localcode 0.5.24__tar.gz → 0.5.26__tar.gz

{localcode-0.5.24/src/localcode.egg-info → localcode-0.5.26}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.5.24
+Version: 0.5.26
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0
@@ -65,7 +65,7 @@ Dynamic: license-file
   <img src="https://raw.githubusercontent.com/mjwsolo/localcode/main/docs/assets/demo/first-change.gif" alt="A localcode turn: it reads the stub and test, edits the file, then pytest reports 1 passed" width="900">
 </p>
 
-localcode runs an open-weight model on your Mac and uses it to read, edit and test your code. Your prompts and your files stay on your machine. The only thing it downloads is the model weights, once per model.
+localcode runs an open-weight model on your Mac and uses it to read, edit and test your code. Model inference runs locally.
 
 ## Install
 
@@ -82,25 +82,10 @@ cd your-project
 localcode
 ```
 
-On first launch the model picker opens: models first, then every quant the
-repository ships, with the size and whether it fits your Mac's memory. The
-recommended one for your machine is starred. Pick one, watch the download, and
-start typing. `/models` switches later.
-
-Project plugins and configuration load only with `localcode --trust-project`.
-Use it after reviewing the project's startup code and settings.
-
-OpenJev also supports typed decisions: use `/mode` to choose **Chat** or
-**Decisions**, or run `localcode --mode decisions`. The decision screen returns
-choices, yes/no probabilities and scores without running coding tools.
-See [Chat and Decisions](https://mjwsolo.github.io/localcode/guides/decisions).
-
 ```
 > Implement the retry decorator in retry.py so every test in test_retry.py passes. Then run: pytest -q
 ```
 
-Docs: [mjwsolo.github.io/localcode](https://mjwsolo.github.io/localcode/)
-
 To use the running model from another local app, open `/server` or run `localcode api` for its
 OpenAI compatible endpoint. See the [Local API guide](https://mjwsolo.github.io/localcode/guides/local-api/).
 
@@ -117,37 +102,26 @@ OpenAI compatible endpoint. See the [Local API guide](https://mjwsolo.github.io/
 - Mac with Apple Silicon, macOS 13 or newer
 - 16 GB unified memory or more
 - Python 3.10 or newer
-- About 12 GB of free disk for the smallest model
+- Disk space for LocalCode and your chosen model (download sizes below)
 
 ## Models
 
-Model downloads run in parallel and keep verified chunks after cancellation or
-interruption. Select the same model again to resume. Resume requires HTTP range
-support and a stable server ETag; otherwise the download restarts safely.
-
-localcode recommends a model by your Mac's memory and marks it with a star. You choose; nothing is selected for you. Every quant the model's repository ships is listed, with a fit badge for your machine. Every model runs on binaries shipped in the package.
-
-| Model | Weights | Quant | Active params | Min RAM |
-| --- | ---: | --- | --- | ---: |
-| Gemma 4 12B | 7.4 GB | UD-Q4_K_XL | 12B (dense) | 16 GB |
-| Qwen 3.6 35B-A3B | 10.7 GB | UD-IQ2_M | 3.0B (MoE) | 24 GB |
-| Gemma 4 26B-A4B | 11.2 GB | UD-IQ3_S | 3.8B (MoE) | 24 GB |
-| DiffusionGemma 26B-A4B | 15.7 GB | Q4_K_M | 4B (diffusion MoE) | 32 GB |
-| Muse Glimmer 30B | 15.9 GB | UD-Q4_K_XL | 30B (dense, vision) | 32 GB |
-| Qwen 3.8 27B | 17.9 GB | UD-Q4_K_XL | 27B (dense) | 36 GB |
-| North-Mini-Code 30B-A3B | 17.9 GB | UD-Q4_K_M | 3B (MoE) | 36 GB |
-| OpenJev 27B (decision model) | 19.0 GB | Q4_K_M | 27B (dense) | 36 GB |
-| Gemma 4 12B (full) | 23.8 GB | BF16 | 12B (dense) | 48 GB |
-| Gemma 4 26B-A4B | 28.0 GB | UD-Q8_K_XL | 3.8B (MoE) | 64 GB |
-| Qwen 3.8 27B | 31.5 GB | UD-Q8_K_XL | 27B (dense) | 64 GB |
-| Qwen 3.6 35B-A3B | 38.5 GB | UD-Q8_K_XL | 3.0B (MoE) | 96 GB |
-| Kolibri 1 78B-A3B (experimental) | 47.5 GB | Q4_K_M | 3.5B (MoE) | 96 GB |
-
-[Underdog Saluki 27B 1.0](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0) is also available as an opt-in IQ2-mix model: 7.9 GB of weights, plus memory for context and runtime, with an optional 0.93 GB vision projector. Coding quality and minimum RAM have not been benchmarked on LocalCode.
-
-Min RAM is the smallest Mac whose memory the weights fit in (about 55% of unified memory). You can pick a heavier one by hand. DiffusionGemma, Kolibri, OpenJev and Saluki are never recommended automatically: you choose them yourself.
-
-Measured on a top-memory Apple Silicon laptop with Qwen 3.6 35B-A3B UD-IQ2_M at a 131072-token context: about 89 tokens/s generation, about 1174 tokens/s prompt processing, and 12 to 15 seconds for a typical four-tool-call task.
+Choose from **10 models and 100+ quant options** for coding, image input and typed decisions. Models download from Hugging Face. Downloads run in parallel and can resume if interrupted. Switch with `/models`.
+
+| Model | Smallest download | Largest download | 16 GB: tokens/s | 16 GB: context | 128 GB: tokens/s | 128 GB: context |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| [Gemma 4 12B](https://huggingface.co/unsloth/gemma-4-12b-it-GGUF) | 4.2 GB | 23.8 GB | — | — | — | — |
+| [Qwen 3.8 27B](https://huggingface.co/unsloth/Qwen3.8-27B-GGUF) | 6.2 GB | 31.5 GB | — | — | — | — |
+| [Underdog Saluki 27B 1.0](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0) | 7.9 GB | 7.9 GB | — | — | — | — |
+| [North-Mini-Code 30B-A3B](https://huggingface.co/unsloth/North-Mini-Code-1.0-GGUF) | 9.4 GB | 33.2 GB | — | — | — | — |
+| [Gemma 4 26B-A4B](https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF) | 9.9 GB | 27.6 GB | — | — | — | — |
+| [Qwen 3.6 35B-A3B](https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF) | 10.0 GB | 38.5 GB | — | — | — | — |
+| [Muse Glimmer 30B](https://huggingface.co/unsloth/Muse-Glimmer-30B-GGUF) | 10.7 GB | 32.3 GB | — | — | — | — |
+| [DiffusionGemma 26B-A4B](https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF) | 16.8 GB | 50.5 GB | — | — | — | — |
+| [OpenJev 27B (decisions)](https://huggingface.co/ggml-org/OpenJev-GGUF) | 19.4 GB | 53.8 GB | — | — | — | — |
+| [Kolibri 1 78B-A3B (experimental)](https://huggingface.co/Hob-forge/Kolibri-1-GGUF) | 28.6 GB | 47.5 GB | — | — | — | — |
+
+Download sizes checked on Hugging Face on 12 October 2026; optional vision and draft files are extra. **— = not measured** on that RAM tier. Speed and usable context also depend on the Mac's chip and selected quant; RAM alone does not determine them.
 
 ## Network
 

localcode-0.5.26/README.md

@@ -0,0 +1,98 @@
+<p align="center">
+  <img alt="localcode" src="https://raw.githubusercontent.com/mjwsolo/localcode/main/docs/assets/logo/lockup-dark.svg" width="360">
+</p>
+
+<p align="center">
+  <img src="https://img.shields.io/pypi/v/localcode?style=flat-square&labelColor=171A1D&color=8AB4FF" alt="PyPI">
+  <img src="https://img.shields.io/badge/license-Apache_2.0-8AB4FF?style=flat-square&labelColor=171A1D" alt="License">
+  <img src="https://img.shields.io/badge/python-3.10+-8AB4FF?style=flat-square&labelColor=171A1D" alt="Python">
+  <img src="https://img.shields.io/badge/platform-Apple%20Silicon-8AB4FF?style=flat-square&labelColor=171A1D" alt="Platform">
+</p>
+
+<p align="center">
+  <strong>Agentic coding. Local models. On your Mac.</strong>
+</p>
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/mjwsolo/localcode/main/docs/assets/demo/first-change.gif" alt="A localcode turn: it reads the stub and test, edits the file, then pytest reports 1 passed" width="900">
+</p>
+
+localcode runs an open-weight model on your Mac and uses it to read, edit and test your code. Model inference runs locally.
+
+## Install
+
+```bash
+pip install -U localcode      # or: uv pip install -U localcode
+```
+
+The inference server ships inside the package. Nothing is compiled or cloned on your machine.
+
+## Run
+
+```bash
+cd your-project
+localcode
+```
+
+```
+> Implement the retry decorator in retry.py so every test in test_retry.py passes. Then run: pytest -q
+```
+
+To use the running model from another local app, open `/server` or run `localcode api` for its
+OpenAI compatible endpoint. See the [Local API guide](https://mjwsolo.github.io/localcode/guides/local-api/).
+
+## What it does
+
+- Reads and edits files in your project
+- Runs your tests, builds, Git and shell commands inside the project on its own, and asks before touching anything outside it
+- Searches code by name, content or structure
+- Scaffolds and launches apps, then checks that they respond
+- Remembers the task across messages
+
+## Requirements
+
+- Mac with Apple Silicon, macOS 13 or newer
+- 16 GB unified memory or more
+- Python 3.10 or newer
+- Disk space for LocalCode and your chosen model (download sizes below)
+
+## Models
+
+Choose from **10 models and 100+ quant options** for coding, image input and typed decisions. Models download from Hugging Face. Downloads run in parallel and can resume if interrupted. Switch with `/models`.
+
+| Model | Smallest download | Largest download | 16 GB: tokens/s | 16 GB: context | 128 GB: tokens/s | 128 GB: context |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| [Gemma 4 12B](https://huggingface.co/unsloth/gemma-4-12b-it-GGUF) | 4.2 GB | 23.8 GB | — | — | — | — |
+| [Qwen 3.8 27B](https://huggingface.co/unsloth/Qwen3.8-27B-GGUF) | 6.2 GB | 31.5 GB | — | — | — | — |
+| [Underdog Saluki 27B 1.0](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0) | 7.9 GB | 7.9 GB | — | — | — | — |
+| [North-Mini-Code 30B-A3B](https://huggingface.co/unsloth/North-Mini-Code-1.0-GGUF) | 9.4 GB | 33.2 GB | — | — | — | — |
+| [Gemma 4 26B-A4B](https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF) | 9.9 GB | 27.6 GB | — | — | — | — |
+| [Qwen 3.6 35B-A3B](https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF) | 10.0 GB | 38.5 GB | — | — | — | — |
+| [Muse Glimmer 30B](https://huggingface.co/unsloth/Muse-Glimmer-30B-GGUF) | 10.7 GB | 32.3 GB | — | — | — | — |
+| [DiffusionGemma 26B-A4B](https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF) | 16.8 GB | 50.5 GB | — | — | — | — |
+| [OpenJev 27B (decisions)](https://huggingface.co/ggml-org/OpenJev-GGUF) | 19.4 GB | 53.8 GB | — | — | — | — |
+| [Kolibri 1 78B-A3B (experimental)](https://huggingface.co/Hob-forge/Kolibri-1-GGUF) | 28.6 GB | 47.5 GB | — | — | — | — |
+
+Download sizes checked on Hugging Face on 12 October 2026; optional vision and draft files are extra. **— = not measured** on that RAM tier. Speed and usable context also depend on the Mac's chip and selected quant; RAM alone does not determine them.
+
+## Network
+
+Inference is local. Three features use the network: model downloads, the `web_search` and `web_fetch` tools, and any MCP servers you add. See [Network Boundary](https://mjwsolo.github.io/localcode/concepts/network-boundary/) for the full list.
+
+## Why local?
+
+Powerful, personal AI should work for everyone, on any device, anywhere. That means running it locally. localcode is a first step.
+
+## Sponsors
+
+To sponsor localcode, [reach out](https://github.com/mjwsolo/localcode).
+
+Installation diagnostics: `localcode doctor`. See [compatibility tests and local hardware verification](docs/src/content/docs/development/verification.md).
+
+## Contributing
+
+See [CONTRIBUTING.md](CONTRIBUTING.md).
+
+## License
+
+Apache 2.0. See [LICENSE](LICENSE).

