Skip to content

an open-source alternative to the dots, bots and muses of the world, run an office of claude code/codex like agents on your laptop, sandboxes or anywhere, uses your existing subscriptions

release(v0.3.6): version bump, changelog, release notes

Chaitanya Giri7ccbd90e5255
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 638a021b..b9035954 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,46 @@ All notable changes to this project are documented here. The format is based on
[Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to
[Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+## [0.3.6] — 2026-08-08
+
+**A machine with nothing installed on it can now run agents.**
+Every fix in this release is about the same failure: the app assumed the user's machine
+already had things on it — a shell to expand `~`, a `node` on PATH, an npm to install
+with — and when it didn't, agents died with a bare exit code and no explanation. Plus a
+long-standing office bug where the floor went blank and never came back.
+
+### Fixed
+- **`~/dev/foo` no longer fails with "cwd does not exist".** Only a shell expands `~`;
+ Node treats it as a literal directory, so a typed `~/…` path failed every existence
+ check and the agent never spawned. `~` is now expanded once at ingestion, so the
+ registry only ever stores an absolute cwd — which also repairs existing `~` entries
+ that had been reading as "not absolute" forever.
+- **Hooks stopped silently dying with exit 127.** Agent CLIs run hooks through `sh -c`
+ with a bare `PATH=/usr/bin:/bin:/usr/sbin:/sbin` — nvm's node isn't there, so every
+ hook payload was lost: no live status, no Stop→inbox drain, no session ids. Every hook
+ shim now runs under the Node the app already bundles.
+- **`node` is on every agent's PATH.** An MCP server declared as `node ./server.js`, a
+ provider CLI that shells out, a `.cjs` an agent wrote itself — all died with 127 on a
+ machine with no system node. The bundled runtime is now *appended* to the agent's PATH,
+ never prepended, so anyone with their own node keeps their own version.
+- **The office floor comes back after losing its GPU context.** Chromium caps live WebGL
+ contexts and silently evicts the oldest — always the office, since it starts first,
+ once enough agent terminals are open. Pixi reported nothing, so the floor simply went
+ blank until you restarted the app. It now detects the loss and rebuilds the scene.
+- **God no longer messages agents that no longer exist.** A live roster of the floor is
+ pushed into the orchestrator's context on session start and on every prompt, instead of
+ relying on it to re-read `fleet.json` — which is why it went stale across restarts.
+
+### Added
+- **Node and npm install themselves when they're missing.** Selecting an engine on a
+ machine with no Node used to print `npm install -g …` and run it anyway, so the user
+ watched `npm: command not found` scroll past. The app now fetches the latest Node LTS
+ straight from nodejs.org, verifies it against the official `SHASUMS256.txt` before
+ anything executes, installs it visibly in that agent's terminal, and then installs the
+ CLI. If you already have Node 20 or newer, it is left completely alone.
+- **An honest dead end instead of a doomed command.** When no installer can succeed, the
+ banner names the missing piece and runs nothing at all.
+
## [0.3.5] — 2026-08-06
**The queue always has an escape hatch, and the app updates itself for the first time.**
diff --git a/RELEASE.md b/RELEASE.md
index b0ce994e..6cc93a2b 100644
--- a/RELEASE.md
+++ b/RELEASE.md
@@ -1,4 +1,4 @@
-# Munder Difflin v0.3.5
+# Munder Difflin v0.3.6
**A local hive of Claude Code, Antigravity, Codex, Grok & Copilot agents that run themselves** — messaging,
routing, and remembering, coordinated by a GOD orchestrator you talk to. Local-first and open source.
@@ -7,7 +7,45 @@ routing, and remembering, coordinated by a GOD orchestrator you talk to. Local-f
---
-## What's new in 0.3.5 — *The queue always has an escape hatch*
+## What's new in 0.3.6 — *A machine with nothing on it can run agents*
+
+Every fix here is the same failure wearing different clothes: the app assumed your machine
+already had things on it — a shell to expand `~`, a `node` on PATH, an npm to install with
+— and when it didn't, agents died with a bare exit code and no explanation. Plus one
+long-standing office bug where the floor went blank and never came back.
+
+- **Node and npm install themselves.** Pick an engine on a machine with no Node and the app
+ now fetches the latest Node LTS straight from nodejs.org, **verifies it against the
+ official `SHASUMS256.txt` before anything runs**, installs it visibly in that agent's own
+ terminal, and then installs the CLI. Already have Node 20 or newer? It's left completely
+ alone. And when no installer can possibly succeed, the banner names the missing piece and
+ runs *nothing* — instead of the old behaviour of running `npm install -g …` on a machine
+ with no npm and letting you watch `command not found` scroll past.
+- **Hooks stopped silently dying with exit 127.** Agent CLIs run hooks through `sh -c` with
+ a bare `PATH=/usr/bin:/bin:/usr/sbin:/sbin` — nvm's node isn't on it — so every hook
+ payload was being lost: no live status, no Stop→inbox drain, no session ids. Every shim
+ now runs under the Node the app already ships. This is also why the orchestrator kept
+ going stale across restarts.
+- **`node` is on every agent's PATH.** An MCP server declared as `node ./server.js`, a
+ provider CLI that shells out, a `.cjs` an agent wrote for itself — all died with 127 with
+ no system node. The bundled runtime is now *appended*, never prepended, so if you have
+ your own node you keep your own version.
+- **`~/dev/foo` no longer fails with "cwd does not exist."** Only a shell expands `~`; Node
+ treats it as a literal folder name. Paths are expanded once on the way in, so the registry
+ only ever holds absolute paths — which also repairs existing `~` entries that had been
+ reading as invalid forever.
+- **Michael stops messaging agents that don't exist.** A live roster of the floor is pushed
+ into the orchestrator's context on session start and on every prompt, rather than trusting
+ it to go re-read `fleet.json` on its own.
+- **The office floor survives losing its GPU context.** 0.3.4 leased WebGL contexts for
+ *terminals* when Chromium started evicting the oldest past ~16 — the office floor was the
+ remaining victim, and always the first to go since it's created at startup. Pixi reported
+ nothing at all, so the floor just went blank until you restarted. It now notices and
+ rebuilds itself.
+
+---
+
+## What was new in 0.3.5 — *The queue always has an escape hatch*
A fast-follow to yesterday's big 0.3.4 wave — and the **first release the app delivers to
itself**: if you're on 0.3.4, it downloads in the background and offers "Restart to
diff --git a/package.json b/package.json
index 6a5e3a33..7e450445 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
{
"name": "munder-difflin",
- "version": "0.3.5",
+ "version": "0.3.6",
"private": true,
"description": "Local multi-agent harness for Claude Code: autonomous agents that message, route, and remember — coordinated by a GOD orchestrator and visualized as avatars at work.",
"main": "out/main/index.js",