localcode 0.3.38__tar.gz → 0.3.40__tar.gz

{localcode-0.3.38/src/localcode.egg-info → localcode-0.3.40}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.38
+Version: 0.3.40
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0
@@ -121,7 +121,6 @@ LocalCode is early software. Hardware support is expected to broaden, but only t
 | M4 MacBook | 16 GB | Tested | Validated at the 16 GB floor |
 | M1/M2/M3 Apple Silicon | 16 GB+ | Not yet tested | Expected to work, but needs validation |
 | M4 Apple Silicon | 24 GB+ | Not yet tested | Expected to support larger contexts, but needs validation |
-| Intel Mac | Any | Not supported | LocalCode targets Apple Silicon |
 
 > Linux installs and runs in CI for development, but Apple Silicon is the
 > supported target — the Metal-accelerated inference path is Mac-only.
@@ -136,9 +135,11 @@ On launch, LocalCode recommends the best model for **your Mac's RAM** — there'
 | Gemma 4 26B-A4B | 11.2 GB (Q3) | 3.8B (8/128 experts) | 24 GB | gemma4-iswa |
 | Qwen 3.6 35B-A3B | 10.7 GB (Q2) | 3.0B (8+1/256) | 24 GB | qwen35moe |
 | DiffusionGemma 26B-A4B † | 15.7 GB (Q4) | 4B (diffusion MoE) | 32 GB | diffusion_gemma |
+| Qwen 3.8 27B (vision) | 17.9 GB (Q4) | 27B (dense) | 32 GB | qwen35 |
 | North-Mini-Code 30B-A3B † | 17.9 GB (Q4) | 3B (30B MoE) | 36 GB | cohere2_moe |
 | Gemma 4 12B (full) | 23.8 GB (BF16) | 12B (dense) | 48 GB | gemma4-iswa |
 | Gemma 4 26B-A4B | 28 GB (Q8) | 3.8B (8/128 experts) | 64 GB | gemma4-iswa |
+| Qwen 3.8 27B (vision) | 31.5 GB (Q8) | 27B (dense) | 64 GB | qwen35 |
 | Qwen 3.6 35B-A3B | 38.5 GB (Q8) | 3.0B (8+1/256) | 96 GB | qwen35moe |
 
 *Min RAM* is the threshold for auto-recommendation (weights ≤ ~55% of unified memory, leaving room for KV cache + OS); you can still pick a heavier model manually. **†** experimental — pickable but **not** auto-recommended (DiffusionGemma needs a separate runner; `cohere2_moe` is unvalidated on this stack).

{localcode-0.3.38 → localcode-0.3.40}/README.md

@@ -74,7 +74,6 @@ LocalCode is early software. Hardware support is expected to broaden, but only t
 | M4 MacBook | 16 GB | Tested | Validated at the 16 GB floor |
 | M1/M2/M3 Apple Silicon | 16 GB+ | Not yet tested | Expected to work, but needs validation |
 | M4 Apple Silicon | 24 GB+ | Not yet tested | Expected to support larger contexts, but needs validation |
-| Intel Mac | Any | Not supported | LocalCode targets Apple Silicon |
 
 > Linux installs and runs in CI for development, but Apple Silicon is the
 > supported target — the Metal-accelerated inference path is Mac-only.
@@ -89,9 +88,11 @@ On launch, LocalCode recommends the best model for **your Mac's RAM** — there'
 | Gemma 4 26B-A4B | 11.2 GB (Q3) | 3.8B (8/128 experts) | 24 GB | gemma4-iswa |
 | Qwen 3.6 35B-A3B | 10.7 GB (Q2) | 3.0B (8+1/256) | 24 GB | qwen35moe |
 | DiffusionGemma 26B-A4B † | 15.7 GB (Q4) | 4B (diffusion MoE) | 32 GB | diffusion_gemma |
+| Qwen 3.8 27B (vision) | 17.9 GB (Q4) | 27B (dense) | 32 GB | qwen35 |
 | North-Mini-Code 30B-A3B † | 17.9 GB (Q4) | 3B (30B MoE) | 36 GB | cohere2_moe |
 | Gemma 4 12B (full) | 23.8 GB (BF16) | 12B (dense) | 48 GB | gemma4-iswa |
 | Gemma 4 26B-A4B | 28 GB (Q8) | 3.8B (8/128 experts) | 64 GB | gemma4-iswa |
+| Qwen 3.8 27B (vision) | 31.5 GB (Q8) | 27B (dense) | 64 GB | qwen35 |
 | Qwen 3.6 35B-A3B | 38.5 GB (Q8) | 3.0B (8+1/256) | 96 GB | qwen35moe |
 
