Skip to content

feat(claude-code-review)!: default to a GitHub-hosted runner - #80

Merged
nkg merged 1 commit into
mainfrom
ci/review-on-hosted-runners
Sep 11, 2026
Merged

nkg merged 1 commit into
mainfrom
ci/review-on-hosted-runners

Conversation

@nkg

@nkg nkg commented Sep 11, 2026

Copy link
Copy Markdown
Owner

Fixes the review job's queueing and its ~1-in-6 hard failures — one change here instead of 27 consumer PRs.

The problem, measured

The review job reads a PR diff and calls the Claude API. It needs no private-network access, no local toolchain and no persistent state — so a self-hosted runner buys it nothing. On a contended pool it cost a lot:

run created job started finished queued actually ran
19:37:04 20:54:07 20:59:14 77 min 5m 07s
22:18:25 23:02:23 23:05:28 44 min 3m 05s
17:25:19 18:26:13 18:29:42 61 min 3m 29s

Two consequences, both observed repeatedly:

  1. Reviews land after the PR has merged. Findings arrive on a closed PR where nothing surfaces them. In one recent batch every round's feedback arrived post-merge and needed a fresh PR to address.
  2. It fails outright roughly 1 run in 6conclusion=failure with steps_run=0, dying at runner assignment. Across three sampled repos: failure=1 success=5 each.

Both are queueing artefacts. Neither is a problem with the review.

Why the default, not per-repo

27 repos consume this workflow. 26 pass no runs-on at all and inherit the default here; only sproncy/monitoring_stack sets it explicitly. Changing the default fixes all 26 at once and leaves the one deliberate opt-out untouched.

⚠️ Breaking, and two things to know

This changes behaviour for every consumer that does not pass runs-on — currently all but monitoring_stack. The escape hatch is unchanged: pass runs-on: '["self-hosted", "linux", "x64"]'.

Consumers pin @v3, and v3 currently points at main (97080ab). Nothing reaches them until the v3 tag is re-pointed at the release containing this commit — worth being deliberate about, given the blast radius.

Hosted minutes are metered on private repos. A review is 3–6 minutes. For context, sproncy is at 653 of 2,000 for September and HordiaLabs at 315, so there is room — but this moves real volume onto that meter across every consuming org.

The image is pinned to ubuntu-24.04 rather than ubuntu-latest so a GitHub-side image bump is an intentional, reviewable change rather than a silent environment shift.

🤖 Generated with Claude Code

https://claude.ai/code/session_0114gaSKj6YinQCkmxVbDACv

The review job reads a PR diff and calls the Claude API. It needs no
private-network access, no local toolchain and no persistent state, so a
self-hosted runner buys it nothing -- while costing a great deal on a
contended pool.

Measured on the sproncy pool: the job queued 40-77 minutes to do 3-6 minutes
of work, so reviews routinely posted after the PR had already merged, and it
failed outright at runner assignment (zero steps executed) in roughly one run
in six. Both are queueing artefacts rather than anything wrong with the
review.

Moving the default also returns capacity to the pool for the jobs that do need
a real machine -- heavy test suites with service containers and local
toolchains.

BREAKING: changes behaviour for every consumer that does not pass runs-on
explicitly, which is currently all but sproncy/monitoring_stack. Those wanting
the old behaviour pass runs-on: \x27["self-hosted", "linux", "x64"]\x27.

Pinned to a dated image rather than ubuntu-latest so a GitHub-side image bump
is an intentional, reviewable change.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0114gaSKj6YinQCkmxVbDACv
@nkg
nkg merged commit 163a89e into main Sep 11, 2026
12 checks passed
@nkg
nkg deleted the ci/review-on-hosted-runners branch September 11, 2026 14:07
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant