mu-agent 0.1.14 → 0.2.0

package/CHANGELOG.md

@@ -2,6 +2,25 @@
 
 ## [Unreleased]
 
+## [0.2.0] - 2026-10-23
+
+**0.2.0 starts the era of mu's own models: `mu judge mude install` turns on mude 1.0 in one command.**
+
+### Added
+
+- `mu judge mude install` makes mude 1.0, mu's own decision model optimized for coding, the judge on a Mac with Apple silicon (macOS 14 or later, 8 GB of memory or more). The steps:
+  - It downloads the MLX 8-bit edition from huggingface.co/qybaihe/mude-1.0-mlx-8bit at a pinned revision: 31 files, 1.8 GB. Every file is checked against the sha256 pinned in `src/mude/manifest.ts`, and an interrupted download resumes with a range request.
+  - It makes a Python 3.13 environment with uv: the machine's own, or a pinned release mu downloads. The packages come from `kyrn/local-judge/mude-requirements.lock.txt` with `--require-hashes`.
+  - It starts the package's own System One server on 127.0.0.1:8790 (another free port when that one is taken) with a random key.
+  - It puts `mude` first in `tiers`. The key goes in the .env as `MU_JUDGE_MUDE_KEY`.
+- What happens when mude can't answer:
+  - By default, a decision it can't answer (its server is down, or a yes/no between 0.2 and 0.8) goes to the judges that were there before. Those are hosted, so what they read leaves the machine. `tool.description` and `tool.injection` also keep going to them (`routes`): mude did worse than Jev on the first in mu's internal benchmark, and the second was not measured.
+  - With `--local-only`, mude is the only judge.
+- Other commands: `enable [--local-only | --with-fallback]`, `disable` (back to the judges from before), `start`, `stop`, `status` and `uninstall`. `--json` prints one JSON object per line, for the desktop app.
+- A session starts the server in the background when `tiers`, a route or `MU_JUDGE` names mude, on macOS only; the desktop app's host does too, with Electron as Node. It is shared by every session like Laya's, and uses about 4 GB of memory until stopped.
+- The built-in judge `mude`: `typesafe` at 127.0.0.1:8790, key `MU_JUDGE_MUDE_KEY` (never sent to TypeSafe), 4 s timeout, trusted on every capability.
+- `mu doctor` checks the mude server's `/health` and its key.
+
 ## [0.1.14] - 2026-10-23
 
 ### Added

package/README.md

@@ -2,7 +2,7 @@
 
 **A coding agent with a judgment kernel.** It hands its routine decisions to a small, fast judge model, so the main model can keep its attention on the work.
 
-> **Early development.** The authors use mu every day, but names, settings and file formats may still change. Source, docs and the desktop app: [github.com/qybaihe/mu](https://github.com/qybaihe/mu) (README in English, 简体中文, 繁體中文, 日本語 and 한국어).
+> Source, docs and the desktop app: [github.com/qybaihe/mu](https://github.com/qybaihe/mu) (README in English, 简体中文, 繁體中文, 日本語 and 한국어).
 
 ## Install
 
@@ -32,6 +32,7 @@ Out of the box, the judge is Jev 1.13, TypeSafe's hosted judge model, free on Op
 - **Jev** (hosted): set `TYPESAFE_API_KEY` (TypeSafe), `MU_JUDGE_OPENROUTER_API_KEY` (Jev on OpenRouter), `AI_GATEWAY_API_KEY` (Vercel AI Gateway), `OPENCODE_API_KEY` (OpenCode Zen) or `CLOUDFLARE_API_KEY` with `CLOUDFLARE_ACCOUNT_ID` (Cloudflare Workers AI) in your environment, or add it to `~/.mu/.env` as `KEY=value`. `mu setup` asks for a key and writes it there. mu reads that file as data, never prints it and starts using the key automatically.
 - **Another classifier** from pi's catalog: `/mu judge classifier:<provider>/<model>`, for example Cloudflare's Clef (`clef`, `clef-flash`) or a System One model on OpenRouter, with that provider's key.
 - **CLM** (self-hosted; its encoder needs a GPU): start `clm-serve` from [Contrastive-LM/CLM](https://github.com/Contrastive-LM/CLM), then run `/mu judge clm`. For a server on another machine, add `"judges": { "clm": { "type": "clm", "baseUrl": "http://<host>:8700" } }` to `~/.mu/agent/mu.json`. If the server was started with `CLM_API_KEY`, set the same value in `MU_JUDGE_CLM_API_KEY`. `mu doctor` checks the server.
+- **mude 1.0** (local, a Mac with Apple silicon, macOS 14+), mu's own decision model, optimized for coding: `mu judge mude install` downloads it (1.8 GB), sets it up in `~/.mu/local-judge/mude`, starts its server on 127.0.0.1 and puts it first. What it can't answer goes to the judges you had before, which are hosted; `mu judge mude enable --local-only` keeps everything on your Mac. Its server uses about 4 GB of memory until `mu judge mude stop`.
 - **Laya** (local, macOS on Apple silicon): `mu judge setup` installs it into `~/.mu/local-judge`. It downloads about 930 MB, and only when you run that command.
 - **Any model you already use:** `/mu judge llm:<provider>/<model>`.
 
@@ -45,6 +46,7 @@ mu "prompt"              start a session with this prompt
 mu -p "prompt"           run one prompt, print the answer and exit
 mu -c | -r               continue the last session | pick a session to resume
 mu setup                 connect a model and choose a judge (see mu setup --help for scripting and services)
+mu judge mude <cmd>      mude 1.0 on this Mac: install | enable | disable | start | stop | status | uninstall
 mu judge <cmd>           manage the local judge, Laya (macOS on Apple silicon): setup | start | stop | status | run
 mu ledger [n] [--json]   print the judge's verdicts from the last n sessions
 mu import <file>...      import Claude Code, Codex and ZCode conversations; --list shows them (mu import --help)