 *Min RAM* is the threshold for auto-recommendation (weights ≤ ~55% of unified memory, leaving room for KV cache + OS); you can still pick a heavier model manually. **†** experimental — pickable but **not** auto-recommended (DiffusionGemma needs a separate runner; `cohere2_moe` is unvalidated on this stack).

{localcode-0.3.38 → localcode-0.3.40}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "localcode"
-version = "0.3.38"
+version = "0.3.40"
 description = "High-performance AI coding on consumer hardware."
 readme = "README.md"
 requires-python = ">=3.10"

{localcode-0.3.38 → localcode-0.3.40}/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__ = "0.3.38"
+    __version__ = "0.3.40"

{localcode-0.3.38 → localcode-0.3.40}/src/localcode/models_catalog.py

@@ -323,6 +323,36 @@ CHOICES: list[ModelChoice] = [
         mmproj_size_gb=0.9,
         mmproj_hf_filename="mmproj-F16.gguf",
     ),
+    ModelChoice(
+        key="qwen38",
+        name="Qwen 3.8 27B (Q4)",
+        hf_repo="unsloth/Qwen3.8-27B-GGUF",
+        filename="Qwen3.8-27B-UD-Q4_K_XL.gguf",
+        size_gb=17.9,
+        active_params="27B (dense)",
+        architecture="qwen35",
+        license="Apache 2.0",
+        humaneval_pass_at_1=None,
+        notes="Qwen 3.8 27B — a DENSE vision-language model (not MoE): understands images, thinking on by default (per-request toggle), native 262K context. Because it's dense, decode speed scales with ALL 27B params, so it's heavier than the 3B-active MoE entries — recommend >=32 GB unified memory (comfortable on 48 GB+). UD-Q4_K_XL keeps attention/embedding layers above 4-bit for quality. Pair with the F16 mmproj for image input. qwen35 arch is natively supported by the TurboQuant server (turbo4 KV + context checkpoints).",
+        mmproj_filename="mmproj-Qwen3.8-27B-F16.gguf",
+        mmproj_size_gb=0.9,
+        mmproj_hf_filename="mmproj-F16.gguf",
+    ),
+    ModelChoice(
+        key="qwen38-q8",
+        name="Qwen 3.8 27B (Q8)",
+        hf_repo="unsloth/Qwen3.8-27B-GGUF",
+        filename="Qwen3.8-27B-UD-Q8_K_XL.gguf",
+        size_gb=31.5,
+        active_params="27B (dense)",
+        architecture="qwen35",
+        license="Apache 2.0",
+        humaneval_pass_at_1=None,
+        notes="High-RAM workstation pick (>=64 GB unified memory; sweet spot on 128 GB Apple Silicon). Near-lossless UD-Q8_K_XL of the dense Qwen 3.8 27B vision-language model (~99% of BF16 quality). Thinking on by default, native 262K context, image input via the F16 mmproj. Dense 27B -> heavier decode than the MoE entries; the extra RAM buys quality, not speed.",
+        mmproj_filename="mmproj-Qwen3.8-27B-F16.gguf",
+        mmproj_size_gb=0.9,
+        mmproj_hf_filename="mmproj-F16.gguf",
+    ),
 ]
 
 

