Skip to content

feat(blacksmith): migrate a repo's GitHub Actions compute to Blacksmith runners - #53

Merged
jtomaszewski merged 1 commit into
mainfrom
feat/blacksmith-skill
Sep 9, 2026
Merged

feat(blacksmith): migrate a repo's GitHub Actions compute to Blacksmith runners#53
jtomaszewski merged 1 commit into
mainfrom
feat/blacksmith-skill

Conversation

@jtomaszewski

@jtomaszewski jtomaszewski commented Sep 9, 2026

Copy link
Copy Markdown
Contributor

Problem

One FSH repo was moved to Blacksmith by hand. The rules that came out of it — which jobs to move, how to rewrite image builds for the sticky-disk cache, what to leave alone and why, how to measure without fooling yourself — lived in one PR thread. The next repos would re-derive them.

What this adds

/fsh:blacksmith — run inside any repo:

  1. Gate: the Blacksmith GitHub App must be installed on the org (without it blacksmith-* jobs queue until GitHub times them out).
  2. Inventory + baseline: scripts/inventory.sh prints one row per job with the classification signals and a suggested verdict; scripts/ci-timing.sh prices the last 14 days per workflow and runner label.
  3. Classify by one question: does it build, run or test the app, and take long enough to matter?
    • Migrate at the same vCPU: image builds (swap to useblacksmith/*@v2, constant cache-key, drop cache-from/cache-to), unit/lint/typecheck, service-container tests, browser e2e, CPU-bound bundle steps.
    • Stay: required check and if: always() gates, notifiers, PR commenters, paths-filter, deploys and health checks that wait on a cluster, self-hosted jobs, machine-local or network-scoped credentials.
    • Ask once: fused build+deploy jobs (recommend splitting); test jobs holding cloud-federation creds (default migrate, listed so a client-repo owner can veto).
  4. Rewrite, actionlint with the bundled config template, timeout-minutes on every migrated job, then /deliver --no-merge with the decision table and baseline in the PR body.

references/policy.md carries the measurements the policy rests on (hardware −53%, sticky disk a further −33% on one image build; n=1 misled twice).

Verified

  • shellcheck + bash -n clean on both scripts.
  • Inventory on fsh-monorepo: 2 image builds + 3 test/build jobs → migrate; every deploy, reaper and notifier → stay; 3 callers resolved to their callees.
  • Inventory on the already-migrated repo's main: every moved job reports DONE, the required-check aggregator reports STAY, and a self-triggering workflow idle for 63 days is flagged as the one whose edit will wake a pre-existing failure.
  • ci-timing.sh against fsh-monorepo (5 runs/workflow) returns priced rows.
  • The actionlint template takes label errors on a 57-workflow repo from 36 to 0.

plugin.json 1.6.0 → 1.7.0 (new skill).

🤖 Generated with Claude Code


View with [code]smith Autofix with [code]smith
Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.

…th runners

One FSH repo was moved to Blacksmith by hand and the rules that came out of it were
worth keeping: move only what builds, runs or tests the app, at the same vCPU count;
give image builds the sticky-disk Docker actions and drop the external layer cache
(measured 12.4 → 3.6 min on one build, hardware −53% and the disk a further −33%);
leave glue, deploy, self-hosted and credential-bound jobs where they are, because
faster cores buy them nothing and the merge gate should sit on the most boring
infrastructure available.

The skill gates on the GitHub App being installed (without it blacksmith-* jobs
queue forever), inventories every job with a suggested verdict it must still judge,
captures a priced timing baseline before editing, and hands off to deliver --no-merge
with the decision table in the PR. It refuses pilot scaffolding (runner inputs,
label-guarded dual step paths) — the PR is the rollback unit — and treats a
self-triggering workflow that has been dormant as something to expect a pre-existing
failure from, not to fix.

Bundled: inventory.sh (PyYAML, read-only, dormancy check via gh), ci-timing.sh
(per-workflow wall-clock and cost at both vendors' rates), an actionlint.yaml
template so the labels don't read as typos, and references/policy.md with the
measurements the policy rests on.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@jtomaszewski
jtomaszewski merged commit fb2816f into main Sep 9, 2026
2 checks passed
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