[justinwlin]

Moves the PyPI Publish job in release-please.yml from Blacksmith to runpod-4vCPU-ubuntu-2404, the org's GitHub-hosted runner with static, allow-listed IPs. PyPI trusted publishing is built for GitHub-hosted runners, and Blacksmith counts as self-hosted. The label is literal because the org variable RUNNER_4VCPU_UBUNTU_2404 isn't shared with this repo. A probe job on this branch confirmed the runner accepts jobs from runpod/flash. The probe was then removed.

This doesn't fix the current Release failure. Since #399 the token step asks for workflows: write, and the release app (RELEASE_APP_ID) isn't granted it: The permissions requested are not granted to this installation. An org owner has to add Workflows: Read and write to the app and accept it on the installation. Then rerun Release on main. That publishes 1.21.1, which also needs runpod-workers/flash#112 released first, because of #390.

[OneAsterisk]

Claude review, round 1

Verdict: Approved — a one-line runner move that does what the description says; the two notes below are non-blocking.
Approval gate: 0 Critical, 0 High, 1 Medium+Low → approved.

PR description vs code

Claim	Status	Where
PyPI Publish moves from Blacksmith to runpod-4vCPU-ubuntu-2404	✅	release-please.yml:52
The label is the org's GitHub-hosted runner with static, allow-listed IPs	✅ GitHub-hosted: the probe log shows runner group runpod-4vcpu-ubuntu-2404 and image ubuntu-24.04 from actions/runner-images. ⚠️ Static-IP and allow-list config needs org admin to read (403 here); consistent with runpod/KaOS and runpod/spinnaker-platform calling it "the org's static-IP runner".	job 124146750304
Trusted publishing is built for GitHub-hosted runners; Blacksmith counts as self-hosted	⚠️ Mostly. pypa/gh-action-pypi-publish's README: self-hosted runners are "best effort", "Trusted Publishing is expected to work on self-hosted runners", and publishing from GitHub-provided runners is "the recommended/supported way". Supported path, not a hard requirement. See nit.	README, Non-goals
RUNNER_4VCPU_UBUNTU_2404 isn't shared with this repo	✅ repos/runpod/flash/actions/organization-variables returns 0 variables	API
A probe confirmed the runner accepts jobs from runpod/flash, then was removed	✅ Run 41080143161's sibling runner-probe on 204c67f succeeded on that runner group; 2507107 deletes runner-probe.yml; head has only ci.yml, e2e.yml, release-please.yml	run 41080143737
This doesn't fix the current Release failure (workflows: write not granted)	✅ The latest main run fails at the token step with exactly The permissions requested are not granted to this installation	run 40991928237
An org owner must grant Workflows: Read and write, then rerun Release	⚠️ Plausible; the app installation is not readable from here
1.21.1 needs runpod-workers/flash#112 released first, because of #390	✅ #390's warning names runpod-workers/flash#113 (merged 2026-10-27, on main since v1.6.0); #112 is the 1.6.1 release PR whose tag would include it. ⚠️ #112's changelog is dated 2026-10-23 and does not list #113, so release-please there has not regenerated it since.

Findings

Critical

None.

High

None.

Medium

None.

Low

• Literal runner label instead of the org variable — release-please.yml:52. The org's own tooling treats a written-out label as an error: runpod/spinnaker-platform's check-workflow-runners.sh rejects any runs-on that is not ${{ vars.RUNNER_4VCPU_UBUNTU_2404 }} ("named by an organization variable, so the label changes in one place"), and runpod/KaOS and runpod/internal-workflows use the variable. If the runner group is renamed, the next release's publish job waits for a runner that never arrives. The body already explains why the literal is needed today. The fix is the same ask as the app permission: have the org admin share RUNNER_4VCPU_UBUNTU_2404 with runpod/flash, then switch to the variable. Fine to merge as-is and follow up.

Nits

• release-please.yml:52 "trusted publishing needs that" overclaims; pypa says it is expected to work on self-hosted runners and GitHub-hosted is the supported path. Suggested wording inline.

Simplification opportunities

None; one-line change.

