@carter-mcalister/pi-rlm

RLM-first extension for Pi Coding Agent.

This package now exposes two distinct surfaces:

• rlm / /rlm — a true recursive language model surface with planner, worker, and synthesizer nodes
• rlm_task / /rlm-task — named reusable RLM workflows backed by Ax agents

What it is for

Use rlm for open-ended tasks that may benefit from recursive decomposition:

• repo architecture analysis
• risk reviews
• broad planning tasks
• multi-part investigations
• synthesis across several subproblems

Use rlm_task for repeatable rubric-driven workflows:

• incident reviews
• RFC / design review checks
• long-document analysis
• manager-ready summaries with evidence

Install

pi install npm:@carter-mcalister/pi-rlm

For local development in this monorepo:

pi install /absolute/path/to/pi-packages/packages/pi-rlm

Recursive Tool

The rlm tool and /rlm command support:

• op: "start" — start a recursive run
• op: "status" — inspect a run or recent runs
• op: "wait" — wait for a run to finish
• op: "cancel" — cancel a running job

Example:

rlm({
  task: "Analyze the reliability risks in this repo",
  mode: "auto",
  maxDepth: 2,
  maxNodes: 12,
  maxBranching: 3
})

Async example:

rlm({ task: "Review this architecture", async: true })
rlm({ op: "status" })
rlm({ op: "wait", id: "<run-id>" })

Command example:

/rlm "Analyze the reliability risks in this repo" --maxDepth 2
/rlm "Summarize this repo" --async
/rlm-list

Seeing What RLM Is Doing

Use /rlm-list to browse running and completed runs.

• It opens a native Pi selection list.
• Selecting a running run shows current status and recent activity.
• Selecting a completed run shows the final result and artifacts.

Both sync run output and completed run summaries include a Recent Activity section so you can see the planner and child-node progress without opening the artifact files.

Artifacts are written under:

/tmp/pi-rlm-runs/<runId>/

Named Workflow Surface

Slash commands

/rlm-task list
/rlm-task show incident-review
/rlm-task run incident-review --file docs/incidents/checkout.md --audience "engineering managers"
/rlm-task doctor

Tool calls

rlm_task({ task: "incident-review", inputs: { file: "docs/incidents/checkout.md", audience: "engineering managers" } })

Built-in workflows

• incident-review
• rfc-quality-check

Project-local workflows

Add local workflow definitions under:

.pi/rlm/tasks/*.ts

Design approach

This package is RLM-first:

• public surfaces talk about RLM workflows and recursive runs
• bundled workflows center agent(...) and RLM runtime behavior
• Ax is the implementation layer, not the package identity

The bundled workflows follow current Ax guidance:

• use agent(...) rather than manual new AxAgent(...)
• default to contextPolicy: { preset: 'checkpointed', budget: 'balanced' }
• use AxJSRuntime() deliberately for long-context, stateful analysis
• keep mode: 'simple' unless recursive delegated sub-agents are actually needed

Development

mise install
bun install
bun run --filter @carter-mcalister/pi-rlm check