Orca Cook Plan: run an AgentKit plan through Orca, one fresh worker session per phase

# Orca Cook Plan: run an AgentKit plan through Orca, one fresh worker session per phase Source: https://modernengineer.dev/en/library/orca-cook-plan > **In short:** Orca Cook Plan runs an AgentKit plan you already have through Orca. The coordinator orders phases by dependency, gives each phase to a fresh worker session running `ak-cook`, asks you when a decision is needed, checks the evidence and checkpoints. The guide is free to read; buying the files gets you the skill (version 1.2.3) for Claude Code, Codex, Cursor, Grok CLI, Pi or OMP, plus the overview page and an interactive diagram of how a run works. Note: the current AgentKit beta (since 3.0.0-beta.35, October 21, 2026) changed skill names from `ak:<name>` to `ak-<name>`, and the commands in this post use the new names: Claude Code calls `/ak-plan`, Codex calls `$ak-plan`. On an older version, swap `ak-` back to `ak:` to get the matching command. Running a whole multi-phase plan in one session tends to go wrong in the same three ways: workers ask about things outside the scope, the coordinator loses track of progress after compaction, and every run starts by rediscovering the project's tools. Orca Cook Plan splits the work: the coordinator only coordinates, each phase runs in its own worker session, and the plan stays the anchor. ## What the skill fixes - **Stays in scope.** With `--auto`, the coordinator decides within the plan you gave it. Proposals to widen the scope are recorded and never block the running phase. - **Remembers pending work.** A checkpoint keeps the Run, Task/Dispatch, decisions, evidence and next step. After compaction or a fallback, the coordinator recovers in the same session and Run and checks Orca so it never dispatches twice or reruns an accepted phase. - **Uses the tools you have.** Workers get the project's entrypoint, commands, working directory and checks. A fix that was verified once is reused by later workers. - **Picks runtime and model.** `--select-agent` asks once per Run, finds the CLIs in the worker's environment and checks the integration before running. The guide comes with the full overview page (the “Overview” tab) and an interactive diagram of the coordinator, worker and advisor (the “Step by step” tab). The diagram labels are in Vietnamese. ## What you get - The `orca-cook-plan` 1.2.3 skill files: `SKILL.md`, `README.md`, references (`recovery`, `runtimes`, `selection`, `omp`), the runtime discovery script, the OMP Context Guard extension and tests. - The overview page with the flags, Context Guard and the skill's changelog. - The interactive diagram of the three roles. The files are for your own use; please do not share them. Subscribers download the files with their plan without buying them separately, and can install them over MCP so their agent picks up new versions. ## Three roles  *The coordinator owns the Run, a worker does one phase, the advisor only challenges.* | Role | Does | Does not | | --- | --- | --- | | Coordinator | Owns the Run: orders phases, dispatches through Orca, handles questions, checks evidence, accepts results, updates the plan and saves the checkpoint. Commits in Git, or records file-scope evidence when the workspace has no Git. | Implement, or re-review all the code when the report and evidence are enough. | | Worker | Does one phase in its own session: implements, tests and reviews per cook, then reports `worker_done`. Prefers the project's tools and asks through Orca when a decision is needed. | Stage, commit or start the next phase. A repair uses a new worker session. | | Advisor | Challenges the approach and suggests changes. | Own the Run, dispatch or commit. | `worker_done` does not mean the phase passed. The coordinator checks the evidence, accepts the phase and writes the checkpoint before dispatching anything that depends on it. A phase with no changes is recorded instead of producing an empty commit. ## Before you install 1. **Orca.** The coordinator uses Orca to create Tasks, open worker sessions and track progress. 