Repository navigation
ci: allow the bun release-asset host in the fork review lanes - #11976
Conversation
|
Intent: Let the six fork review lanes install bun again, by allowing the one GitHub host that serves bun's release asset, so Not a goal: changing any action pin, changing what the reviewers review, or answering the separate question of which upstream change made this start on 2026-09-18. This PR only widens each affected job's egress allowlist by that one host, and pins the coupling with a test. |
Design Review (Fable 5) — 🟡 CONCERNSDesign-level review of Design-Verdict: CONCERNS Undocumented egress widening: the diff adds a second host, WatchThe description documents exactly one change ("Change | allow that one host", "Reach | 6 jobs across 6 files", tests "returns 12 passed"), but the diff also adds [DESIGN-REVIEWED] 87cf02d |
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
First Principles Review (Fable 5) — 🟡 CONCERNSPremise-level review of First-Principles-Verdict: CONCERNS The description declares one host; the diff allows two — the Not justified as shipped
What this change shipsInventory (4 items) — 2 justifiedIntent: unwedge fork PR merging by letting the six fork review lanes download bun through their blocking egress allowlists — a FIX. (Grepped
[FIRST-PRINCIPLES-REVIEWED] 87cf02d |
Opus 4.8 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
231cfa7 to
4360de2
Compare
|
One more host is needed alongside this, or the fork GPT 5.6 lane stays red after it merges: To be clear about scope, since the four green lane comments above are the same-repo lanes on this in-repo branch: the breakage is only on the three seconds before the model call's first reconnect, then fails closed. Its allowlist carries the classic It is GPT-specific — the fork Opus, Design and First-Principles lanes fail on Since this PR already touches @pepmach — could you take the review? Every model-driven lane on every fork PR is currently reporting |
Two hosts are missing from the fork lanes' blocking egress allowlists, and each one takes a different set of lanes down. claude-code-action runs oven-sh/setup-bun, which downloads bun from a GitHub release asset. GitHub serves that from release-assets.githubusercontent.com, a different host from the objects.githubusercontent.com the six lanes already allow, so harden-runner refused the download, bun never installed, and the reviewer step died `bun: command not found` (exit 127). That is the Opus, Design, UX, First-Principles and Security-Scope lanes. The GPT lane breaks earlier and on a different host. Its two review passes run a CLI whose amazon-bedrock provider posts to bedrock-mantle.us-east-1.api.aws, which the allowlist's classic bedrock-runtime and bedrock entries do not cover. harden-runner logs `domain not allowed: bedrock-mantle.us-east-1.api.aws`, both passes then retry five times and exit 1 with `Connection failed: error sending request`. Either way the lane posts `review incomplete`. PR Readiness aggregates those lanes, so no fork PR can reach a mergeable state. Only the jobs that make each call get the endpoint. fork-security-scope-review.yml blocks egress in four jobs and runs the model in one, so its other three allowlists are unchanged, and the mantle host goes to the GPT lane alone. Tests pin both couplings in both directions, because workflow_run lanes execute the default branch's yaml and cannot exercise this change. Fixes #11911
4360de2 to
87cf02d
Compare
|
@rnoack1 Accepted and folded in. Head is now I verified it independently before pushing rather than taking the host on trust:
So your scoping is right, and it also explains something I had left as an open discrepancy: the GPT lane was the one lane with no bun signature. Two notes on how I implemented it. The test is a second class, That same run also blocked On folding it in here rather than a follow-up: this PR has no approvals to invalidate, it already touches that file, and shipping it alone would have left the outage half in place. The PR body now states both root causes, both failure signatures and the credit. |