Verified clean

• Workflow-level id-token: write (line 11) still applies to the job, and PyPI's publisher match (repository, release-please.yml, environment pypi-production) does not depend on the runner, so the OIDC exchange is unaffected.
• actions/setup-python, astral-sh/setup-uv with enable-cache, and pypa/gh-action-pypi-publish are all in their supported configuration on a GitHub-hosted runner.
• Build Package is skipped by guard, not failed: compute_should_build only fires for packaging paths (ci.yml:118-126), and this PR touches none. Every other check passes at 2507107.
• Both commits use ci:, hidden from the changelog; the repo squash-merges.
• The other three jobs stay on Blacksmith, which already works under the org IP allow list (#397), so nothing else changes.
• Merging adds one more failing Release run on main until the app permission is granted; the body says so.

Questions for the author

• Can the org-admin request for the release app's Workflows permission also include sharing RUNNER_4VCPU_UBUNTU_2404 with runpod/flash? That closes the Low above.
• runpod-workers/flash#112 still describes main as of 2026-10-23 and omits #113. Is that repo's release automation also stalled? The tag would still include 95de97c9, but the changelog would not.

---

_{Reviewed by Claude (claude-fable-5-1) via the pr-review skill. Tokens: 290 in / 31,674 out / 653,971 cache read / 95,402 cache write = 781,337. Estimated cost: $3.66 at $10 / $50 / $0.25 / $20 per MTok (input / output / cache read / 1-hour cache write), as of posting. Round 1 of this PR. This is automated assistance; a human owns the final call.}

[OneAsterisk]

Claude review, round 2

Verdict: Approved — the only change since round 1 is the comment wording on line 52, applied from the round 1 suggestion; nothing new to raise.
Approval gate: 0 Critical, 0 High, 1 Medium+Low → approved. Round 1's approval at 2507107 stands; this round re-approves at ccbb51a.

Status of prior findings

#	Finding	Severity	Status	Evidence
R1-1	Literal runner label instead of ${{ vars.RUNNER_4VCPU_UBUNTU_2404 }}	Low	Not addressed (blocked on org admin, not the author)	release-please.yml:52 still literal at ccbb51a; organization-variables for runpod/flash still returns 0
R1-2	Comment claimed trusted publishing needs a GitHub-hosted runner	Nit	Fixed	ccbb51a applied the suggestion verbatim; thread resolved

Existing feedback

• No human or bot comments on this PR; the two threads are round 1's own, answered above.

PR description vs code

Re-checked the claims that could have moved since round 1; the rest are unchanged and still hold.

Claim	Status	Where
PyPI Publish runs on runpod-4vCPU-ubuntu-2404	✅	release-please.yml:52
The probe job was removed	✅ head has only ci.yml, e2e.yml, release-please.yml	.github/workflows/
RUNNER_4VCPU_UBUNTU_2404 isn't shared with this repo	✅ still 0 org variables visible	API
This doesn't fix the current Release failure	✅ the newest Release run on main (40991928237, 02:28Z today) still fails at the token step; nothing newer has run	run 40991928237

Findings

None new. The diff since round 1 is one comment string on one line.

Simplification opportunities

None.

Verified clean

• The runs-on label itself is byte-identical to the one the probe exercised, so the round 1 probe result still applies.
• The new comment is valid YAML (space before #, no quotes to break).
• ccbb51a's subject (Update .github/workflows/release-please.yml) has no conventional-commit type, so release-please ignores it in the squash body; nothing leaks into the changelog.
• CI at ccbb51a: every check passes (run 41087170098). Build Package stays skipped by guard (should_build only fires for packaging paths); pre-check, quality-gates, and nix run in full mode on every PR regardless, as ci.yml:205,227,324 gate them.

Questions for the author

• Same two from round 1, still open: can the org-admin ask bundle sharing RUNNER_4VCPU_UBUNTU_2404 with the Workflows permission? And is runpod-workers/flash#112's changelog going to pick up #113 before it ships?

---

_{Reviewed by Claude (claude-fable-5-1) via the pr-review skill. Tokens: 358 in / 11,206 out / 803,627 cache read / 62,849 cache write = 878,040. Estimated cost: $2.02 at $10 / $50 / $0.25 / $20 per MTok (input / output / cache read / 1-hour cache write), as of posting. Round 2 of this PR. This is automated assistance; a human owns the final call.}