@jmcombs/pi-relay

Relay roles for the Pi coding agent: run any Pi subagent on an external coding agent instead of a local model — just by setting its model. Relay registers pi providers (relay-claude, relay-grok, relay-cursor); a subagent whose model is relay-claude/opus, relay-grok/grok-4.5, relay-cursor/cursor, relay-cursor/opus-4.8, or relay-cursor/opus-5.5 routes through relay to a headless Claude (claude -p), Grok Build (grok -p), or Cursor Agent (cursor-agent -p), which runs its own tool loop and returns the final result.

Not affiliated with or endorsed by Anthropic, xAI, or Anysphere. Claude and Opus are trademarks of Anthropic, PBC; Grok is a trademark of xAI; Cursor is a trademark of Anysphere, Inc.

Fixes

• oh-my-pi system prompts are normalized. oh-my-pi supplies systemPrompt as a string[]; relay joins those sections before building the backend prompt.
• oh-my-pi tasks terminate through local yield. When the host context exposes oh-my-pi's task-only yield tool, relay returns the backend text plus one synthetic terminal yield call. oh-my-pi therefore finalizes after one external CLI run instead of issuing text-only reminders and repeating the request.

How It Works

A relay role is an existing pi-subagent (its persona .md + referenced SKILL.mds). Nothing about the subagent changes except the processor:

• Trigger + model — set a subagent's model to relay-claude/opus, relay-claude/opus-5.5, relay-grok/grok-4.5, relay-cursor/cursor, relay-cursor/opus-4.8, or relay-cursor/opus-5.5. pi's native resolveModel routes the completion to relay's registered provider → claudeDriver / grokDriver / cursorDriver → claude -p … --model opus|claude-opus-5-5 --effort <level> / grok -p … --model grok-4.5 --reasoning-effort <level> / cursor-agent -p … --model composer-2.5|claude-opus-4-8-thinking-high|claude-opus-5-5-medium (Pi thinking selects Claude/Grok effort flags, or the listed Cursor id).
• Persona + skills — when pi runs a subagent it assembles the persona body + a skill injection into the (child) session's system prompt, where skills are <available_skills> references (name/description/location). Relay reads each referenced SKILL.md and inlines its full content into the prompt it writes via the backend's own system-prompt mechanism (claude's --system-prompt-file, grok's inline --system-prompt-override/--rules, or prepended onto Cursor's user prompt — Cursor has no system-prompt flag), so the methodology is guaranteed present (deterministic — no model re-echo, no drift).
• Tools — each driver maps the subagent's pi tools onto its backend's own permission model (read → Read, bash → Bash, edit → Edit, write → Write, grep → Grep, find → Glob); pi-only tools with no external equivalent (e.g. subagent, ls) are dropped. Claude gets --allowedTools; Grok gets one --allow <Tool> flag per tool plus --permission-mode dontAsk (fail-closed — unlisted tools are silently declined, never a hang or a blanket bypass). The map is a driver function (D10). Cursor has no --allowedTools argv flag; cursorDriver maps the same pi names onto allow/deny rules (Read(**/*), Shell(*), Write(**/*)), copies the user's Cursor config home into a temp dir, overlays those rules on cli-config.json, and points CURSOR_CONFIG_DIR at it. A sparse temp dir (only the generated config) hangs cursor-agent -p on the first tool call. --force / --yolo are never passed.
• Single external run — the relayed subagent has no external equivalent for pi/oh-my-pi orchestration tools. The external agent runs its own tool loop once and returns final text. pi consumes that text directly; oh-my-pi receives the same text plus a local terminal yield call generated by relay.

The flagship consumer is phase verification: the verifier subagent runs as a relayed subagent (model: relay-claude/opus, read-only tools) — no bespoke tool, no inline prompt.

Backend

The verify quality bar is Claude Opus only (D1), reached through the subscription claude -p CLI (billed to your Claude subscription via oauthAccount — never the Anthropic API, never a local model). The verify role is read-only: claude is invoked with a scoped --allowedTools allowlist and never with --dangerously-skip-permissions. On a cut run (idle-cap or abort) relay surfaces an UNVERIFIED error result — it never auto-passes.

relay-grok (Grok Build, grok -p) is a second live driver available for generic subagent dispatch — it does not change the verify quality bar. Per D1, a new backend only becomes verify-eligible after it clears the accuracy benchmark; until then, route the verifier role to relay-claude/opus and use relay-grok for other subagents. Grok is invoked with --permission-mode dontAsk plus one --allow <Tool> per allowed tool (verified fail-closed and non-interactive — never --always-approve or --permission-mode auto/bypassPermissions).

relay-cursor (Cursor Agent, cursor-agent -p) is a third live driver. Cursor is invoked with --output-format stream-json and --trust (skip the workspace-trust prompt). --force / --yolo (Cursor's permission bypass) and --sandbox are never passed. Pi model → Cursor --model map:

Pi id	Cursor --model (thinking off / default)	Notes
relay-cursor/auto	auto	Thinking does not change the id
relay-cursor/cursor	composer-2.5	Cursor's own Composer 2.5 model
relay-cursor/opus-4.8	claude-opus-4-8-high	Renamed from opus; :high → …-thinking-high
relay-cursor/opus-5.5	claude-opus-5-5-medium	Effort baked into the listed id; :high → …-high

The mapper strips the relay-cursor/ provider prefix first; an id that does not resolve to a listed id is rejected up front rather than forwarded. Legacy relay-cursor/opus still resolves in the driver to the Opus 4.8 High family but is no longer in the provider catalog. Cursor has no system-prompt flag, so persona + skills are prepended onto the user prompt. Tool scoping overlays allow/deny rules on a seeded temp CURSOR_CONFIG_DIR (copy of the user's Cursor config home, not a lone cli-config.json — that hangs headless tool calls until the wall-cap; see #254). No-tools roles still seed that dir and fail-closed deny Write(**/*) so Cursor cannot fall through to ~/.cursor.

