| | --- |
| | name: gso-babysit-external-agents |
| | description: Hand Gunship Origins implementation, research, or planning to a cheap in-process babysitter (Sonnet in Claude Code, Terra in Codex) that drives external Jam Session agents (Muse, Cursor Grok, Grok, Devin, and others) from a written plan, so the interactive session keeps its context for Jamon. Use when Jamon asks for a babysitter, cheap agent babysitter, or external agents to do work while he keeps talking with the lead. Do not use for a short task the lead can finish directly, the night-shift queue, or a single summoned agent. |
| | --- |
| | # Babysit External Agents |
| | The interactive session is the supervisor in $jamsession-orchestrate-agent-work; the babysitter is its manager; Jam Session agents are workers. This skill adds the Gunship shape. Use concise plain English at every level and require it of every agent. |
| | ## Lead: before launch |
| | 1. Record Jamon's decisions and the authorized scope in the task worksheet first. The lead owns that worksheet's decision log; the babysitter appends only to its progress or evidence section. |
| | 2. Check the repository state the workers will inherit: uncommitted human edits, partial renames, known breakage, an open Godot editor. Put what you find in the plan. |
| | 3. Write .agents/scratch/YYYY-MM/<task>/PLAN.md (or BRIEF.md for research or planning). It must stand alone: |
| | - Authorized scope, plus an explicit list of what is not authorized. |
| | - Inherited state, and what workers must preserve (Jamon's uncommitted edits, files never to touch). |
| | - Ordered steps. Each step lists exact files, the change, the order and behavior checks to report back, done checks, and the local commit message. |
| | - The provider assignment table: implementer and a reviewer from a different provider for each step. Use Jamon's named providers; otherwise choose with $jamsession-model-recommendations. |
| | - The text every worker prompt must carry, and the text every reviewer prompt must carry. |
| | - Validation commands: lint, project load, docs lint, focused tests, and 2P pairs as $gso-test-change selects. Name the known inherited failures. |
| | - Commit authority (local commits on the current branch, never push, unless Jamon says otherwise) and the REPORT.md path. |
| | 4. Launch one background babysitter: Claude Code uses the Agent tool with model: sonnet; Codex uses a Terra subagent. Its prompt points at the plan and the required skills, and asks for a short final summary. Give it nothing the plan does not already hold. |
| | ## Babysitter rules (put them in its prompt) |
| | - Run jamsession status and jamsession models <provider> first. Record an unavailable provider in REPORT.md; never substitute one silently. |
| | - Write shared worker prompt text once (common.md, reviewer-common.md) and add each step's spec to it. |
| | - Run one write-capable worker per checkout at a time. Read-only research or review may run in parallel. Run each long jamsession run in the background, with output written to the scratch folder, and wait for its completion notice. |
| | - Verify every worker claim with its own diff, grep, and gate runs before accepting the step. Make tiny verified fixes itself; send larger ones back to the same worker session. |
| | - Treat HEAD moving, or files changing, as Jamon working in parallel. Never revert, restore, reset, or use bare stash. |
| | - Keep REPORT.md current after each step: sessions, commits, gates, review findings and how they were handled, and questions for Jamon. |
| | ## Lead: while it runs |
| | - Read REPORT.md and git log for status. Never read the babysitter's transcript file. |
| | - To add or change scope, record the decision in the worksheet, then message the running babysitter (SendMessage in Claude Code). Do not start a second writer in the same checkout. |
| | - A second babysitter may run at the same time only when it is read-only, or owns separate files or a separate workspace. |
| | - If the harness flags a babysitter (for example a security warning), inspect its prompts, worker output, REPORT.md, and git log before relaying anything to Jamon. |
| | - Relay results in a few lines: commits, gates, findings, and the questions only Jamon can answer. |