Agent collab

Agent Collab

Overview

Use Agent Collab to run multiple AI agents on one repository without shared working-directory edits, hidden state, duplicated work, or unreviewed merges. The repository is the coordination surface: task files, branches, worktrees, handoffs, reviews, test reports, ADRs, and status files must capture everything future agents need.

Core Rules

• Use a dedicated git worktree and branch for each agent role or independently owned task slice.
• Never edit protected branches such as main, master, production, staging, or release branches directly.
• Keep tasks small, scoped, reviewable, testable, and reversible.
• Write durable coordination state under .agent/; do not rely on private chat context.
• Require implementation handoff, review evidence, and verification evidence before merge readiness.
• Escalate destructive, production-affecting, security-sensitive, billing, auth, licensing, or ambiguous architectural decisions to the user.
• Do not claim checks passed unless they were run.
• Do not silently resolve overlapping agent work; document conflicts under .agent/risks/.

Standard Workflow

1. Initialize coordination files if .agent/ does not exist:

   agent-collab init
   

2. Create or read the task file before assigning agents:

   agent-collab new-task --id TASK-001 --title "API Client Refactor"
   

3. Decide whether work can run in parallel. Parallelize only when agents have disjoint file ownership or a documented coordination plan.

4. Create one worktree per agent from the protected base branch.

5. Give each agent a role, task file, branch, worktree, owned files, non-goals, required checks, and handoff expectation.

6. Require each implementation agent to inspect status, branch, task, existing patterns, and relevant handoffs before editing.

7. Require handoff after implementation, review after handoff, test report after review, and integration check before merge.

8. Ask the user before destructive, irreversible, production, security, auth, billing, or licensing changes.

Coordination Files

Use this repository structure:

.agent/
  README.md
  status.json
  tasks/
  handoffs/
  decisions/
  reviews/
  test-reports/
  risks/
  protocols/
  scratch/

Run commands from the target repository root after installing the package. When coordinating another repository from elsewhere, pass --root <repo-root> to the command.

Create durable artifacts with:

agent-collab handoff --task TASK-001-api-client-refactor --role implementer --branch agent/impl/TASK-001-api-client-refactor --worktree ../repo-implementer
agent-collab review --task TASK-001-api-client-refactor --branch agent/impl/TASK-001-api-client-refactor
agent-collab test-report --task TASK-001-api-client-refactor --branch agent/impl/TASK-001-api-client-refactor
agent-collab adr --number 1 --title "Use Server-Side API Wrapper"
agent-collab conflict --task TASK-001-api-client-refactor --title "API Client and Auth Changes"
agent-collab file-ownership --task TASK-001-api-client-refactor
agent-collab human-decision --task TASK-001-api-client-refactor --title "Choose Auth Boundary"
agent-collab merge-recommendation --task TASK-001-api-client-refactor

Read references/artifacts-and-templates.md for file names, required fields, source-of-truth priority, and complete template guidance.

Worktree Protocol

Create sibling worktrees from the base branch:

git fetch origin
git worktree add ../repo-implementer -b agent/impl/TASK-001-api-client-refactor origin/main
git worktree add ../repo-reviewer -b agent/review/TASK-001-api-client-refactor origin/main
git worktree add ../repo-tester -b agent/test/TASK-001-api-client-refactor origin/main
git worktree list

Branch naming:

agent/<role>/<task-id>-<short-description>
agent/impl-frontend/<task-id>-<short-description>
agent/impl-backend/<task-id>-<short-description>

Before editing, every agent must run:

git status
git branch --show-current
git log --oneline -5

Read references/worktree-and-branch-protocol.md before creating, sharing, rebasing, integrating, or removing worktrees.

Role Selection

Use the smallest role set that makes the task safe:

Role	Use for
Coordinator	Scoping, task flow, worktree assignment, handoff routing, merge recommendation
Architect	Solution design, boundaries, ADRs, split strategy, testing strategy
Implementer	Scoped code changes, tests, commits, implementation handoff
Reviewer	Diff review for correctness, architecture, security, reliability, maintainability, tests
Tester	Running checks, reproducing flows, smoke testing, test report
Documentation	README/docs/examples/migration notes aligned with implemented behavior
Integration	Conflict resolution, branch freshness, final checks, merge readiness

Read references/role-playbooks.md for responsibilities, forbidden actions, workflows, and role prompts.

Quality Gate

Implementation is not ready for merge until:

• Task acceptance criteria are satisfied.
• Scope is controlled and unrelated churn is absent.
• Implementation handoff exists.
• Review exists and has no unresolved blocking findings.
• Required checks were run or skipped with clear justification.
• Test report exists for non-trivial changes.
• Secrets and high-risk changes were checked.
• Rollback path is documented.
• Human approval exists where required.

Read references/quality-review-testing-merge.md for coding, review, testing, commit, PR, and merge rules.

Security And Escalation

Treat auth, authorization, identity, sessions, cookies, CSRF, CORS, billing, payments, webhooks, admin panels, database migrations, file uploads, multi-tenant data access, PII, logging of user data, deployment secrets, and infrastructure as high-risk.

If a secret is committed, stop and escalate. Do not merely delete it in a later commit and continue.

Read references/security-and-escalation.md for high-risk areas, human approval triggers, database/API/runtime rules, failure handling, and escalation templates.