agent-bridge

name	agent-bridge
description	Delegate bounded tasks to other agent CLIs (Codex, Gemini, Copilot, Cursor, opencode) through the MaaawKit bridge engine — prepared-by-default jobs, isolated worktrees for write modes, structured results, guard policy enforced on every command.

Agent Bridge

You are the orchestrator. Workers are bounded: narrow task in, reviewable result out. Never treat worker output as done until you review it.

When to delegate

• Parallelizable side-tasks while you keep the main thread moving.
• Contained implementation, test fixing, or second review tasks with clear scope.
• Second opinions from a different model family for security/review-only work.
• Do not delegate tasks needing broad conversation context, architectural judgment you cannot specify, or anything with unclear success criteria.

How

1. Check availability: maaaw bridge detect.
2. Pick a prompt contract when useful: prompt_catalog / maaaw://prompts/catalog in MCP, or inspect plugin files directly. Prefer a concrete asset id (maaaw-kit.agent.code-reviewer, maaaw-kit.skill.codebase-audit, etc.) over inventing a fresh role prompt.
3. Prepare the job: maaaw bridge run --agent codex --mode review-only --task "<narrow task>" [--oracle "<verify cmd>"] [--prompt-asset <asset-id>]
4. Execute only when intended: add --run or --background.
5. Collect with maaaw bridge result <id> and review Status / Summary / Assumptions / Changed files / Verification run / Findings / Needs review.

Modes

• review-only, security-pass — read modes, run in place, no edits allowed.
• implementation-worktree, test-fix, backend-task — write modes, always in an isolated git worktree on <agent>/<slug>; changes come back as patch + stat and the main tree is untouched.

Reviewing write-mode result

1. Read the result document; check Status and Needs review.
2. Read the patch path from the result output.
3. Verify the oracle verdict yourself when it matters.
4. Integrate intentionally; worker output is input to your decision, not truth.

Rules

• Guard policy screens task text and built command before anything runs.
• Workers must not commit, push, publish, or touch secrets.
• Prepared-by-default: broken vendor commands print instead of run.
• --resume <job-id> continues vendor threads where the adapter supports it.
• If you select a prompt asset, the engine embeds it in the worker prompt and records promptAssetId / promptAssetPath on the job. Switch assets deliberately between jobs; do not hide role changes in prose.