hive PRs can go red with no change of their own and hive runs unpinned crux shell; pinned SHA = reviewed/reproducible, but needs a bump on every crux baseline change (the failure that produced #17149 and #17221 within a week).
I have not closed or approved anything. Whichever is chosen, the other PR and its issue should be closed as superseded so the two "decision taken by hive" notes stop contradicting each other.
— hive: agent=reviewer-2 backend=copilot model=claude-fable-5.1 copilot=1.0.88
|
Thank you for your contribution! Your PR has been merged. We'd love to hear how your experience was: share feedback |
The pinned crux commit 13372ca7 (2026-10-29) drifted from crux main's baselines; crux lowers baselines on every sync that shrinks a file (crux#124). A PR within hive's stale headroom was green here and broke crux main for 4h on 2026-11-07. Crux is the source of truth for the baseline so hive should not carry a second copy that can drift.
Closes #17221
— hive: agent=scanner backend=copilot model=claude-fable-5.1 copilot=1.0.88