2. **AgentKit Engineer Kit** in the worker's runtime, since workers run `ak-cook`. For OMP: `ak kit init engineer --target omp --yes`. Get AgentKit through [my referral link](https://agentkit.best/?ref=OMG49S8R) for 30% off. 3. **The runtimes** you plan to use: OMP if you want Context Guard, or Claude Code, Codex, Cursor, Grok CLI, Pi. The coordinator and workers can use different runtimes. 4. **Python 3**, and **Node.js** for the OMP extension. ## Install from the file pack Subscribers who connected the [modernEngineer MCP server](https://modernengineer.dev/en/mcp) can skip the manual steps: paste the prompt from the file pack section into your agent, which downloads, installs and updates the skill when a new version is out. Download the file pack at the end of this guide and unzip it. The skill folder is `orca-cook-plan/orca-cook-plan/` inside the zip. Copy **the whole folder**, including `extensions/` and `references/`, into your runtime's skills folder: | Runtime | Per project | Global | | --- | --- | --- | | Claude Code | `.claude/skills/` | `~/.claude/skills/` | | Codex | `.agents/skills/` | `~/.agents/skills/` | | OMP | `.omp/skills/` | `~/.omp/agent/skills/` | | Pi | `.pi/skills/` | `~/.pi/agent/skills/` | | Cursor | `.cursor/skills/` | `~/.cursor/skills/` | | Grok CLI | `.grok/skills/` | `~/.grok/skills/` | A global install for Claude Code: ```bash unzip orca-cook-plan-v1.2.3.zip mkdir -p ~/.claude/skills test ! -e ~/.claude/skills/orca-cook-plan && cp -R orca-cook-plan/orca-cook-plan ~/.claude/skills/ ``` If the target already has `orca-cook-plan`, stop and diff before replacing it. Using `orca-cook-plan` across several projects or worktrees with Claude Code or Codex? [Let your agent install and update AgentKit and skill packs through MCP](https://modernengineer.dev/en/library/agent-control-plane-mcp-first) shows how to have your agent keep one master copy, link it into each project and update it when a new version is out. ### One more step on OMP Copying the skill folder does not load Context Guard by itself. Add a loader file in the same scope: | Scope | Skill | Loader | | --- | --- | --- | | Project | `.omp/skills/orca-cook-plan/` | `.omp/extensions/subscriber-orca-cook-plan.js` | | Global | `~/.omp/agent/skills/orca-cook-plan/` | `~/.omp/agent/extensions/subscriber-orca-cook-plan.js` | The loader contains: ```js export { default } from "../skills/orca-cook-plan/extensions/context-guard.mjs"; ``` Open a new OMP session and call the skill. The extension starts idle; the coordinator configures and activates it when a Run starts. Check that it loaded with `/orca-cook-context-status`. Do not load a second copy through `--extension`. If you remove the skill, remove the loader too, or OMP fails on the import. ## Calling it Open a new session in the project and call the skill with the real plan path (absolute or relative, optionally prefixed with `@`): | Runtime | Syntax | | --- | --- | | Claude Code, Cursor, Grok CLI | `/orca-cook-plan plan.md --auto` | | Pi, OMP | `/skill:orca-cook-plan plan.md --auto` | | Codex | `$orca-cook-plan plan.md --auto` | Some combinations (OMP syntax): ```text /skill:orca-cook-plan plan.md --advice --auto /skill:orca-cook-plan plan.md --runtime omp --omp-advisor all --advice --auto /skill:orca-cook-plan plan.md --runtime omp --omp-advisor phase-02,phase-04 --auto --tdd /skill:orca-cook-plan plan.md --select-agent --parallel --auto ``` Without flags, workers use the current runtime and cook's default mode, and decisions that need approval go to you through the coordinator. ## Flags There are two layers: orchestration flags handled by Orca Cook Plan, and flags passed down to `ak-cook` for each phase. These are skill-invocation conventions, not a CLI parser; the coordinator checks them against the real cook in the worker's runtime before dispatching. **Orchestration flags** | Flag | Meaning | | --- | --- | | `--runtime <agent-id>` | The worker's runtime; the coordinator's runtime does not change. Defaults to the current one. Accepts `--runtime omp`, `--runtime=omp`, `--runtime=current`. | | `--select-agent` | Asks for the runtime and primary model before the first dispatch and keeps the choice for the Run and recovery. `--auto` still waits for this answer. | | `--auto` | The coordinator decides in-plan approvals. Workers do not get cook's `--auto` and still ask through Orca. It never widens scope, skips access checks, publishes or deploys. | | `--omp-advisor` | OMP workers only: `all`, `none`, or phases such as `phase-02,phase-04` (their repairs included). | | `--parallel` | Prefers running independent phases in parallel. Dependencies and write conflicts still hold; at most two workers by default. | **Flags passed to `ak-cook`** | Flag | Meaning | | --- | --- | | `--advice` | AgentKit's advisory workflow (kongming) for every phase and repair. Independent of `--omp-advisor`; both can be on. | | `--tdd` | Tests for current behavior first, checked again after implementing. | | `--interactive` | The full workflow with review gates; questions still go through the coordinator. | | `--fast` | Skips research: scout, plan, code. It does not skip review or tests, or widen the phase. | | `--no-test` | Skips cook's test step. The report must say what was not verified. | | `--yagni` | Lets cook consider cutting scope the outcome does not need, never a plan acceptance criterion. | | `--skip-journal` | Skips the automatic journal entry at finalize. | ## Context Guard (OMP only) The guard is an extension on OMP sessions. While active, it reminds the session of its role, Run and checkpoint; asks for a checkpoint when telemetry reaches the context threshold; asks for recovery after compaction or a fallback; and blocks every tool except reading and writing the checkpoint when the session runs a model outside its role's `allowedModels`. It never changes models, saves the checkpoint itself or wakes a stopped runtime. Claude Code, Codex, Cursor, Grok CLI and Pi still get the workflow, the checkpoint and the recovery guide, without a hook that watches the context. If the coordinator runs Codex and the workers run OMP, only the OMP workers have the guard. ## Common questions **Does the coordinator continue if I answer late?** The skill stores the question with the pending work. When you answer, or the session recovers, the coordinator applies that decision and continues. It does not wake a runtime that has exited. **Does compaction or a fallback start a new coordinator?** No. The coordinator re-reads the checkpoint, checks the Run, Task, Dispatch, pending questions and evidence, and continues in the same session and Run. If it cannot recover, it reports the blocker and the checkpoint path. **When can a worker be replaced?** Only when recovery inside the worker's session cannot continue. The new worker finishes the remaining acceptance work of the same Task; it does not redo the phase. **Why is the guard idle in OMP?** Loaded is not active. Idle is normal before a Run starts or after it ends. `unknown` telemetry means not enough data, not plenty of context left. ## Known limits The skill has been checked for logic, an OMP smoke test and some Orca lifecycles. I have not run a full cook end to end on all six runtimes, on native Windows, or with a real exhausted quota or full context. Model fallback is managed by your runtime and its configuration. Never read silence as done.

unzip orca-cook-plan-v1.2.3.zip
mkdir -p ~/.claude/skills
test ! -e ~/.claude/skills/orca-cook-plan &&
  cp -R orca-cook-plan/orca-cook-plan ~/.claude/skills/

export { default } from "../skills/orca-cook-plan/extensions/context-guard.mjs";

/skill:orca-cook-plan plan.md --advice --auto
/skill:orca-cook-plan plan.md --runtime omp --omp-advisor all --advice --auto
/skill:orca-cook-plan plan.md --runtime omp --omp-advisor phase-02,phase-04 --auto --tdd
/skill:orca-cook-plan plan.md --select-agent --parallel --auto

Read "Orca Cook Plan: run an AgentKit plan through Orca, one fresh worker session per phase" on modernEngineer and apply it to this repo.
1. Call the content_get tool of the modernengineer MCP server with locale "en", slug "orca-cook-plan". If MCP is not connected or access is refused, read the page https://modernengineer.dev/en/library/orca-cook-plan instead.
2. Summarize the main steps of the piece and point out which apply to this repo.
3. Propose concrete changes and ask me before editing any file.

claude mcp add --transport http modernengineer \
  https://api.modernengineer.dev/mcp \
  --header "Authorization: Bearer <API key>"