action-run

name	action-run
description	Trigger and monitor any GitHub Actions workflow through the GitHub MCP server's actions toolset (no gh CLI). Use when the user asks to run, trigger, or execute a GitHub Action (e.g. "/action-run pypi publish", "/action-run generate migration"). Uses the github MCP server bundled with this plugin (requires GH_TOKEN).

Action Run Skill

Trigger any GitHub Actions workflow_dispatch workflow by name, gather required inputs, monitor it to completion, and fix failures automatically — driven through the GitHub MCP server's actions toolset (mcp__plugin_devpowers_github__actions_*, mcp__plugin_devpowers_github__get_job_logs), exactly the way /check-pr drives PR/CI status through mcp__plugin_devpowers_github__pull_request_read. No gh CLI.

Invocation Format

/action-run <workflow description> [key=value] [flag] ... [branch=<name>]

Examples:

/action-run hello
/action-run hello name=Mark shout
/action-run pypi publish tag=v0.0.8 pre-release
/action-run pypi publish tag=v0.0.8 branch=main
/action-run generate migration message="add users table"

Step 0 — Ensure the GitHub MCP actions toolset is available

This skill drives everything through the actions toolset of the github MCP server bundled with this plugin — its .mcp.json loads as mcp__plugin_devpowers_github__*, and the X-MCP-Toolsets: all header turns the actions tools on.

• If mcp__plugin_devpowers_github__get_job_logs (and the mcp__plugin_devpowers_github__actions_* tools) are visible in your tool list, proceed.
• If they're missing, the plugin's bundled github MCP didn't connect — almost always because GH_TOKEN isn't set (it authenticates with Bearer ${GH_TOKEN}). Tell the user to export GH_TOKEN=<pat> (a long-lived PAT) and restart Claude Code (or run /reload-plugins), then re-invoke /action-run. Stop until the server loads.

Triggering workflow_dispatch needs a token with Actions: write (classic workflow scope / fine-grained Actions: write). A read-only token can list and inspect runs, but run_workflow will fail with 403.

Tool shape: the hosted server groups Actions operations into method-based tools — mcp__plugin_devpowers_github__actions_run_trigger (methods run_workflow, rerun_failed_jobs, rerun_workflow_run, cancel_workflow_run), mcp__plugin_devpowers_github__actions_list (list_workflows, list_workflow_runs, list_workflow_jobs), mcp__plugin_devpowers_github__actions_get (get_workflow_run), and the standalone mcp__plugin_devpowers_github__get_job_logs. Server versions may rename these or shift methods; map the intent below to whatever mcp__plugin_devpowers_github__* actions tools your session actually exposes, and read each tool's registered schema for exact filter/pagination parameters.

Step 1 — Resolve repo and branch

git remote get-url origin
git rev-parse --abbrev-ref HEAD

Parse owner/repo from the remote URL. Accept github.com[:/]<owner>/<repo>(.git) and proxied /git/<owner>/<repo> forms. If parsing fails, surface the remote URL and stop — don't guess.

The default ref is the current branch. A branch=<name> argument overrides it — it controls which branch the workflow runs on, it is not a workflow input.

Step 2 — Identify the workflow file

ls .github/workflows/

Match the user's description to a workflow file (fuzzy match by name/content). When ambiguous, ask the user to clarify. The actions tools accept the workflow file name (e.g. hello.yaml) as workflow_id.

Step 3 — Discover required inputs

Read the matched workflow file (Read tool). Find the on.workflow_dispatch.inputs section. For each input:

• required: true — must be provided before running
• required: false / no required field — optional, has a default

Step 4 — Gather missing inputs

Check which required inputs were not supplied in the invocation. Ask the user for any that are missing.

Inputs already provided become entries in the inputs object (send values as strings — GitHub coerces booleans/numbers):

• tag=v0.0.8 → "tag": "v0.0.8"
• pre-release (bare flag) → "pre-release": "true"

Step 5 — Check for an already-running workflow

Don't re-trigger if one is already in flight. Call mcp__plugin_devpowers_github__actions_list with method list_workflow_runs, owner, repo, and resource_id: "<workflow>" (the file name); use the tool's run filter to narrow to active runs (status in queued / in_progress / waiting). If any active run comes back, take the newest one's id and skip to Step 7 (monitor it) instead of triggering. Otherwise continue.

