modelsdesk 0.0.0-stage → 0.2.0

package/bin/modelsd.js

@@ -0,0 +1,33 @@
+#!/usr/bin/env node
+// modeldesk bin shim (platform-binary pattern): resolve this platform's
+// native binary from the matching @modelsdesk/<platform>
+// optionalDependency and hand it the argv verbatim. Clear, actionable error
+// when the binary is missing — never a raw stack trace.
+import { spawnSync } from "node:child_process";
+import { createRequire } from "node:module";
+
+const SUPPORTED = new Set(["darwin-arm64"]);
+
+const platform = `${process.platform}-${process.arch}`;
+const require = createRequire(import.meta.url);
+
+let bin;
+try {
+  bin = require.resolve(`@modelsdesk/${platform}/bin/modelsd`);
+} catch (err) {
+  console.error(`modelsd: no native binary for ${platform}.`);
+  if (!SUPPORTED.has(platform)) {
+    console.error(
+      `  ${platform} is not supported yet (supported: ${[...SUPPORTED].sort().join(", ")}).`,
+    );
+    console.error(`  Linux/Windows follow the provider-agnostic core — track releases.`);
+  } else {
+    console.error(
+      `  Install the platform package: npm install -g @modelsdesk/${platform}`,
+    );
+  }
+  process.exit(1);
+}
+
+const result = spawnSync(bin, process.argv.slice(2), { stdio: "inherit" });
+process.exit(result.status ?? 1);

package/package.json

@@ -1,6 +1,28 @@
 {
   "name": "modelsdesk",
-  "version": "0.0.0-stage",
-  "stub": true,
-  "description": "Temporary package placeholder for staged publishing"
-}
+  "version": "0.2.0",
+  "description": "One CLI to find, install, serve, and monitor local LLMs — macOS first, provider-agnostic core.",
+  "license": "GPL-3.0-or-later",
+  "type": "module",
+  "bin": {
+    "modelsd": "bin/modelsd.js"
+  },
+  "optionalDependencies": {
+    "@modelsdesk/darwin-arm64": "0.1.0"
+  },
+  "engines": {
+    "node": ">=20"
+  },
+  "files": [
+    "bin/",
+    "skills/"
+  ],
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/talha7k/modelsdesk.git"
+  },
+  "homepage": "https://github.com/talha7k/modelsdesk#readme",
+  "bugs": {
+    "url": "https://github.com/talha7k/modelsdesk/issues"
+  }
+}