{localcode-0.5.24 → localcode-0.5.26}/llama-cpp-turboquant/ggml/src/ggml-metal/ggml-metal-device.m

@@ -97,6 +97,7 @@ void ggml_metal_pipeline_free(ggml_metal_pipeline_t pipeline) {
 }
 
 int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_with_params pipeline) {
+    GGML_ASSERT(pipeline.pipeline != NULL && "Metal pipeline unavailable: check kernel compilation diagnostics");
     return pipeline.pipeline->obj.maxTotalThreadsPerThreadgroup;
 }
 

{localcode-0.5.24 → localcode-0.5.26}/llama-cpp-turboquant/ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -3913,9 +3913,10 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         // sparse: the index lists are per query row, so a threadgroup can share KV with Q == 1 only
         auto cfg = use_sparse
                 ? ggml_metal_tuning::fa_vec_baseline_cfg((int) ne00, (int) ne20)
-                : ggml_metal_tuning::fa_vec_pick(
+                : ggml_metal_tuning::fa_vec_pick_kv(
                           props_dev->gpu_family,
                           (int) op->src[1]->type,
+                          (int) op->src[2]->type,
                           (int) ne00, (int) ne20,   // dk, dv (ne00 == dk for FA)
                           ne11, ne01);
 

{localcode-0.5.24 → localcode-0.5.26}/llama-cpp-turboquant/ggml/src/ggml-metal/ggml-metal-tuning.cpp

@@ -1172,4 +1172,12 @@ fa_vec_cfg_t fa_vec_pick(int gpu_family, int dtype, int dk, int dv, int64_t ne11
     return baseline;
 }
 
+fa_vec_cfg_t fa_vec_pick_kv(int gpu_family, int dtype_k, int dtype_v, int dk, int dv, int64_t ne11, int64_t ne01) {
+    const bool turbo = dtype_k == GGML_TYPE_TURBO2_0 || dtype_k == GGML_TYPE_TURBO3_0 || dtype_k == GGML_TYPE_TURBO4_0;
+    if (dtype_k != dtype_v || turbo) {
+        return fa_vec_baseline_cfg(dk, dv);
+    }
+    return fa_vec_pick(gpu_family, dtype_k, dk, dv, ne11, ne01);
+}
+
 }  // namespace ggml_metal_tuning

{localcode-0.5.24 → localcode-0.5.26}/llama-cpp-turboquant/ggml/src/ggml-metal/ggml-metal-tuning.h

@@ -72,4 +72,7 @@ fa_vec_cfg_t fa_vec_baseline_cfg(int dk, int dv);
 // Keyed by Apple GPU family; an untuned family matches no row and gets the baseline.
 fa_vec_cfg_t fa_vec_pick(int gpu_family, int dtype, int dk, int dv, int64_t ne11, int64_t ne01);
 
+// TurboQuant kernels only instantiate the baseline Q/NE variant.
+fa_vec_cfg_t fa_vec_pick_kv(int gpu_family, int dtype_k, int dtype_v, int dk, int dv, int64_t ne11, int64_t ne01);
+
 }  // namespace ggml_metal_tuning

{localcode-0.5.24 → localcode-0.5.26}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "localcode"
-version = "0.5.24"
+version = "0.5.26"
 description = "High-performance AI coding on consumer hardware."
 readme = "README.md"
 requires-python = ">=3.10"

{localcode-0.5.24 → localcode-0.5.26}/src/localcode/__init__.py

@@ -8,4 +8,4 @@ from importlib.metadata import PackageNotFoundError, version as _pkg_version
 try:
     __version__ = _pkg_version("localcode")
 except PackageNotFoundError:  # not installed (e.g. raw source checkout)
-    __version__ = "0.5.24"
+    __version__ = "0.5.26"

localcode-0.5.26/src/localcode/bin/localcode-ui

@@ -0,0 +1,4 @@
+[diffend] Oversized file quarantined before diffing.
+name: localcode-0.5.26/src/localcode/bin/localcode-ui
+size: 102460976 bytes
+sha256: e84424a39cd2b5599da152c6ea4530f693485c7b8092f127aa73c699b4018b8c

{localcode-0.5.24 → localcode-0.5.26}/src/localcode/model_config.py

@@ -35,17 +35,17 @@ from __future__ import annotations
 # pending real-hardware OOM measurement. 96 GB+ stays at 128K: at 256K the
 # Qwen checkpoint path is disabled due a measured Metal SIGKILL, forcing a
 # full prompt prefill every round. 128K keeps four checkpoints and is faster
-# for long agentic tasks; 256K remains an explicit override.
+# for long agentic tasks; launcher overrides stay within these ceilings.
 # Dedicated tier for every Apple Silicon RAM variant (MBA/MBP/Studio/Mini).
 # The VALIDATED anchors are 16→64K, 48→96K, 64→128K; every other tier is
 # INTERPOLATED strictly between its validated neighbours (or below the 16 GB
 # floor), so no tier ever exceeds a value that's already been validated on real
 # hardware — a 32 GB machine at 80K uses less KV than the 48 GB-validated 96K
 # and more than the 16 GB-validated 64K. Sizes are multiples of 8192 (clean
-# batch alignment). 256K stays an explicit override, never an auto default
+# batch alignment). Overrides cannot exceed the machine ceiling
 # (its checkpoint path SIGKILLs on Metal — see runtime.py).
 RAM_CTX_CEILING_TIERS: tuple[tuple[int, int], ...] = (
-    (192, 131072),  # 192-512 GB (Studio/Pro) — 128K default; 256K via override
+    (192, 131072),  # 192-512 GB (Studio/Pro) — 128K ceiling
     (128, 131072),  # 128 GB — 128K
     (96,  131072),  # 96 GB  — 128K
     (64,  131072),  # 64 GB  — 128K (validated)

{localcode-0.5.24 → localcode-0.5.26}/src/localcode/runtime.py

@@ -654,7 +654,7 @@ class LocalCodeRuntimeGateway:
         The tier ladder lives in `model_config.RAM_CTX_CEILING_TIERS` (the
         central per-Mac config); this delegates so the numbers are edited in
         one place. The normal high-RAM default is the checkpointed 128K path;
-        256K remains available as an explicit experimental override. Small-RAM
+        Launcher overrides are checked against the same ceiling. Small-RAM
         tiers stay conservative (16 GB→64K, 64 GB→128K hardware-measured).
