action-runner
スキルの推奨を実際の実行行動に変える――チケットを開き、問題を提出し、更新を投稿する――安全に:予行プレビュー、リスク分類、承認ゲート、そして脳に記録する。
取り込み時のスキャン結果 · 2026-08-22
- 指示の混入利用者に代わって承認を先取りする記述があります
SKILL.md:55
ルールに基づく静的スキャンの結果です。検出がないことは安全を保証するものではありません。 本文と同梱スクリプトは全文を閲覧できるため、実行前に内容をご確認ください。
Action Runner Skill
The library is great at recommending work. This skill executes it — the action layer of the Professional Brain (Phase 2). A skill says "open a ticket per checklist item"; this turns that into real GitHub/Linear/Slack actions, safely: previewed, risk-rated, approved, then recorded. The cardinal rule: nothing acts silently.
What This Skill Produces
- A dry-run actions plan — every proposed action with its target, operation, and risk.
- After approval, the executed actions (via the connected action MCP) — outbound/destructive ones gated individually.
- A record back to the brain of what was actually done, with provenance.
Required Inputs
Ask for (if not already provided):
- The recommendations to act on (a launch checklist, PRD requirements, postmortem follow-ups…).
- The connected action MCP and targets — which GitHub repo / Linear project / Slack channel. Scope is limited to what the user names; never act outside it.
- Approval posture — what may run with a single OK vs. what needs per-action confirmation.
How it works
recommend → build an actions plan (JSON) → preview + risk-gate → approve → execute → record
- Build the plan — express each action as JSON:
{"target","op","args","why","risk?"}. - Preview + gate — run the helper; it prints a dry-run, classifies risk (🟢 low / 🟡 medium /
🔴 high), and refuses to proceed while any 🔴 outbound/destructive action is unapproved:
echo '<plan json>' | python3 scripts/action_preview.py - # after the user approves the risky ones: echo '<plan json>' | python3 scripts/action_preview.py - --allow-high - Approve — low/medium can run on a single confirmation; every 🔴 (post, send, delete, deploy, merge, charge…) needs explicit per-action approval. Default is do nothing until told.
- Execute — only approved actions, only via the connected action MCP (e.g. Composio/GitHub
create_issue). One target at a time; stop and report on the first failure. - Record — append what was actually done to the brain so the loop closes:
python3 ../professional-brain/scripts/brain_write.py ./brain decisions "Filed launch tickets" \ --tag external --body "Opened 7 issues in acme/app from the launch checklist" --commit
Supported action targets
Any action MCP can be wired in; these are the common targets, with example operations and the default risk the gate applies. Reads are 🟢; anything outbound, destructive, or that spends is 🔴.
| Target | Example operations | Default risk |
|---|---|---|
| GitHub | create_issue, comment, open_pr · (merge_pr, close 🔴) | 🟡 (🔴 for merge/close) |
| Linear / Jira | create_issue, update_status, comment | 🟡 |
| Slack | post_message, reply_in_thread (outbound → always confirm) | 🔴 |
| Notion | append_block, create_page, update_property | 🟡 (🔴 if it overwrites) |
| Email / Gmail | send_email (outbound) | 🔴 |
| Calendar | create_event, invite (outbound) | 🟡 (🔴 if it emails invitees) |
Pick the narrowest target and op that does the job, scope to exactly what the user named, and let the risk gate decide what needs explicit approval. Outbound messages (Slack/email) are 🔴 by default — the model never posts on someone's behalf without a per-action yes.
Safety rules (non-negotiable)
- Dry-run by default. The plan is shown before anything runs.
- Approval-gated. No execution without a yes; 🔴 actions are confirmed one by one.
- Scope-limited. Only the repos/channels/projects the user named.
- Logged. Every executed action is recorded to the brain with an
[external]tag and a link. - No silent retries, no bulk outbound. If a step fails, stop and surface it.
The contract for other skills
An action-aware skill adds a short "Proposes Actions" section: after producing its artifact,
it lists the actions it could take (target · op · why), then hands off to action-runner —
which previews, gates, executes, and records. The skill never executes directly.
Output Format
- Proposed actions — a table: # · target · operation · why · risk.
- Gate result — the preview output; the 🔴 actions needing approval called out explicitly.
- Executed (after approval) — what ran, with links/IDs returned by the MCP.
- Recorded to the brain — the line(s) appended, with provenance.
Quality Checks
- A dry-run plan is shown before anything executes
- Every action has a risk level; 🔴 actions are individually approved
- Execution stays within the named scope and uses only the connected MCP
- Each executed action is recorded back to the brain with an
[external]tag - On failure, it stops and reports rather than retrying blindly
Anti-Patterns
- Executing anything without showing the dry-run plan first
- Treating an outbound/destructive action (post, email, delete, deploy) as low-risk
- Acting outside the scope the user named, or fanning out to many targets
- "Helpfully" doing more than was approved
- Forgetting to record what was done — the brain must reflect reality