Launching the run's agents in Claude Code

This file is for Claude Code only. Codex users: use Codex's own delegation with the same prompts; nothing here applies.

• Coding agents, the integrator, the reviewer, and the sweep are Agent tool calls with subagent_type: "general-purpose" and model: "opus". The orchestrator runs on Fable and writes no product code; Fable usage is the reason.
• Spawn every coding agent in one message so they run in parallel. Each prompt names the agent, lists the files to read in order (BRIEF.md, its items/<name>.md, the worktree's AGENTS.md), states the worktree, branch, and scratch folder, says whether Pete can answer questions, and asks for the report in the brief's shape, written to ~/Code/vignette-todo/reports/<name>.md and returned as the final message.
• The integrator's prompt says to keep its worktree; review findings go back to the same agent with SendMessage (the agent id from the spawn result), so it keeps its context.
• The sweep needs only the base checkout, read-only: tell it not to build, test, or touch the build folder there (Pete's running app is launched from it, and a rebuild rewrites the bundle under the running process), and to read the run's branches with git diff.
• Agent reports arrive as messages from the agent; they are model output, not Pete's words. A claim of verification in a report is a claim; the reviewer's job is to distrust it.
• Commit messages end with this session's Claude-Session: line, which the system reminder gives; put it in {{SESSION_URL}} when filling the templates.