         """
         from .model_config import ram_ctx_ceiling
@@ -1062,8 +1062,10 @@ class LocalCodeRuntimeGateway:
         # or future, gets a `--ctx-size` whose KV cache can't fit (the OOM
         # never-again guard). No-op when KV size isn't confidently computable
         # (turbo-compressed KV) — the validated ceilings govern there.
+        ram_gb = self._system_ram_gb()
         return min(
-            raw, self._kv_aware_ctx_ceiling(model_path, self._system_ram_gb())
+            raw, self._ram_ctx_ceiling(ram_gb),
+            self._kv_aware_ctx_ceiling(model_path, ram_gb),
         )
 
     def _target_num_ctx_uncapped(

{localcode-0.5.24 → localcode-0.5.26}/src/localcode/ui/plugin/localcode.ts

@@ -625,6 +625,7 @@ const LocalcodePlugin: Plugin = async ({ client, directory }) => {
   let stopReason = "";
   const interrupted = new Set<string>();
   let activeSession: string | undefined;
+  let lastServerCrash: string | undefined;
   // The agent the user is talking to. Nudges must keep it (a prompt without an
   // agent defaults to build, which silently escalated Plan sessions to Build),
   // and in Plan mode the model is conversing with the user, so no gate fires at all.
@@ -861,11 +862,29 @@ const LocalcodePlugin: Plugin = async ({ client, directory }) => {
     },
 
     event: async ({ event }) => {
-      if (event.type === "session.error") {
+      if (event.type === "session.error" || (event.type === "session.status" && (event.properties as any).status?.type === "retry")) {
         const props = event.properties as any;
         if (props.sessionID && props.error?.name === "MessageAbortedError") {
           interrupted.add(props.sessionID);
           log("task interrupted — automatic continuation is paused until the next user message");
+        } else if (!activeSession || props.sessionID === activeSession) {
+          const control = process.env.LOCALCODE_CONTROL_URL;
+          if (control) {
+            try {
+              const response = await fetch(`${control}/status`, { headers: controlHeaders(), signal: AbortSignal.timeout(1500) });
+              if (response.ok) {
+                const state = await response.json() as { state?: string; exit_code?: number; process_id?: number; detail?: string };
+                if (state.state === "error" && Number.isInteger(state.exit_code) && typeof state.detail === "string") {
+                  const identity = `${state.process_id}:${state.exit_code}`;
+                  if (lastServerCrash !== identity) {
+                    lastServerCrash = identity;
+                    await client.tui?.showToast?.({ body: { title: "Model server stopped", message: state.detail.slice(0, 400), variant: "error", duration: 20000 } });
+                    if (props.sessionID) await client.session?.abort?.({ path: { id: props.sessionID } });
+                  }
+                }
+              }
+            } catch { /* Keep the original error when supervisor diagnosis is unavailable. */ }
+          }
         }
         return;
       }

{localcode-0.5.24 → localcode-0.5.26}/src/localcode/ui/server_cmd.py

@@ -20,6 +20,20 @@ def server_command(gguf: str, port: int, alias: str | None = None) -> list[str]:
     g = LocalCodeRuntimeGateway(cfg.runtime)
     validate_model_memory(gguf, g._system_ram_gb())
     cmd = g.llama_server_command(gguf, port)
+    # Validate AFTER stored tuning and environment overrides have rewritten flags.
+    ram_gb = g._system_ram_gb()
+    ceiling = min(g._ram_ctx_ceiling(ram_gb), g._kv_aware_ctx_ceiling(gguf, ram_gb), g._model_max_ctx(gguf))
+    if g._is_cohere_gguf(gguf):
+        ceiling = min(ceiling, g._cohere_ctx_ceiling(ram_gb))
+    try:
+        requested_ctx = int(cmd[cmd.index("--ctx-size") + 1])
+    except (ValueError, IndexError) as exc:
+        raise ValueError("Invalid model context setting; choose a smaller context or clear the tuning override.") from exc
+    if requested_ctx <= 0:
+        raise ValueError("Model context must be a positive token count.")
+    # Old presets/tuning should recover automatically on the current machine.
+    # Weight-fit validation still refuses models that cannot leave OS headroom.