Claude Code and Cursor name the same Anthropic models differently:

Goal	Claude Code (claude -p)	Cursor Agent (cursor-agent -p)
Opus 4.8	--model opus or --model claude-opus-4-8 + optional --effort	Listed id claude-opus-4-8-high / claude-opus-4-8-thinking-<level>
Opus 5.5 Medium	--model claude-opus-5-5 --effort medium	Listed id claude-opus-5-5-medium
Composer 2.5	n/a	--model composer-2.5

Confirm Cursor account availability with an authenticated cursor-agent --list-models (catalog is server-driven). Claude Code 2.1.280+ accepts claude-opus-5-5; the rolling opus alias remains Opus 4.8 on the Anthropic API per Claude Code model-config docs.

Claude and Grok keep --model as the alias / pinned id and apply Pi thinking as --effort / --reasoning-effort. Relay catalogs these models with reasoning: true so Pi's thinking UI matches what the drivers send.

A driver/adapter seam (AgentDriver in drivers/claude.ts) keeps the provider backend-agnostic. claudeDriver, grokDriver, and cursorDriver are the live implementations, each owning its own pi→backend tool-name map (D10); drivers/codex.ts is a documented seam-only stub (codex exec, -s read-only) for a future OpenAI Codex backend. roles/resolver.ts is backend-neutral: it inlines skill references to full content (expandSkillReferences) and resolves a persona+skills role from disk (used off the pi-subagents path). The provider streams the completion through pi's own createAssistantMessageEventStream() (@earendil-works/pi-ai).

Requirements

• Pi (loads the extension via jiti — no build step)
• Node >= 22.19.0
• The claude CLI on PATH, authenticated via your Claude subscription (oauthAccount), for relay-claude
• The grok (Grok Build) CLI on PATH, authenticated (grok login or XAI_API_KEY), for relay-grok
• The cursor-agent CLI on PATH, authenticated (cursor-agent login or CURSOR_API_KEY), for relay-cursor

Configuration

• PI_RELAY_WALL_MS — idle-cap backstop for a single relayed run, in milliseconds (default 600000). Any stdout or stderr byte resets the timer; true silence still SIGTERMs the child. On a cut run relay reports an UNVERIFIED error result.
• PI_RELAY_HEARTBEAT_MS — fallback no-op stream beat while a backend emits nothing (json print mode). Stream backends forward real text_delta / tool-progress instead. Default 20000; set 0 to disable. Each beat surfaces as a pi message_update, advancing the parent's activity clock; the verdict still rides only on the terminal result, so the beats never affect it.

Install

# Globally (recommended)
pi install npm:@jmcombs/pi-relay

# For a single session, without installing
pi -e npm:@jmcombs/pi-relay

# Oh My Pi
omp plugin install @jmcombs/pi-relay

See the Pi packages documentation for git, local path, project-scoped install, and filtering options.

Usage

Relay registers the relay-claude, relay-grok, and relay-cursor providers; you use any of them by pointing a subagent (or a whole session) at it through model:

# Route a whole session through the relay provider
pi --model relay-claude/opus "…"
pi --model relay-claude/opus-5.5 "…"
pi --model relay-grok/grok-4.5 "…"
pi --model relay-cursor/auto "…"
pi --model relay-cursor/cursor "…"
pi --model relay-cursor/opus-4.8 "…"
pi --model relay-cursor/opus-5.5 "…"

To run an existing subagent through relay, set its model frontmatter to relay-claude/opus, relay-claude/opus-5.5, relay-grok/grok-4.5, relay-cursor/cursor, relay-cursor/opus-4.8, or relay-cursor/opus-5.5 and make relay discoverable in the subagent's child pi (an installed package, or the agent's extensions field).

oh-my-pi discovers custom task agents from ~/.omp/agent/agents/*.md and .omp/agents/*.md. Set the agent's model normally:

---
name: relay-reviewer
description: Review a change through Claude Code.
model: relay-claude/opus
tools: read, grep
---

oh-my-pi keeps yield local; relay never forwards it to Claude, Grok, or Cursor.

Extending — adding a driver

Relay is backend-agnostic through the AgentDriver seam (D10). claudeDriver, grokDriver, and cursorDriver are the live implementations; drivers/codex.ts is a documented seam-only stub for a future OpenAI Codex backend. To add a driver for another coding agent (Codex, Gemini CLI, …) — the AgentDriver API, the pi→backend tool-name mapping, the read-only/fail-safe constraints, and a step-by-step guide — see CONTRIBUTING.md in this package.

Development

This package lives in the pi-extensions monorepo. See the repo-root CONTRIBUTING.md for project conventions, and this package's CONTRIBUTING.md for the driver seam.

# From the repo root
npm ci
npm run check                       # full quality gate
node packages/relay/scripts/harness.mjs   # manual provider proof vs. real `claude -p`
node packages/relay/scripts/harness.mjs --model relay-grok/grok-4.5   # same, vs. real `grok -p`
node packages/relay/scripts/harness.mjs --model relay-cursor/auto     # same, vs. real `cursor-agent -p`

# Unreleased worktree only (--no-extensions -e ./packages/relay):
./packages/relay/scripts/prove-thinking-map.sh catalog   # list-models vs expected context/max/thinking
./packages/relay/scripts/prove-thinking-map.sh argv      # Pi thinking → real CLI flags (six -p runs)

License

MIT © Jeremy Combs