Skip to content

ci: allow the Claude Code installer hosts in the fork review lanes - #12201

Closed
Skymore wants to merge 1 commit into
kirodotdev:mainfrom
Skymore:ci/fork-lanes-allow-claude-installer
Closed

Skymore wants to merge 1 commit into
kirodotdev:mainfrom
Skymore:ci/fork-lanes-allow-claude-installer

Conversation

@Skymore

@Skymore Skymore commented Sep 20, 2026

Copy link
Copy Markdown
Contributor

Problem / Motivation

Every fork pull request's Opus lane has posted review incomplete since 2026-09-18 21:37Z, and PR Readiness aggregates that lane, so no fork PR can reach a mergeable state. The failing step is the Claude Code install inside anthropics/claude-code-action, before any model call:

Sat, 20 Sep 2026 03:53:52 GMT:domain not allowed: claude.ai.
curl: (7) Failed to connect to claude.ai port 443 after 1 ms: Couldn't connect to server
curl: (7) Failed to connect to claude.ai port 443 after 3 ms: Couldn't connect to server
curl: (7) Failed to connect to claude.ai port 443 after 1 ms: Couldn't connect to server
##[error]Action failed with error: Failed to install Claude Code after 3 attempts: Error: Install failed with exit code 7

Taken from Fork Opus 4.8 Review run 35487678930 on #12126. The same step fails identically in the Design, UX, First Principles and Security Scope fork lanes; those lanes are advisory, so they finish neutral instead of red.

Why it matters

Across the fork-opus-review.yml run history there is no completed review after 2026-09-18 21:25Z (run 35396661118, the last success) -- more than 150 non-skipped runs since, all failed. Outside contributors cannot land anything, and re-running a lane cannot help because the refused host is the same every time.

What changed (motivation -> approach -> change)

claude-code-action installs its runtime by running the installer script served from https://claude.ai/install.sh. That script sets DOWNLOAD_BASE_URL="https://downloads.claude.ai/claude-code-releases" and fetches the version manifest and the binary from there. Neither host is a GitHub or npm host, so the six fork lanes' blocking egress allowlists, which carry only Bedrock, STS, GitHub and npm hosts, refuse the first fetch.

Why now: the last successful run executed under step-security/harden-runner v2.20.1, which let this fetch through. #8179 bumped the six lanes to v2.21.1 at 21:37Z the same day, and nothing has completed since. #11976 restored the bun release-asset host that the same bump started refusing; with bun installing again, the lane runs one step further and stops on this host instead.

The change adds claude.ai:443 and downloads.claude.ai:443 to the allowlist of each job that runs anthropics/claude-code-action: the single model job in fork-opus-review.yml, fork-gpt-review.yml, fork-design-review.yml, fork-ux-review.yml and fork-first-principles-review.yml, and the generate job in fork-security-scope-review.yml. The three non-model jobs of the security-scope lane keep their narrower lists, following the least-privilege split #11976 established.

workflow_run lanes execute the default branch's copy of the yaml, so this PR cannot exercise its own change; the effect is only visible after merge. This is a fork PR touching .github/workflows/, so it needs the allow-fork-workflow-change label from a maintainer.

Backwards compatibility

Compatible: two hosts are added to six existing allowlists; no endpoint is removed and no job outside the model jobs changes.

Tests

test/test_ai_review_workflows.py::TestForkLaneClaudeInstallerEgress (12 parametrized cases) pins the coupling in both directions, mirroring TestForkLaneBunEgress: every blocking-egress job that runs claude-code-action allows both installer hosts, and every blocking-egress job that runs no model does not. Without the workflow edits the first test fails on all six lanes; with them all 12 pass. The full test_ai_review_workflows.py and test_pr_quality_gates.py (835 tests) pass on this branch.