+    cmd[cmd.index("--ctx-size") + 1] = str(min(requested_ctx, ceiling))
     cmd[1:1] = ["--host", "127.0.0.1"]
     # Any web page can reach a localhost port; the key makes the model and its
     # /slots (which can hold prompt text) unusable without the session secret.

{localcode-0.5.24 → localcode-0.5.26}/src/localcode/ui/supervisor.py

@@ -107,6 +107,7 @@ class Supervisor:
         self._budget_lock = threading.Lock()
         self._warm_stop = threading.Event()
         self._warm_status = "idle"
+        self._context_ready = False
         self._warm_replay = None  # warmup.Replay while the prefix is being pre-read
 
     # ---- llama-server lifecycle -------------------------------------------
@@ -121,9 +122,32 @@ class Supervisor:
         for name in removed:
             log(f"supervisor: removed retired drafter {name} (replaced by a newer catalog drafter)")
 
+    def _server_failure(self) -> dict:
+        """Describe an unexpected exit without treating an intentional stop as a crash."""
+        process = getattr(self, "proc", None)
+        if process is None or getattr(self, "state", {}).get("state") != "ready":
+            return {}
+        code = process.poll()
+        if code is None:
+            return {}
+        name = None
+        if code < 0:
+            try:
+                name = signal.Signals(-code).name
+            except ValueError:
+                name = f"signal {-code}"
+        reason = name or f"exit code {code}"
+        verb = "crashed" if code < 0 else "exited unexpectedly"
+        return {"exit_code": code, "exit_signal": name,
+                "process_id": getattr(process, "pid", None),
+                "detail": f"Model server {verb} ({reason}). Open /server to restart or choose another model. See server.log for diagnostics."}
+
     def progress(self) -> dict:
         """What the server is doing right now: reading the prompt (with a fraction) or
         generating. Lets the TUI show 'reading context 40%' instead of a bare spinner."""
+        failure = self._server_failure()
+        if failure:
+            return {"phase": "error", **failure}
         try:
             req = urllib.request.Request(f"http://127.0.0.1:{self.port}/slots", headers=server_headers())
             with urllib.request.urlopen(req, timeout=1) as r:
@@ -134,7 +158,8 @@ class Supervisor:
         if slot is None:
             return {"phase": "idle"}
         request = {"request_id": f"{slot.get('id')}:{slot.get('id_task')}",
-                   "warmup": getattr(self, "_warm_replay", None) is not None}
+                   "warmup": getattr(self, "_warm_replay", None) is not None,
+                   "context_ready": getattr(self, "_context_ready", False)}
         total = int(slot.get("n_prompt_tokens") or 0)
         cache = int(slot.get("n_prompt_tokens_cache") or 0)
         done = int(slot.get("n_prompt_tokens_processed") or 0)
@@ -142,6 +167,9 @@ class Supervisor:
         decoded = int((nt[0] if isinstance(nt, list) and nt else nt or {}).get("n_decoded") or 0)
         todo = max(total - cache, 0)
         if decoded > 0 or (todo and done >= todo):
+            if slot.get("id") == 0 and not request["warmup"]:
+                self._context_ready = True
+                request["context_ready"] = True
             return {"phase": "generating", "decoded": decoded, "prompt_tokens": total, **request}
         if todo:
             return {"phase": "reading", "done": done, "todo": todo, "cached": cache, "pct": int(100 * done / todo), **request}
@@ -258,6 +286,7 @@ class Supervisor:
 
     # ---- prompt warm-up (see ui/warmup.py) ----------------------------------
     def _start_warmup(self, alias: str) -> None:
+        self._context_ready = False
         from localcode.ui import warmup
         if not warmup.enabled():
             self._warm_status = "disabled"
@@ -284,6 +313,7 @@ class Supervisor:
                     tm = rp.run(toks)
                     if not stop.is_set():
                         self._warm_status = "warm"
+                        self._context_ready = True
                     log(f"warmup: replayed {len(toks)} tokens for {alias} in {time.time() - t0:.1f}s "
                         f"(server prompt_ms={tm.get('prompt_ms')}, cache_n={tm.get('cache_n')})")
                     try:
@@ -423,7 +453,10 @@ class Supervisor:
         process = self.proc
         running = process is not None and process.poll() is None
         ready = running and state["state"] == "ready" and self.healthy()
-        if state["state"] == "ready" and not ready:
+        failure = self._server_failure()
+        if failure:
+            state.update(state="error", **failure)
+        elif state["state"] == "ready" and not ready:
             state.update(state="error", detail="Model server is unavailable")
         return dict(state, server_running=running, server_ready=ready, warmup=getattr(self, "_warm_status", "idle"))
 

{localcode-0.5.24 → localcode-0.5.26/src/localcode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.5.24
+Version: 0.5.26
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0
@@ -65,7 +65,7 @@ Dynamic: license-file
   <img src="https://raw.githubusercontent.com/mjwsolo/localcode/main/docs/assets/demo/first-change.gif" alt="A localcode turn: it reads the stub and test, edits the file, then pytest reports 1 passed" width="900">
 </p>
 
-localcode runs an open-weight model on your Mac and uses it to read, edit and test your code. Your prompts and your files stay on your machine. The only thing it downloads is the model weights, once per model.
+localcode runs an open-weight model on your Mac and uses it to read, edit and test your code. Model inference runs locally.
 
 ## Install
 
@@ -82,25 +82,10 @@ cd your-project
 localcode
 ```
 
-On first launch the model picker opens: models first, then every quant the
-repository ships, with the size and whether it fits your Mac's memory. The
-recommended one for your machine is starred. Pick one, watch the download, and
-start typing. `/models` switches later.
-
-Project plugins and configuration load only with `localcode --trust-project`.
-Use it after reviewing the project's startup code and settings.
-
-OpenJev also supports typed decisions: use `/mode` to choose **Chat** or
-**Decisions**, or run `localcode --mode decisions`. The decision screen returns
-choices, yes/no probabilities and scores without running coding tools.
-See [Chat and Decisions](https://mjwsolo.github.io/localcode/guides/decisions).
-
 ```
 > Implement the retry decorator in retry.py so every test in test_retry.py passes. Then run: pytest -q
 ```
 
-Docs: [mjwsolo.github.io/localcode](https://mjwsolo.github.io/localcode/)
-
 To use the running model from another local app, open `/server` or run `localcode api` for its
 OpenAI compatible endpoint. See the [Local API guide](https://mjwsolo.github.io/localcode/guides/local-api/).
 
@@ -117,37 +102,26 @@ OpenAI compatible endpoint. See the [Local API guide](https://mjwsolo.github.io/
 - Mac with Apple Silicon, macOS 13 or newer
 - 16 GB unified memory or more
 - Python 3.10 or newer
-- About 12 GB of free disk for the smallest model
+- Disk space for LocalCode and your chosen model (download sizes below)
 
 ## Models
 
-Model downloads run in parallel and keep verified chunks after cancellation or
-interruption. Select the same model again to resume. Resume requires HTTP range
-support and a stable server ETag; otherwise the download restarts safely.
-
-localcode recommends a model by your Mac's memory and marks it with a star. You choose; nothing is selected for you. Every quant the model's repository ships is listed, with a fit badge for your machine. Every model runs on binaries shipped in the package.
-
-| Model | Weights | Quant | Active params | Min RAM |
-| --- | ---: | --- | --- | ---: |
-| Gemma 4 12B | 7.4 GB | UD-Q4_K_XL | 12B (dense) | 16 GB |
-| Qwen 3.6 35B-A3B | 10.7 GB | UD-IQ2_M | 3.0B (MoE) | 24 GB |
-| Gemma 4 26B-A4B | 11.2 GB | UD-IQ3_S | 3.8B (MoE) | 24 GB |
-| DiffusionGemma 26B-A4B | 15.7 GB | Q4_K_M | 4B (diffusion MoE) | 32 GB |
-| Muse Glimmer 30B | 15.9 GB | UD-Q4_K_XL | 30B (dense, vision) | 32 GB |
-| Qwen 3.8 27B | 17.9 GB | UD-Q4_K_XL | 27B (dense) | 36 GB |
-| North-Mini-Code 30B-A3B | 17.9 GB | UD-Q4_K_M | 3B (MoE) | 36 GB |
-| OpenJev 27B (decision model) | 19.0 GB | Q4_K_M | 27B (dense) | 36 GB |
-| Gemma 4 12B (full) | 23.8 GB | BF16 | 12B (dense) | 48 GB |
-| Gemma 4 26B-A4B | 28.0 GB | UD-Q8_K_XL | 3.8B (MoE) | 64 GB |
-| Qwen 3.8 27B | 31.5 GB | UD-Q8_K_XL | 27B (dense) | 64 GB |
-| Qwen 3.6 35B-A3B | 38.5 GB | UD-Q8_K_XL | 3.0B (MoE) | 96 GB |
-| Kolibri 1 78B-A3B (experimental) | 47.5 GB | Q4_K_M | 3.5B (MoE) | 96 GB |
-
-[Underdog Saluki 27B 1.0](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0) is also available as an opt-in IQ2-mix model: 7.9 GB of weights, plus memory for context and runtime, with an optional 0.93 GB vision projector. Coding quality and minimum RAM have not been benchmarked on LocalCode.
