Code Orchestrator
Code Orchestrator is two Agent Skills for repository work that needs several phases. One writes an implementation-ready phase plan from a direction you already chose. The other carries out an existing plan: separate workers implement each phase, independent reviewers check it, and one final integration brings the phases together.
Try it
Ask:
Turn this approved architecture into an implementation-ready phase plan.
write-phase-plan returns ordered phases, each with an owner, the files it
may change, acceptance criteria, dependencies, validation, and who integrates
it. It stops after the plan; it does not start implementation or reopen the
architecture decision.
How it works
approved direction → phase plan → implementation → independent review → phase join → final integration
| Skill | Result |
|---|---|
write-phase-plan | A dependency-ordered plan with enough context that workers without the conversation can carry it out |
phase-orchestrator | Reviewed phase delivery with a report from each worker, repairs, status, and one final integration result |
Worker roles live under phase-orchestrator/references/workers/. They are
reference files, not skills an agent can discover, so the implementer and
reviewer roles never compete with the parent skill for a request.
Interfaces
The two skills are skills/write-phase-plan/SKILL.md and
skills/phase-orchestrator/SKILL.md. Host manifests expose the same files in
Codex, Cursor, Copilot, and Agent Plugins hosts.
Checks
skills/write-phase-plan/references/plan-contract.mddefines the plan format and when phases can join.skills/phase-orchestrator/references/orchestration-protocol.mddefines execution, review, convergence, and final integration.- Separate implementer, reviewer, and final-reviewer references keep review independent of implementation.
- The repository’s routing fixtures separate phased delivery from ordinary single-owner work, open architecture questions, and host scheduling.
Attribution and pinned upstream sources are recorded in
PROVENANCE.md.
Limits
- Use phased delivery only when the implementation direction is already chosen and the work splits into phases that can be delivered separately.
- Do not fan out a one-line fix or other ordinary single-owner task.
- Settle open architecture choices before writing or running the plan.
- The plugin plans the work and assigns each part one owner. It never replaces
the target repository’s
AGENTS.md, compute scheduler, validation, CI, review, merge, release, or deployment policy.
Questions
Do I need a phase plan before execution?
Yes. Use an existing implementation-ready plan or create one with
write-phase-plan; phase-orchestrator does not make up a plan of its own.
Are implementer and reviewer separate skills?
No. They are worker instructions inside phase-orchestrator, so an agent
choosing a skill sees exactly two.
Does the orchestrator manage machine capacity?
No. It leaves scheduling to the host and gives each heavy command or external wait one owner.
Start
Read the plan contract or the orchestration protocol, then install
code-orchestrator through one of the paths in the
Hraness Agent Skills README.