@rol1n/subagent 0.1.0 → 0.1.1

package/README.md

@@ -32,4 +32,4 @@ Without a usable multiplexer, the worker backend falls back to a background proc
 
 ## Other delegation packages
 
-Use [multi-harness](https://github.com/RolinShmily/pi-kit/tree/main/extensions/multi-harness) for Codex, Claude, or OpenCode peers. That package shares multiplexer helpers but provides a separate tool and does not automatically enable this extension's entrypoint. The bundled subagent skill file is reference material and is not registered by the package manifest.
+Use [multi-harness](https://github.com/RolinShmily/pi-kit/tree/main/multi-harness) for Codex, Claude, or OpenCode peers. That package shares multiplexer helpers (`mux.ts`) while exposing a separate tool.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rol1n/subagent",
-  "version": "0.1.0",
+  "version": "0.1.1",
   "type": "module",
   "description": "subagent extension from pi-kit",
   "keywords": [
@@ -9,7 +9,7 @@
   "repository": {
     "type": "git",
     "url": "git+https://github.com/RolinShmily/pi-kit.git",
-    "directory": "extensions/subagent"
+    "directory": "subagent"
   },
   "homepage": "https://github.com/RolinShmily/pi-kit#readme",
   "bugs": {
@@ -21,10 +21,10 @@
   "publishConfig": {
     "access": "public"
   },
-  "files": [
-    "extensions/subagent",
-    "extensions/subagent/README.md"
-  ],
+  "exports": {
+    ".": "./index.ts",
+    "./mux": "./mux.ts"
+  },
   "peerDependencies": {
     "@earendil-works/pi-ai": "*",
     "@earendil-works/pi-agent-core": "*",
@@ -34,7 +34,7 @@
   },
   "pi": {
     "extensions": [
-      "./extensions/subagent/index.ts"
+      "./index.ts"
     ],
     "skills": [],
     "prompts": [],

package/extensions/subagent/README.md

@@ -1,35 +0,0 @@
-# subagent
-
-Delegate tasks to independent, full-powered Pi workers. The `subagent` tool returns after launch; completion or questions are delivered back to the parent session asynchronously.
-
-## Install
-
-```sh
-pi install npm:@rol1n/subagent
-```
-
-Also included in `@rol1n/pi-kit`. The parent and workers require the `pi` executable on PATH and their normal Pi authentication.
-
-## Use
-
-```json
-{"action":"spawn","name":"reviewer","cwd":"/existing/project","task":"Review the changed files and report findings."}
-{"action":"send","target":"reviewer","task":"Also inspect the error handling."}
-{"action":"stop","target":"reviewer"}
-```
-
-`model` optionally selects a worker model. `/subagent` shows active workers. Each worker inherits its own normal Pi capabilities; this is a delegation runtime, not a restricted execution environment.
-
-## Backends
-
-| Variable | Behavior |
-|---|---|
-| `PI_SUBAGENT_BACKEND=process` | Run silent background processes |
-| `PI_SUBAGENT_MUX=tmux` | Prefer tmux when running inside an available tmux session |
-| `PI_SUBAGENT_MUX=zellij` | Prefer Zellij when running inside an available Zellij session |
-
-Without a usable multiplexer, the worker backend falls back to a background process. Native Windows supports PowerShell launch commands and Zellij panes; tmux is for POSIX environments. Session transcripts and task artifacts are stored under `~/.pi/agent/`.
-
-## Other delegation packages
-
-Use [multi-harness](https://github.com/RolinShmily/pi-kit/tree/main/extensions/multi-harness) for Codex, Claude, or OpenCode peers. That package shares multiplexer helpers but provides a separate tool and does not automatically enable this extension's entrypoint. The bundled subagent skill file is reference material and is not registered by the package manifest.

package/extensions/subagent/skills/subagent/SKILL.md

@@ -1,70 +0,0 @@
----
-name: subagent
-description: Dispatch, message, and collaborate with parallel, full-powered Pi subagent workers running in split terminal panes or silent background processes.
----
-
-# Subagent
-
-Use the native `subagent` tool to delegate tasks to independent, full-powered Pi instances ("workers"). Workers run in dedicated split terminal panes (inside tmux or Zellij) or run silently as background processes (in standard terminals, VSCode, CI/CD, or headless environments). Every worker inherits all standard tools, skills, and extensions, and operates completely asynchronously.
-
-## Spawning Workers
-
-Call the `subagent` tool directly. The tool returns **immediately** with an acknowledgement. You do **not** need to wait or poll:
-
-```json
-subagent({
-  "task": "Investigate the authentication middleware in src/auth.ts and propose fixes.",
-  "name": "auth-audit"
-})
-```
-
-- `task`: Clear, comprehensive instructions for the worker.
-- `name`: Short, descriptive identifier (e.g. `frontend`, `db-migration`, `tester`).
-- `model`: (Optional) Model override (e.g. a faster model for simple exploration).
-- `cwd`: (Optional) Working directory for the worker.
-
-### Parallel Workers
-
-You can dispatch **multiple workers in parallel in a single turn**:
-
-```json
-[
-  { "name": "subagent", "arguments": { "task": "Audit frontend router in src/routes/", "name": "frontend" } },
-  { "name": "subagent", "arguments": { "task": "Audit database models in src/db/", "name": "backend" } }
-]
-```
-
-All workers run concurrently in separate split panes. As each worker finishes, the system automatically injects its final result into your conversation as a **steer message**, waking you up to review the findings.
-
-## Non-Blocking Principle
-
-- **DO NOT** write bash scripts, `sleep` commands, `tail` logs, or polling loops to check on a worker.
-- **DO NOT** fabricate or assume worker results before receiving the steer message.
-- After calling `subagent`, either end your turn immediately or work on other independent tasks. The harness will wake you with the result as soon as it is ready.
-
-## Answering Worker Questions
-
-Workers have an `ask_question` tool. When a worker encounters ambiguity or needs a decision:
-1. The worker pauses its work and marks itself as `waiting: ask`.
-2. You receive a steer notification: `[Subagent "<name>" question] ...`.
-3. Answer the worker by targeting its name:
-
-```json
-subagent({
-  "target": "auth-audit",
-  "task": "Use the JWT verification method rather than session cookies."
-})
-```
-
-The worker absorbs your guidance, resumes execution, and reports back when done.
-
-## Stopping a Worker
-
-If a worker is no longer needed or is off-track, terminate it:
-
-```json
-subagent({
-  "target": "auth-audit",
-  "action": "stop"
-})
-```