agent-ci

name	agent-ci
description	Run this repo's GitHub Actions locally with Agent-CI before pushing CI-sensitive changes.
user-invocable	true
allowed-tools	Bash, Read, Edit
model	claude-haiku-4-5
context	fork
metadata	{"internal":true}

agent-ci

Run the repo's CI pipeline locally before pushing. CI was green before you started, so any failure the local run surfaces comes from your changes.

RedwoodJS wrote the upstream tool and skill (MIT, https://github.com/redwoodjs/agent-ci). The fleet pins @redwoodjs/agent-ci in the wheelhouse catalog and wires it as the ci:local package script (resolved via node_modules/.bin, never pnpm exec/npx). Edit only in socket-wheelhouse/template/; the cascade refreshes downstream copies.

Requirements

• Docker must be running — each job runs in a container. On macOS the fleet uses OrbStack (open -a OrbStack; recommended over Docker Desktop). If the daemon is down, agent-ci fails fast with couldn't use a Docker socket at /var/run/docker.sock … missing or a dangling symlink and exit 1 — that's the daemon, not a workflow failure. Start the provider, confirm with docker info, re-run. No daemon and can't start one → fall back to greening-ci (push + watch remote).
• The dep is already installed — @redwoodjs/agent-ci is a fleet devDependency (catalog:), provisioned by pnpm install.
• --github-token for remote reusable workflows — every socket-* repo's ci.yml calls a SocketDev/socket-registry/.github/workflows/… reusable workflow. agent-ci can't fetch it without a token; pass --github-token (no value → auto-resolves via gh auth token). Omitting it makes a remote-reusable CI silently fail to resolve.
• macOS jobs (runs-on: macos-*) run in a throwaway VM and need tart + sshpass on an Apple Silicon host (brew install cirruslabs/cli/tart hudochenkov/sshpass/sshpass). Without both, macOS jobs are skipped with a reason — the rest of the run still proceeds.

Run

The blessed entry is the canonical ci:local script — it already carries the full flag set (--all --quiet --pause-on-failure --github-token), and pnpm resolves the agent-ci binary from node_modules/.bin cross-platform:

pnpm run ci:local

--all runs the PR/push workflows for the current branch. --quiet suppresses the live renderer (pipe-safe). --pause-on-failure stops at the first failed step and holds the container open for retry. --github-token (bare → gh auth token) fetches the socket-registry reusable workflow every fleet ci.yml calls. Pipes are safe: when stdout is not a TTY the launcher detaches and the foreground process exits 77 the moment a step pauses, so | tee log and > log.txt work.

There is no --list or dry-run flag — run executes. Args after the subcommand pass through, so a typo'd flag becomes a workflow arg rather than an error.

To resolve the binary from a .mts script (not a package.json script — those resolve node_modules/.bin themselves), use the fleet helper, never a shelled-out which/command -v (which searches the global PATH and resolves the wrong binary — enforced by socket/no-which-for-local-bin):

import { whichSync } from "@socketsecurity/lib-stable/bin/which";

const agentCi = whichSync("agent-ci", {
  path: nodeModulesBinDir,
  nothrow: true,
});

Fix and retry

When a step fails the run pauses (and the run.paused event carries the exact retry_cmd to copy). Fix the code, then retry the paused runner — don't restart the whole pipeline:

node_modules/.bin/agent-ci retry --name <runner-name>

Call the linked binary directly, the fleet form for an ad-hoc bin invocation, same as node_modules/.bin/oxfmt / tsgo in build scripts — never pnpm exec/npx. Re-run from an earlier step with --from-step <N>. Repeat fix → retry until every job passes. Don't push to trigger remote CI when agent-ci can run it locally.

Reference

• Machine-readable --json event stream, the full requirements rationale, and the agent-ci-vs-remote-CI decision matrix: see reference.md.

Handoffs

Use greening-ci-local for the local fix-and-retry loop, and greening-ci when a real runner is required.