package/skills/modeldesk/SKILL.md

@@ -0,0 +1,149 @@
+---
+name: modeldesk
+description: >-
+  Manage local LLM servers with the modeldesk CLI: search Hugging Face,
+  set up models end-to-end, start/stop/health, switch models, sync agent
+  configs, check backend/model freshness. Use when the user runs models
+  locally, asks about local inference, needs a model server for
+  OpenCode/jcode/agents, or mentions modeldesk, mlx-dspark, llama.cpp,
+  ollama, or local model health.
+
+  Trigger phrases include "local model", "local llm", "start the model",
+  "model server", "model health", "switch model", "which models fit",
+  "find a model", "model outdated", "update the backend", "modeldesk"
+---
+
+# ModelDesk
+
+`modelsd` (the modelsdesk CLI) is a Rust CLI that manages local LLM inference servers —
+macOS first (MLX + mlx-dspark), provider-agnostic core (llama.cpp,
+Ollama work cross-platform).
+
+- **Install:** `cargo install modelsdesk` (crates.io) or
+  `cargo install --path .` from https://github.com/talha7k/modelsdesk
+- **Source:** https://github.com/talha7k/modelsdesk (GPL-3.0-or-later)
+
+## Command map
+
+| Task | Command |
+|---|---|
+| Find runnable models on HF | `modelsd search <terms> [--limit N]` |
+| Set up end-to-end (verify → install → download → serve) | `modelsd setup <repo> [--name N] [--provider ID] [--drafter R] [--start]` |
+| What's installed (+versions) | `modelsd providers` |
+| What's downloaded | `modelsd models` |
+| Validate before launch | `modelsd doctor [name] [--fix]` |
+| Start (validated, auto-installs backend) | `modelsd start [--foreground] [--no-install]` |
+| Stop | `modelsd stop` |
+| Restart | `modelsd restart` |
+| Health probe | `modelsd health` (alias `status`) |
+| Switch model (hot-swap + agent re-sync) | `modelsd use <name> [--provider <id>] [--restart]` |
+| Pre-fetch weights | `modelsd pull [name]` |
+| Backend + model freshness | `modelsd outdated` |
+| Sync OpenCode/jcode to active model | `modelsd sync-agents` |
+| Live console | `modelsd watch` (q to quit) |
+| Logs | `modelsd logs [n] [--follow]` |
+| Profiles | `modelsd profile <list\|show\|add\|remove>` |
+| Completions | `modelsd completions <bash\|zsh\|fish\|powershell\|elvish>` |
+
+Append `--json` to health/models/providers/use/doctor/search/outdated/sync-agents
+for machine output. Append `--verbose` for stderr diagnostics.
+`MODELDESK_NO_INSTALL=1` disables auto-install everywhere.
+
+## Setup flow (the happy path)
+
+```bash
+modelsd search qwen3 8b                 # annotated HF hits
+modelsd setup mlx-community/Qwen3-8B-8bit --name qwen3-8b --start
+modeldesk health
+```
+
+`setup` verifies the repo exists on HF, recommends (or takes) the backend,
+installs the backend when missing (brew/pipx, with progress), runs the full
+doctor validation, downloads weights, saves the profile, and optionally
+starts serving. `search` annotates each hit with format, recommended backend,
+backend install status, RAM estimate vs this machine, and cache state.
+
+## Backends and auto-install
+
+| Backend | Weights | Platforms | Install (automatic) |
+|---|---|---|---|
+| `mlx-dspark` | MLX (+ DSpark drafter) | macOS arm64, macOS 14+ | `pipx install mlx-dspark` |
+| `mlx-lm` | MLX | macOS arm64, macOS 14+ | `brew install mlx-lm` |
+| `llama-cpp` | GGUF (auto-downloaded) | macOS/Linux/Windows | brew / apt / dnf / pacman / winget / choco |
+| `ollama` | registry names | macOS/Linux/Windows | brew / install script / winget |
+| `strata` | own tiers (q2_0, iq3_s, gyro-s…) | macOS/Linux/Windows | `curl … install.sh \| sh` (Strata) |
+| `mistral-rs` | GGUF + safetensors | macOS/Linux/Windows | `curl … mistralrs.dev/install.sh \| sh` |
+| `vllm` | safetensors/GGUF | Linux/Windows (CUDA) | `pip install vllm` |
+| `sglang` | safetensors/GGUF | Linux (CUDA) | `pip install 'sglang[all]'` |
+
+`start`, `setup`, and `doctor --fix` install a missing backend automatically
+(OS + distro detected at runtime). Minimum versions are gated (mlx-dspark
+≥0.17, llama.cpp ≥b5000); older backends warn but still run. GGUF weights for
+llama.cpp download into `~/.cache/modeldesk/gguf/` (Q4_K_M preferred).
+
+## Profiles (stored model configs)
+
+Profiles live in `config.toml` (see `modelsd profile list` for the path).
+Each profile: `name + provider + repo + drafter? + kv_bits + context_window? +
+temperature? + mode?`.
+
+```bash
+modeldesk profile add lfm-2.5-8b LiquidAI/LFM2.5-8B-A1B-MLX-4bit \
+  --drafter LiquidAI/LFM2.5-8B-A1B-DSpark
+modeldesk profile add mistral-7b ./models/mistral-7b.Q4_K_M.gguf \
+  --provider llama-cpp --context-window 8192
+modeldesk use lfm-2.5-8b
+```
+
+`profile add` validates format/provider/arch **before saving** — a bad
+combo is rejected with the reason, never stored.
+
+## Freshness (outdated detection)
+
+- `modelsd outdated`: live check of backend releases (PyPI / brew
+  livecheck / GitHub releases) + cached-model staleness (HF `lastModified`
+  vs cache mtime, 24h grace for metadata-only bumps).
+- **Ambient warnings:** every other command prints cached `! outdated: …`
+  stderr warnings when something is stale. Cache is 6 hours, refreshed by a
+  detached background worker that never blocks or fails the command. Skipped
+  for `outdated` itself, completions, and `--json` runs.
+
+## Compatibility rules (also in `references/compatibility.md`)
+
+- Format is detected from the repo name: `-MLX-`/`mlx-` → MLX,
+  `.gguf`/`-GGUF` → GGUF, else Safetensors/Unknown.
+- `mlx-dspark` / `mlx-lm`: MLX weights only. Apple Silicon + macOS 14+ only.
+- `llama-cpp`: GGUF only. `ollama`: own registry names, `ollama pull` first.
+- RAM is estimated from param count × quant + drafter + overhead;
+  `doctor` fails above 85% of machine RAM, warns above 60%.
+- Hot-swap (`POST /admin/load`) is same-backend only (mlx-dspark).
+  A provider change always does stop + full restart.
+
+## Wiring agents to the server
+
+The server is OpenAI-compatible at `http://127.0.0.1:<port>/v1`
+(default port 8090) plus an Anthropic endpoint at `/v1/messages`.
+
+- **OpenCode** (`~/.config/opencode/opencode.json`): provider with
+  `"npm": "@ai-sdk/openai-compatible"`, `"baseURL": "http://127.0.0.1:8090/v1"`,
+  model id = the served repo short name.
+- **jcode**: `jcode provider add <name> --base-url http://127.0.0.1:8090/v1
+  --model <model-id> --api-key-stdin` (any dummy key; the server ignores it,
+  but jcode's discovery probe requires the header). Note: jcode `auth-test`
+  FAILs on its `/v1/models` probe (no bearer sent) while chat completions
+  work — verify with a real `run`, not `auth-test`.
+
+`modelsd sync-agents` rewrites both configs to the active profile;
+`modelsd use` runs it automatically after every switch.
+
+## Safety rules
+
+- `stop` matches backend binaries on modeldesk's port only. A non-backend
+  process (Expo, another tool) can never match — it is reported by name,
+  never killed blindly.
+- `doctor` runs before every `start` when the setup changed.
+- Spinners degrade to plain lines when piped — scripts always see results.
+- Memory pressure on 16 GB Macs is normal with 8B models — suggest closing
+  apps or a smaller quant before blaming the server.
+- Reasoning models put thinking in `reasoning_content`, leaving `content`
+  empty on short replies — agents expecting `content` may show blanks.