…tdev#11976) Two hosts are missing from the fork lanes' blocking egress allowlists, and each one takes a different set of lanes down. claude-code-action runs oven-sh/setup-bun, which downloads bun from a GitHub release asset. GitHub serves that from release-assets.githubusercontent.com, a different host from the objects.githubusercontent.com the six lanes already allow, so harden-runner refused the download, bun never installed, and the reviewer step died `bun: command not found` (exit 127). That is the Opus, Design, UX, First-Principles and Security-Scope lanes. The GPT lane breaks earlier and on a different host. Its two review passes run a CLI whose amazon-bedrock provider posts to bedrock-mantle.us-east-1.api.aws, which the allowlist's classic bedrock-runtime and bedrock entries do not cover. harden-runner logs `domain not allowed: bedrock-mantle.us-east-1.api.aws`, both passes then retry five times and exit 1 with `Connection failed: error sending request`. Either way the lane posts `review incomplete`. PR Readiness aggregates those lanes, so no fork PR can reach a mergeable state. Only the jobs that make each call get the endpoint. fork-security-scope-review.yml blocks egress in four jobs and runs the model in one, so its other three allowlists are unchanged, and the mantle host goes to the GPT lane alone. Tests pin both couplings in both directions, because workflow_run lanes execute the default branch's yaml and cannot exercise this change. Fixes kirodotdev#11911
Problem / Motivation
Every fork pull request has been unmergeable since 2026-09-18 21:25Z. All six fork review lanes finish
review incomplete, andPR Readiness, the only required status onmain, aggregates them. It is red on every fork PR, and re-running a lane does not clear it.No lane reaches a verdict, and two different missing hosts are why. Five lanes die on a missing binary before any model call:
The GPT lane dies at the model call instead, with no bun in sight:
Why it matters
No outside contributor can land anything, and the repository cannot merge its own repairs. #11888 is an already-approved fix for a red macOS shard, and it is stuck behind this.
A fork author has no way out from the branch either. Per #8445 the fork lanes consume no
/ai-review overridemarker, so there is nothing to override.Credit for the diagnosis is not ours. @buluoray found the bun release-asset mechanism in #11911. @rnoack1 found the second host on this PR: without it the fork GPT lane stays red after this merges.
What changed (motivation -> approach -> change)
claude-code-actioninstalls bun before it runs anything, usingoven-sh/setup-bun. That fetches bun as a GitHub release asset, and GitHub serves a release download by redirecting torelease-assets.githubusercontent.com. That host is notobjects.githubusercontent.com, which these allowlists already carried. All six lanes runstep-security/harden-runnerwithegress-policy: block, so the download was refused, bun never installed, and the step exited 127 on the missing binary.The GPT lane fails earlier and on a different host. Its two review passes run a CLI configured with
model_provider = "amazon-bedrock", and that provider posts tohttps://bedrock-mantle.us-east-1.api.aws/openai/v1/responses. The allowlist carried the classicbedrock-runtime.*andbedrock.*hosts but not the mantle one, so both passes retried five times, exited 1, and the lane failed closed. Itsclaude-code-actionuse is only the Opus adjudication sub-step, which is why the bun signature never appears there.review incomplete,PR Readinessred, no fork PR mergeablerelease-assets.githubusercontent.comblocked; bun's release asset is served therebedrock-mantle.us-east-1.api.awsblocked; the GPT passes call Bedrock thereOnly the jobs that make each call get the endpoint. Five lanes have one bun job each.
fork-security-scope-review.ymlblocks egress in four jobs and runs the action in one,generate; itsvalidate,adjudicateandpublishjobs fetch nothing but Actions artifacts, so their allowlists are unchanged. The mantle host goes to the GPT lane alone, since the other five reach Bedrock through the runtime host.Both hosts are named exactly, never as a wildcard.
objects.githubusercontent.comis already allowed and carries the same class of GitHub-owned asset bytes, and the mantle host is the same Bedrock service the lane is already permitted to call, in the same region and under the same role.The alternative for bun was installing it ahead of the action, which needs no allowlist change, since
registry.npmjs.orgis already allowed. It was rejected. The action pins its ownbun-versioninternally, 1.3.14 today, andsetup-bunskips its download only when the version already present matches that pin. A pre-installed bun works until the action bumps the pin, and then all six lanes break again the same way and just as silently. Allowing the host does not depend on a version.Backwards compatibility
Compatible: this adds permitted destinations to fork egress allowlists and removes nothing. No input, field, type or contract is narrowed.
