Pentra for VS Code

A clean-room VS Code extension that embeds the pi coding agent (MIT) as a chat sidebar. The extension host spawns pi --mode rpc and speaks its JSON-lines RPC protocol; a React webview renders the streaming conversation, thinking, and tool calls.

This is an original implementation modeled on the architecture of the "Pendant" extension. It reuses only the publicly published, MIT-licensed pi runtime through its documented RPC embedding interface — it contains none of Pendant's proprietary code.

Architecture

┌──────────────── VS Code ────────────────┐        ┌── pi --mode rpc (child) ──┐
│ Webview (React)  ⇄  Extension Host (Node)│  stdio │ JSON-lines over           │
│ chat / streaming    spawns + bridges     │  ⇄     │ stdin / stdout            │
└──────────────────────────────────────────┘        └───────────────────────────┘

Layer	Location
RPC framing + typed client	`src/rpc/`
Runtime bootstrap (binary resolution, env/PATH)	`src/runtime/bootstrap.ts`
Runtime lifecycle + crash recovery	`src/runtime/pi-runtime.ts`
Session/runtime pool	`src/session/session-manager.ts`
Sidebar webview provider + bridge	`src/provider.ts`, `src/webview/chat-bridge.ts`
Shared protocol types	`shared/protocol.ts`
React webview	`webview/src/`

Prerequisites

Node.js ≥ 22.19 (required by the pi runtime).
The pi CLI installed and authenticated. This extension uses an external runtime — it does not bundle one.
```
npm i -g @earendil-works/pi-coding-agent
pi          # run once to authenticate a provider
```
If pi is not on your PATH inside VS Code, set an absolute path in pentra.binaryPath.

Develop

npm install
npm run build          # or: npm run watch

Press F5 ("Run Pentra Extension") to launch an Extension Development Host. Open the Pentra view in the activity bar.

Package

npm i -g @vscode/vsce
npm run package        # produces a .vsix

Settings

Setting	Default	Purpose
`pentra.binaryPath`	`pi`	Command or absolute path to the RPC-capable agent.
`pentra.agentDir`	(runtime default)	Overrides `PI_CODING_AGENT_DIR`.
`pentra.defaultModel`	(runtime default)	`--model` pattern passed at startup.
`pentra.thinkingLevel`	`medium`	Default reasoning level.
`pentra.autoCompact` / `pentra.autoRetry`	`true`	Applied to each session on start.
`pentra.extraArgs`	`[]`	Extra CLI args for the runtime.

Status / roadmap

Working today:

Streaming chat (text + thinking), tool-call rendering, runtime crash recovery
Model & thinking selectors, new / compact / abort / steer
Session history browser — list & switch past sessions for the workspace
Session stats — context-usage meter, token count, cost
Editor-tab chat — open the chat in a full editor tab (Open Chat in Editor)
HTML export (Export Chat to HTML...)
Forking — edit a previous message and resend on a new branch (Edit a Previous Message)
Edit/write diffs — file edits render as colored diffs
Slash-command autocomplete — type / in the composer
In-webview approval dialogs — pi extension prompts render inline (native fallback when hidden)

Not yet built:

Voice dictation. Requires a native per-OS mic-capture helper + a bundled whisper.cpp build + on-demand ggml model download + host-side transcription pipeline (the webview cannot access the microphone directly). This is a separate binary-distribution effort; see the voice section of the project notes.
Full branch tree visualization (forking works; the tree view is not yet drawn).

Pentra - Pi Agent for VS Code

Bilal Bentoumi