package/skills/modeldesk/references/compatibility.md

@@ -0,0 +1,59 @@
+# ModelDesk compatibility reference
+
+How `modeldesk doctor` and `profile add` decide whether a model runs on a
+backend. Same logic lives in `src/compat.rs` — this file is the human copy.
+
+## Format detection (from repo string, lowercased)
+
+| Signal | Format |
+|---|---|
+| ends with `.gguf`, contains `.gguf` or `-gguf` | GGUF |
+| contains `-mlx`, `mlx-`, `/mlx-` | MLX |
+| ends with `.safetensors` or contains `safetensors` | Safetensors |
+| none of the above | Unknown (treated as MLX/Safetensors with a warning) |
+
+## Provider matrix
+
+| Backend | Accepts | Platforms | Notes |
+|---|---|---|---|
+| `mlx-dspark` | MLX (+ converted Safetensors from `LiquidAI/`, `mlx-community/`) | macOS arm64 only | DSpark speculative decoding; needs a drafter for speedup, runs without one in baseline |
+| `mlx-lm` | MLX (+ converted Safetensors, same rule) | macOS arm64 only | Plain decoding, no speculation |
+| `llama-cpp` | GGUF only | macOS, Linux, Windows | Needs a local `.gguf` file path or HF GGUF repo; `--ctx-size` from profile |
+| `ollama` | Ollama registry names only | macOS, Linux, Windows | Managed daemon: `ollama pull <name>` first; modeldesk does not download for it |
+
+## RAM estimation
+
+`params × bytes-per-param + 1.5 GB overhead + 1.5 GB if drafter`
+
+- Param count parsed from the repo name (`8b` → 8B, `2.6b` → 2.6B …; default 8B).
+- Bytes/param: 4-bit/q4 → 0.55, 8-bit/q8 → 1.05, bf16/fp16/f16 → 2.1, default 1.05.
+- `doctor` FAILS above 85% of machine RAM, WARNS above 60%.
+
+## Hot-swap vs restart
+
+- Same backend + provider supports `POST /admin/load` (mlx-dspark): hot-swap, port stays.
+- Anything else (provider change, llama-cpp, ollama, mlx-lm): full stop + boot.
+- `modeldesk use <name> --restart` forces the full cycle.
+
+## Backend installation (runtime layer)
+
+`start`, `setup`, and `doctor --fix` install a missing backend automatically
+unless `--no-install` / `MODELDESK_NO_INSTALL=1`. OS, version, arch, distro,
+and available package managers are detected at runtime (`src/runtime_env.rs`):
+
+- macOS: brew → pipx → uv (MLX needs macOS 14+ on arm64, enforced)
+- Linux: brew → pipx/uv → apt / dnf / pacman / zypper / apk by distro
+- Windows: winget → choco → scoop
+
+Minimum versions: mlx-dspark ≥0.17, llama.cpp ≥b5000. Older backends warn
+(`old, want ≥X` in `providers`/`doctor`) but still run. Versions resolve via
+binary → pipx → brew since MLX binaries have no `--version`.
+
+## Freshness model
+
+- Backends: PyPI (mlx-dspark, mlx-lm fallback) / `brew livecheck` (mlx-lm,
+  ollama) / GitHub releases (llama.cpp, ollama fallback).
+- Models: HF `lastModified` vs cache snapshot mtime, 24h grace for
+  metadata-only bumps (prevents false "stale" on README edits).
+- `outdated` is always live; ambient warnings use a 6h cache refreshed by a
+  detached worker (`outdated --refresh-cache`, hidden flag).

package/README.md

@@ -1,3 +0,0 @@
-# Temporary Holding Version
-
-This version is a temporary placeholder for this package. An operational version to replace this has been submitted for review and is awaiting a staged release.