-
-Min RAM is the smallest Mac whose memory the weights fit in (about 55% of unified memory). You can pick a heavier one by hand. DiffusionGemma, Kolibri, OpenJev and Saluki are never recommended automatically: you choose them yourself.
-
-Measured on a top-memory Apple Silicon laptop with Qwen 3.6 35B-A3B UD-IQ2_M at a 131072-token context: about 89 tokens/s generation, about 1174 tokens/s prompt processing, and 12 to 15 seconds for a typical four-tool-call task.
+Choose from **10 models and 100+ quant options** for coding, image input and typed decisions. Models download from Hugging Face. Downloads run in parallel and can resume if interrupted. Switch with `/models`.
+
+| Model | Smallest download | Largest download | 16 GB: tokens/s | 16 GB: context | 128 GB: tokens/s | 128 GB: context |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| [Gemma 4 12B](https://huggingface.co/unsloth/gemma-4-12b-it-GGUF) | 4.2 GB | 23.8 GB | — | — | — | — |
+| [Qwen 3.8 27B](https://huggingface.co/unsloth/Qwen3.8-27B-GGUF) | 6.2 GB | 31.5 GB | — | — | — | — |
+| [Underdog Saluki 27B 1.0](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0) | 7.9 GB | 7.9 GB | — | — | — | — |
+| [North-Mini-Code 30B-A3B](https://huggingface.co/unsloth/North-Mini-Code-1.0-GGUF) | 9.4 GB | 33.2 GB | — | — | — | — |
+| [Gemma 4 26B-A4B](https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF) | 9.9 GB | 27.6 GB | — | — | — | — |
+| [Qwen 3.6 35B-A3B](https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF) | 10.0 GB | 38.5 GB | — | — | — | — |
+| [Muse Glimmer 30B](https://huggingface.co/unsloth/Muse-Glimmer-30B-GGUF) | 10.7 GB | 32.3 GB | — | — | — | — |
+| [DiffusionGemma 26B-A4B](https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF) | 16.8 GB | 50.5 GB | — | — | — | — |
+| [OpenJev 27B (decisions)](https://huggingface.co/ggml-org/OpenJev-GGUF) | 19.4 GB | 53.8 GB | — | — | — | — |
+| [Kolibri 1 78B-A3B (experimental)](https://huggingface.co/Hob-forge/Kolibri-1-GGUF) | 28.6 GB | 47.5 GB | — | — | — | — |
+
+Download sizes checked on Hugging Face on 12 October 2026; optional vision and draft files are extra. **— = not measured** on that RAM tier. Speed and usable context also depend on the Mac's chip and selected quant; RAM alone does not determine them.
 
 ## Network
 

{localcode-0.5.24 → localcode-0.5.26}/src/localcode.egg-info/SOURCES.txt

@@ -1241,12 +1241,14 @@ tests/test_comprehensive_upstream_gates.py
 tests/test_comprehensive_voice.py
 tests/test_comprehensive_whisper.py
 tests/test_config.py
+tests/test_context_safety.py
 tests/test_download.py
 tests/test_eval_gates.py
 tests/test_eval_performance.py
 tests/test_eval_suite.py
 tests/test_fresh_install.py
 tests/test_kv_cap.py
+tests/test_metal_kv_selection.py
 tests/test_model_arg_resolution.py
 tests/test_model_config.py
 tests/test_model_verification.py
@@ -1255,6 +1257,8 @@ tests/test_ram_ctx_ladder.py
 tests/test_real_models.py
 tests/test_reasoning_capabilities.py
 tests/test_redaction.py
+tests/test_release_identity.py
+tests/test_release_intent.py
 tests/test_resumable_transfer.py
 tests/test_runtime.py
 tests/test_server_flags_accepted.py

{localcode-0.5.24 → localcode-0.5.26}/tests/test_compatibility.py

@@ -112,3 +112,15 @@ def test_rejected_load_keeps_existing_server(monkeypatch, tmp_path):
     with pytest.raises(ValueError, match='Insufficient RAM'):
         sup.start('too-large')
     assert not stopped
+
+
+def test_reported_qwen_q5_24gib_load_is_refused_before_inference(tmp_path, monkeypatch):
+    from localcode.ui import server_cmd
+    # Exact Hugging Face size of the reported Q5_K_S file. Sparse fixture only.
+    model = tmp_path / "Qwen3.6-35B-A3B-UD-Q5_K_S.gguf"
+    with model.open("wb") as stream:
+        stream.truncate(24_942_050_272)
+    monkeypatch.setattr(server_cmd, "mmproj_for", lambda _: None)
+    monkeypatch.setattr(server_cmd, "catalog_choice", lambda _: None)
+    with pytest.raises(ValueError, match="Insufficient RAM"):
+        server_cmd.validate_model_memory(str(model), 24)