Step 6 — Trigger the workflow

Call mcp__plugin_devpowers_github__actions_run_trigger with method run_workflow:

• owner, repo
• workflow_id: "<workflow>" (file name)
• ref: "<branch>"
• inputs: { ... } (omit when the workflow takes none)

Errors:

• 403 / "Resource not accessible" — the token lacks Actions: write (see Step 0).
• 404 / 422 — bad ref (branch not pushed?) or an input that doesn't match the workflow's declared inputs. Re-check Steps 1 and 3.

Then resolve the new run id — it takes a few seconds to register, so retry a couple of times: call mcp__plugin_devpowers_github__actions_list method list_workflow_runs (owner, repo, resource_id: "<workflow>") filtered to branch=<branch> and event=workflow_dispatch, newest first, and read the top run's id.

Step 7 — Monitor the run to completion

Snapshot the run: call mcp__plugin_devpowers_github__actions_get method get_workflow_run (owner, repo, resource_id: "<run_id>"). If status == "completed", skip to Step 8.

If it's still running, arm a Monitor that emits on status transitions and end your turn. A Monitor runs a shell command and can't invoke MCP tools mid-loop, so it polls the REST run endpoint with the same $GH_TOKEN the MCP server authenticates with — the one shell touchpoint, identical to /check-pr's CI watch. Confirm $GH_TOKEN is set first; if not, surface that (it's the same token the MCP uses).

prev=""
while true; do
  s=$(curl -fsSL \
    -H "Authorization: Bearer $GH_TOKEN" \
    -H "Accept: application/vnd.github+json" \
    "https://api.github.com/repos/<owner>/<repo>/actions/runs/<run_id>" || true)
  cur=$(jq -r '"\(.status) \(.conclusion // "-")"' <<<"$s")
  [ "$cur" != "$prev" ] && echo "$cur"
  prev=$cur
  jq -e '.status=="completed"' <<<"$s" >/dev/null && break
  sleep 20
done

The emitted line carries status and conclusion, covering the success and failure terminal states. Print a one-line "watching run #<run_id> via Monitor — will report when it completes" status and end your turn. When the notification arrives, re-call mcp__plugin_devpowers_github__actions_get (get_workflow_run) for the final conclusion and act per Step 8. If the user cancels mid-watch, use TaskStop with the Monitor's task ID.

Step 8 — Report result and handle failures

On conclusion == "success" — report the run URL (https://github.com/<owner>/<repo>/actions/runs/<run_id>) and stop.

On failure (failure / cancelled / timed_out) — fetch the failed job logs through MCP. Call mcp__plugin_devpowers_github__get_job_logs with:

• owner, repo
• run_id: <run_id>
• failed_only: true
• return_content: true
• tail_lines: 50 (adjust as needed)

This returns the tail of every failed job's log directly — no curl. Present a tight summary (failed step name + the key error lines), not the full dump.

Then:

1. Fix the root cause (failing tests, bad inputs, version conflicts).
2. Commit and push fixes if needed.
3. Re-run this skill with the same or corrected inputs.

Only ask the user for input when the fix genuinely requires a decision (e.g. choosing a new version number). Otherwise fix and retry autonomously. Do NOT loop polling on a failure — stop after reporting and let the user direct.

Notes

• Every GitHub Actions operation — trigger, run lookup, status, logs — goes through the MCP actions toolset (actions_run_trigger / actions_list / actions_get / get_job_logs). The only shell touchpoint is the Step 7 Monitor, which polls REST with the shared $GH_TOKEN because a Monitor can't call MCP tools mid-loop — the same arrangement /check-pr uses for its CI watch.
• The Monitor is session-bound. If the session ends before the run completes, re-run /action-run — Step 5 detects the active run and resumes monitoring.