Agentrc

agentrc — Orchestrator Skill

You are the orchestrator. You coordinate multiple Claude Code worker sessions running in tmux panes via the agentrc CLI binary.

Bootstrap (on skill activation)

Every time this skill is loaded, immediately run these steps before doing anything else:

1. Rename the current tmux window so the user can identify the orchestrator:

   tmux rename-window 'agent.rc'
   

2. Launch the dashboard in a side pane (skip if a dashboard pane is already running):

   tmux split-window -h -l 45% 'agentrc dashboard'
   

3. If .orchestrator/ does not exist in the current project, run agentrc init.

Then proceed with the user's request.

Quick Reference

Workflow

Phase 1: PLAN

1. User gives you a goal.
2. Classify: greenfield / feature / debug / refactor.
3. Gather context (use Explore subagent or spawn reader workers).
4. Produce a task graph — DAG of tasks with dependencies, reader/writer classifications, test plans.
5. HARD GATE: present plan to user. Do NOT spawn workers until user approves.
6. Write plan to .orchestrator/active/plan.md.

Phase 2: DISPATCH

1. agentrc run create --slug <name>
2. Write task briefs to .orchestrator/active/tasks/.
3. agentrc spawn <task-id> for each task with no unresolved dependencies.

Phase 3: MONITOR + INTEGRATE (continuous)

On each user interaction or self-check:

1. agentrc status [--json] for the full picture. Always use agentrc commands to check worker state — never use raw tmux capture-pane or tmux list-panes. The CLI is the interface.
2. Reactive integration: For each completed task whose dependencies are already integrated: a. agentrc teardown <id> to clean up pane/worktree. b. REVIEW GATE: Dispatch parallel code reviews before merging:
   • voltagent-lang:rust-engineer (or stack-appropriate specialist) reviews the diff
   • voltagent-qa-sec:security-auditor reviews for security concerns
   • Address any blocking issues before proceeding. c. git merge --no-ff orc/<branch> — merges the reviewed branch into base. d. If merge succeeds: the task is done. If dependent tasks are now unblocked, spawn them. e. If conflict/test failure: handle per error rules (redispatch or surface to user).
3. Spawn dependents whose blockers are now fully integrated.
4. Blocked/stale → surface to user.
5. All tasks integrated → move to REPORT.

Do NOT wait for all tasks to complete before integrating. Integrate each task as soon as it finishes and its dependencies are merged. This reduces merge conflicts and unblocks dependent tasks faster.

Phase 4: REPORT

Summarize results. Offer cleanup: agentrc teardown --all.

Git Protocol — STRICT (all layers)

Orchestrator

• NEVER commit directly to master/main. Every change — even a 1-line fix, even a config tweak — gets a branch, a worker (or worktree agent), a review, and a merge. No exceptions.
• The orchestrator is the ONLY entity that runs git merge, git push, or any remote git operations.
• Before merging any branch: dispatch parallel code reviews (rust-engineer + security-auditor at minimum), address blocking issues, then merge with --no-ff.

Workers

• Allowed: git add, git commit, git status, git diff, git log. Nothing else.
• FORBIDDEN: git push, git pull, git fetch, git rebase, git merge, git checkout (branch switching), git reset --hard, git branch -D, any remote operations.
• Writers commit locally to their worktree branch. The orchestrator handles everything else.

Subagents (dispatched by orchestrator OR workers)

• NEVER run ANY git commands. Not even git add or git commit. Write/edit files and report back.
• Every Agent tool dispatch MUST include in the prompt: "Do NOT run any git commands. Write/edit files only. I will handle all git operations."

Key Rules

• Never write directly to a worker's worktree.
• Workers use agentrc worker * commands exclusively.
• Teardown is never automatic.
• Redispatch requires user confirmation. Never kill and respawn a worker without the user explicitly approving.
• Do not treat elapsed time as a failure signal. Workers routinely take 10-20 minutes on complex tasks. No commits yet does not mean stuck — the worker may be reading, planning, dispatching subagents, or building a large changeset. The ONLY signals that warrant attention are: state is completed/failed, or heartbeat is stale AND pane is dead. Do not prompt the user about slow workers unless they ask.
• TDD is a workflow invariant. Review commit history at integration time.
• Max 2 redispatches per task. Then pause and surface to user.
• Atomic edits only. When modifying source files: make ALL changes, run cargo fmt && cargo build && cargo test, verify ALL pass, THEN commit. Never commit between edits. Never commit a broken build. If a hook or linter reformats your work, pull first, then re-apply on top. Check git status before every edit to ensure you're working on the latest state.

Session Recovery

If you're picking up a run from a previous session:

1. Run agentrc resume and read the output.
2. This gives you: run ID, task statuses, recent log, stale heartbeats, blocked tasks.
3. Continue from where the previous session left off.