Manual verification

The failing step and the refused host are read from the Fork Opus 4.8 Review job logs on #12126 (run linked above); the installer's download host is read from the installer script itself. The repaired lanes can only be observed on the first fork PR run after this merges.

claude-code-action installs its runtime by running the installer script
served from claude.ai, and that script fetches the release manifest and
binary from downloads.claude.ai. Neither host is in the fork lanes'
blocking egress allowlists, so harden-runner refuses the first fetch
(`domain not allowed: claude.ai`), curl exits 7 on all three attempts,
the action dies before any model call, and the lane posts `review
incomplete`. PR Readiness aggregates the Opus lane, so every fork PR is
red.

The last fork Opus review that completed ran on 2026-09-18 21:25Z under
harden-runner v2.20.1; the bump to v2.21.1 (kirodotdev#8179) landed twelve
minutes later and no fork lane has completed since. kirodotdev#11976 restored the
bun host the same bump started refusing, which moved the failure one
step later, to this one.

Only the six jobs that run claude-code-action get the two hosts; the
three non-model jobs in fork-security-scope-review.yml keep their
narrower list. Tests pin both directions, because workflow_run lanes
execute the default branch's yaml and cannot exercise this change.
@buluoray

Copy link
Copy Markdown
Contributor

This diagnosis was right, and it was first. Recording that plainly, because the PR is now superseded and the reason it lost is mechanical rather than anything about its quality.

Superseded: #12469 landed the same fix at 17:27Z today, and origin/main now carries both claude.ai:443 and downloads.claude.ai:443 in all six fork lanes. So there is nothing left in the workflow half of this PR to merge, and the merge conflict is now literally that main already contains these lines.

The test half is covered too, which I checked rather than assumed before saying so. TestForkLaneBubblewrapBootstrapEgress on main folds both installer hosts into its ENDPOINTS tuple, so its positive check asserts their presence on every claude-code-action job behind a blocking egress policy, and its test_jobs_that_run_no_model_keep_the_narrower_allowlist asserts their absence from every non-model job — the same least-privilege half this PR wrote, including the fork-security-scope-review.yml four-jobs-one-model case your comment calls out. Main additionally forbids code.claude.com and www.anthropic.com, which appear in install.sh only inside its error strings and are never fetched.

You filed this 25 hours earlier. This PR opened 2026-09-20 05:26Z; #12469 opened 2026-09-21 06:30Z. Same two hosts, same root cause.

Why it never got reviewed is worth stating, because it is not about this change. It went CONFLICTING, and a conflicting PR cannot dispatch any pull_request-event workflow — those run on refs/pull/N/merge, which GitHub cannot generate while a conflict exists. Nothing goes red; the count just collapses. This PR carries 4 check-runs where a healthy one in this repo carries around ninety. So no reviewer, human or bot, was ever handed it.

And the reviewers it could not reach were the exact lanes it was fixing. A fork PR repairing the fork review lanes is un-reviewable twice over: once because workflow_run lanes always execute the default branch's copy of the yaml, so the change cannot exercise itself (your docstring says this, and it is why both PRs wrote a test as the only pre-merge guard), and again because the lanes were dead for the very reason this PR identified. The fix for the broken fork reviewers could not be reviewed by the fork reviewers.

Separately: your diagnosis has been confirmed live rather than only by reading. I ran a maintainer re-run of Fast Gate on another fork PR earlier today, specifically to test whether those dead lanes were a transient rate-limit window. The result was three identical curl: (7) Failed to connect to claude.ai port 443 attempts ending in Failed to install Claude Code after 3 attempts, with the bun install from registry.npmjs.org immediately above succeeding — exactly the sequence described here, a day after it was described.

Closing this is the maintainers' call, not mine. Nothing further is needed from you.

@coffeencoke

Copy link
Copy Markdown

Duplicate? #12469

@Skymore

Skymore commented Sep 22, 2026

Copy link
Copy Markdown
Contributor Author

Superseded by #12469, which landed the same two hosts (claude.ai:443, downloads.claude.ai:443) in the fork review lanes' egress allowlists on 09-21. Closing; nothing here remains to merge.

@Skymore Skymore closed this Sep 22, 2026
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Sep 22, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

fork Pull request from a fork (external contributor) merge conflict Branch has merge conflicts with its base — author must resolve before merge

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants