locus-pi — Dynamic Workflows for Pi

Describe a task. Let an agent build a reusable workflow. Run it in Pi.

locus-pi adds a workflow DSL to Pi. Workflows can run agents in parallel, discover more work, and choose the next step from their results. Save the workflow as a readable JavaScript file and run it again when you need it. Model roles let you change models without rewriting the workflow.

flowchart LR
    describe["Describe<br/>a task in plain words"] --> generate["Generate<br/>an agent writes the workflow"]
    generate --> review["Review<br/>the saved workflow source"]
    review --> run["Run in Pi<br/>again, whenever"]

    classDef step fill:#F4EFE4,stroke:#2D4A6E,color:#1A1F2E,stroke-width:2px
    classDef result fill:#EED4CA,stroke:#C0482E,color:#8F3621,stroke-width:2px
    class describe,generate,review step
    class run result

Change models through roles; reuse the same workflow source.

Documentation · DSL reference · Examples

What it adds

Three core capabilities, shipped as Pi extensions.

Model assignments are yours to configure; no provider or model assignments ship with the package.

All six extensions

See the extension guide for commands and tools. Install the extensions together, or choose the resources you need with package filters.

Install

Requires Node.js >=22.19.0, Pi >=0.84.3, and a configured model provider.

From Git:

git clone https://github.com/locus-forge/locus-pi.git
cd locus-pi
npm ci --ignore-scripts
pi install .

From npm:

pi install npm:@locus-forge/locus-pi

The npm release can lag this repository. Use the documentation bundled with your installed version when checking its available workflows and skills.

Windows: use WSL 2 and install Node.js, Pi, and locus-pi inside its Linux environment. Follow Windows setup.

Start a fresh Pi session in your project. If locus-pi is already installed, replace its existing source; keep your other packages. See installation and updates.

Edit only the locus-pi entry in ~/.pi/agent/settings.json or .pi/settings.json. For Git, keep its checkout path as source; the example below uses npm.

{
  "packages": [
    {
      "source": "npm:@locus-forge/locus-pi",
      "extensions": ["extensions/workflows/index.ts"],
      "skills": []
    }
  ]
}

Keep other settings and package entries. skills: [] disables package skills, including the workflow creator; omit that field to retain them. The filter controls what Pi loads. See package filters.

Create a workflow with an agent

Use the ordinary workflow-create lesson in Pi. For an explicitly detailed walkthrough, choose the authoring route. It designs the agent graph, reviews it, builds the source, and checks it:

/skill:locus-pi-workflow-create Create an evaluator-optimizer workflow for the Task in .tasks/example/task.md: implement, independently review, and allow one correction followed by fresh review. Record verified working context and exact result/review paths in an orchestration folder. Build and check the workflow, but do not run it yet.

Review evaluator-optimizer.design.md and evaluator-optimizer.workflow.mjs saved under .locus-pi/workflows/evaluator-optimizer/, then run:

/workflows run evaluator-optimizer -- <original Task and verified working context>
/ps
/workflows result last

Create your first workflow explains the skill and includes complete copyable source. Run and inspect covers progress, results, stopping, fresh runs, and replay. For Codex or Claude Code, see skill installation.

The skill can author a workflow directly; it does not require running a packaged authoring workflow. For a staged authoring process, the task workflows provide task/draft, followed by task/plan or task/plan-light with the complete accepted draft as input. These ship under examples/workflows/task/ and appear in the Package tab of /workflows list; task itself is a group, not a runnable workflow.

Explore the DSL and examples

A workflow connects calls such as agent(), parallel(), and pipeline(). Use agent(..., { choice: [...] }) for a decision; richer results use exact caller-assigned file destinations in prompts, and later agents read those same files. Caller-owned work units come from items(). A stage may request modelRole: "smol"; assign the role through /model-roles independently of the source.

For example, this workflow gathers two perspectives before combining them:

export const meta = { name: "project-summary", profile: "standard" };

export default async function run({ agent, parallel }) {
  const notes = await parallel([
    () =>
      agent("Read README.md. Explain the project. Do not modify files.", {
        label: "purpose",
        title: "Read project purpose",
      }),
    () =>
      agent("Read package.json. Explain the commands. Do not modify files.", {
        label: "commands",
        title: "Read development commands",
      }),
  ]);
  return agent(`Combine these notes into a getting-started guide. Do not modify files.\n${notes.join("\n\n")}`, {
    label: "summary",
    title: "Write getting-started guide",
  });
}

flowchart LR
    subgraph explore["parallel()"]
        purpose["purpose<br/>README.md"]
        commands["commands<br/>package.json"]
    end
    purpose --> summary["summary<br/>getting-started guide"]
    commands --> summary

    classDef step fill:#F4EFE4,stroke:#2D4A6E,color:#1A1F2E,stroke-width:2px
    classDef result fill:#EED4CA,stroke:#C0482E,color:#8F3621,stroke-width:2px
    class purpose,commands step
    class summary result
    style explore fill:#FAF6EC,stroke:#2D4A6E,stroke-dasharray:5 5,color:#2D4A6E

Save it as .locus-pi/workflows/project-summary/project-summary.workflow.mjs and check the source before running it.

• DSL reference — every method, arguments, return values, examples, and availability.
• Workflow file format — metadata, input, and the exported function.
• Source rules — what the creator may generate and what the checker rejects.
• Examples — installed workflows under examples/workflows/ and patterns to adapt.
• Documentation — the entry point for installation, authoring, operation, and deeper topics.

Repository layout

extensions/              extension implementations and co-located manuals
extensions/_shared/      shared host, operator, runtime, model, and agent-runtime layers
extensions/workflows/    workflow runtime
examples/workflows/      installed reusable workflows and their guides
skills/                  bundled workflow authoring, execution, and external-session skills
scripts/                 repository validation and catalog generation
tests/                   focused and integration tests
docs/                    cross-cutting public guides and workflow reference

Project workflows live under .locus-pi/workflows/. Runs write local state under .locus-pi/, including outputs, workspaces, and leases. It may contain prompts, model output, or transcripts, and is ignored by Git. See architecture and repository boundaries.

License and security

Licensed under the MIT License. Run workflows from sources you trust; see the workflow trust guide. Report vulnerabilities through GitHub private vulnerability reporting; use Issues for ordinary defects.