Agents

The agent team that builds an app from botgumi. Instructions are written once and work on both runtimes.

File	What it is
AGENTS.md	The app's agent instructions. Copied to the app's root; its CLAUDE.md is only @AGENTS.md.
orchestrator.md	Plans, briefs, routes reviews, escalates, stops at human gates. Runs as the main session.
workers/<track>.md	One per track: backend, ios, android, landing, dashboard, design, docs. An app has only the platform tracks it chose.
reviewers/reviewer.md	Reviews a PR against review-rules.md and posts it with gh pr review.
review-rules.md	Who reviews, the fixed checks, blocking vs non-blocking, the 2-round limit, agents never merge.
model-policy.md	Default runtime and model per role, the tier rubric, the tokens-low fallback.
intake.md	How tasks arrive (issues, chat, botgumi task), how the runner hears about them, the trust filter.
merge-policy.md	When the orchestrator may merge, and the revert watch after a bot merge.
task-brief.md	The brief the orchestrator writes for each task, with tier, runtime and model.
runtimes/	How each runtime (claude, codex) loads the instructions and runs a role.

Runtimes and models per role are set in botgumi.config.ts (agents): any role can use any runtime and model. Role files carry no model, so the config stays the single source. The AgentRuntime interface is @botgumi/agent-runtime.

botgumi new <app> copies AGENTS.md to the app's root and the role files and docs into .claude/agents/.