{localcode-0.3.38 → localcode-0.3.40}/src/localcode/tui/screens/chat.py

@@ -4008,6 +4008,14 @@ class ChatScreen(Screen):
             # was still running ask() — two agent loops mutating one session.
             if event.worker.name == "run_agent_turn":
                 self._on_turn_done()
+        elif event.state == WorkerState.CANCELLED:
+            # Esc-to-interrupt cancels the worker (`_interrupt_turn` calls
+            # `w.cancel()`), so a cancelled agent turn lands here — NOT in the
+            # SUCCESS branch. Without this the turn never finalizes: the
+            # `◆ thinking…` indicator keeps animating and `_agent_busy` stays
+            # True (input wedged) after a cancel. Mirror the SUCCESS/ERROR cleanup.
+            if event.worker.name == "run_agent_turn":
+                self._on_turn_done()
 
     def _recover_stuck_state(self) -> None:
         """Clear blocking flags after a worker crash so the UI can't wedge.

{localcode-0.3.38 → localcode-0.3.40/src/localcode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.38
+Version: 0.3.40
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0
@@ -121,7 +121,6 @@ LocalCode is early software. Hardware support is expected to broaden, but only t
 | M4 MacBook | 16 GB | Tested | Validated at the 16 GB floor |
 | M1/M2/M3 Apple Silicon | 16 GB+ | Not yet tested | Expected to work, but needs validation |
 | M4 Apple Silicon | 24 GB+ | Not yet tested | Expected to support larger contexts, but needs validation |
-| Intel Mac | Any | Not supported | LocalCode targets Apple Silicon |
 
 > Linux installs and runs in CI for development, but Apple Silicon is the
 > supported target — the Metal-accelerated inference path is Mac-only.
@@ -136,9 +135,11 @@ On launch, LocalCode recommends the best model for **your Mac's RAM** — there'
 | Gemma 4 26B-A4B | 11.2 GB (Q3) | 3.8B (8/128 experts) | 24 GB | gemma4-iswa |
 | Qwen 3.6 35B-A3B | 10.7 GB (Q2) | 3.0B (8+1/256) | 24 GB | qwen35moe |
 | DiffusionGemma 26B-A4B † | 15.7 GB (Q4) | 4B (diffusion MoE) | 32 GB | diffusion_gemma |
+| Qwen 3.8 27B (vision) | 17.9 GB (Q4) | 27B (dense) | 32 GB | qwen35 |
 | North-Mini-Code 30B-A3B † | 17.9 GB (Q4) | 3B (30B MoE) | 36 GB | cohere2_moe |
 | Gemma 4 12B (full) | 23.8 GB (BF16) | 12B (dense) | 48 GB | gemma4-iswa |
 | Gemma 4 26B-A4B | 28 GB (Q8) | 3.8B (8/128 experts) | 64 GB | gemma4-iswa |
+| Qwen 3.8 27B (vision) | 31.5 GB (Q8) | 27B (dense) | 64 GB | qwen35 |
 | Qwen 3.6 35B-A3B | 38.5 GB (Q8) | 3.0B (8+1/256) | 96 GB | qwen35moe |
 
 *Min RAM* is the threshold for auto-recommendation (weights ≤ ~55% of unified memory, leaving room for KV cache + OS); you can still pick a heavier model manually. **†** experimental — pickable but **not** auto-recommended (DiffusionGemma needs a separate runner; `cohere2_moe` is unvalidated on this stack).

{localcode-0.3.38 → localcode-0.3.40}/src/localcode.egg-info/SOURCES.txt

@@ -929,6 +929,7 @@ tests/test_agent_state_machine.py
 tests/test_anchor_guard.py
 tests/test_architecture.py
 tests/test_bootstrap.py
+tests/test_chat_worker_cancel.py
 tests/test_checkpoint.py
 tests/test_churn_detection.py
 tests/test_clipboard_image.py

localcode-0.3.40/tests/test_chat_worker_cancel.py

@@ -0,0 +1,37 @@
+"""Regression: an Esc-interrupted (cancelled) agent turn must finalize.
+
+_interrupt_turn calls `worker.cancel()`, so the run_agent_turn worker lands in
+WorkerState.CANCELLED — not SUCCESS. Before the fix, on_worker_state_changed
+handled only SUCCESS/ERROR, so a cancel left the thinking indicator animating
+and `_agent_busy` stuck True (input wedged). The CANCELLED branch must call
+_on_turn_done() for the agent-turn worker (and only that worker).
+"""
+from __future__ import annotations
+
+import types
+
+from textual.worker import WorkerState
+
+from localcode.tui.screens.chat import ChatScreen
+
+
+def _event(state, name):
+    return types.SimpleNamespace(
+        state=state,
+        worker=types.SimpleNamespace(name=name, error=None),
+    )
+
+
+def test_cancelled_agent_turn_calls_on_turn_done():
+    calls = {"done": 0}
+    fake = types.SimpleNamespace(_on_turn_done=lambda: calls.__setitem__("done", calls["done"] + 1))
+    ChatScreen.on_worker_state_changed(fake, _event(WorkerState.CANCELLED, "run_agent_turn"))
+    assert calls["done"] == 1
+
+
+def test_cancelled_non_agent_worker_does_not_finalize_turn():
+    calls = {"done": 0}
+    fake = types.SimpleNamespace(_on_turn_done=lambda: calls.__setitem__("done", calls["done"] + 1))
+    # a cancelled helper worker (e.g. a model-swap probe) must not end the turn
+    ChatScreen.on_worker_state_changed(fake, _event(WorkerState.CANCELLED, "some_helper"))
+    assert calls["done"] == 0

{localcode-0.3.38 → localcode-0.3.40}/tests/test_comprehensive_models_matrix.py

@@ -91,7 +91,7 @@ def test_every_choice_is_wellformed(choice):
     assert choice.filename and choice.filename.endswith(".gguf")
     assert choice.size_gb and choice.size_gb > 0
     assert choice.architecture in {
-        "gemma4-iswa", "qwen35moe", "diffusion_gemma", "cohere2_moe",
+        "gemma4-iswa", "qwen35moe", "qwen35", "diffusion_gemma", "cohere2_moe",
     }, f"unknown arch {choice.architecture!r}"
     # by_filename must round-trip to a choice carrying the same architecture
     # (this is the dispatch key the runtime keys off).

{localcode-0.3.38 → localcode-0.3.40}/tests/test_models_catalog.py

@@ -78,6 +78,8 @@ _EXPECTED_VISION = {
     "qwen": True,
     "gemma-q8": True,
     "qwen-q8": True,
+    "qwen38": True,
+    "qwen38-q8": True,
     "diffusiongemma": False,
     "north-mini-code": False,
 }
@@ -243,3 +245,18 @@ class TestEstimateDecodeTokS:
         assert 60 <= result <= 105, (
             f"Gemma IQ3 estimated {result} tok/s on M5 Max — expected 60-105 range"
         )
+
+
+def test_qwen38_dense_vision_entry_wired():
+    """Qwen 3.8 27B: dense qwen35 arch (natively served), vision + thinking,
+    and it resolves to the QWEN sampler/family so it gets Qwen's tight profile."""
+    from localcode.models_catalog import by_filename
+    from localcode.model_families import infer_family_from_profile, ModelFamily
+    for fn in ("Qwen3.8-27B-UD-Q4_K_XL.gguf", "Qwen3.8-27B-UD-Q8_K_XL.gguf"):
+        c = by_filename(fn)
+        assert c is not None, f"{fn} missing from catalog"
+        assert c.architecture == "qwen35"          # dense arch the TurboQuant server ships
+        assert c.supports_vision is True            # mmproj sidecar present
+        assert c.supports_thinking is True          # qwen thinking channel
+        assert c.mmproj_hf_filename == "mmproj-F16.gguf"
+        assert infer_family_from_profile(c.key) == ModelFamily.QWEN