Tests
test/test_ai_review_workflows.pycarries two classes, 19 tests, each pinning a coupling in both directions:TestForkLaneBunEgress, parametrized over all six lanes: a job that runsclaude-code-actionbehindegress-policy: blockmust allow the release-asset host, and a job that runs no model must not.TestForkGptLaneMantleEgress: the job that configures theamazon-bedrockprovider must allow the mantle host, and every lane job that configures no such provider must not.All four assertions are mutation-checked. Dropping either endpoint fails the positive test for that host; adding the release-asset host to the security-scope
validatejob, or the mantle host tofork-opus-review.yml, fails the matching negative test; restoring returns 19 passed.test/test_ai_review_workflows.pyplustest/test_pr_quality_gates.py: 823 passed. flake8 and isort clean. mypy reports 8 errors in this file on this branch and the same 8 onorigin/main, all at lines 15 to 9281, none inside the added code, which starts at line 11751.Manual verification
Three facts were checked by hand, because these lanes cannot check themselves.
The bun download URL: requesting
https://github.com/oven-sh/bun/releases/download/bun-v1.3.14/bun-linux-x64.zipanswers302withlocation: https://release-assets.githubusercontent.com/.... That is the exact URLsetup-bunlogs.The bun refusal: in
Fork Opus 4.8 Reviewrun 35422579467, harden-runner logsdomain not allowed: release-assets.githubusercontent.comat04:57:05, the same second the step logs its bun download. The same triple of download,bun: command not foundand exit 127 appears in the design, UX and first-principles lanes.The mantle refusal: in
Fork GPT 5.6 Reviewrun 35442578828, on fork PR #11027, harden-runner logsdomain not allowed: bedrock-mantle.us-east-1.api.awsat12:21:55. The discovery pass starts reconnecting at12:21:59and endsConnection failed: error sending requestat12:22:20; the falsification pass repeats it and the lane fails closed. That run logs no bun error at all.One loose end, stated rather than hidden. That same run also blocked
chatgpt.comin the same second, and this PR does not allow it. The CLI is pinned tomodel_provider = "amazon-bedrock", so its default ChatGPT backend should not carry a model call, and allowing a host on that suspicion would widen egress for nothing. If a fork GPT lane still fails after this merges withchatgpt.comas the only blocked host in its log, that host is the next line, and the log will say so plainly.This PR's own board cannot prove any of it.
workflow_runlanes always execute the default branch's copy of the workflow, so any lane that runs against this PR reads the old allowlist. A green board here is not evidence for the fix, and a red one is not evidence against it. Confirmation is the first fork PR to run a lane after this merges.Related Issues
Fixes #11911
Pattern harvest
Rule candidate: review-prompt
Pattern: one symptom on a shared surface can have more than one cause. Five lanes and a sixth shared
review incomplete, and the sixth needed a different host. Reading one failing lane's log and generalizing would have shipped a fix that left the outage half in place. Read one log per distinct lane shape.Two suggestions beyond this fix.
First, a correction to the timing theory in #11911. The
claude-code-actionbump is probably not the trigger.fork-security-scope-review.ymlis still on the older action pin, and that pin runs the sameoven-sh/setup-bunat the samebun-version: 1.3.14as the newer one the other five use. So the download predates the bump, and the change is either in GitHub's release-asset serving or in harden-runner enforcement. Which one is not isolated here.Second, the evidence for this outage was sitting in every failed run. harden-runner's post-step prints
domain not allowed: <host>and nothing reads it. A lane that fails while that line is present could name the host in its own summary, which would have turned a day of triage into one line, and would have caught the second host on day one.Checklist
feat|fix|docs|refactor|perf|test|chore|ci|build|revert: ...)docs/ci/ci-and-reviews.mddescribes these lanes as using "a narrow endpoint allowlist" and does not enumerate the hosts, so it does not go stale.