localcode 0.3.42__tar.gz → 0.3.44__tar.gz

{localcode-0.3.42/src/localcode.egg-info → localcode-0.3.44}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.42
+Version: 0.3.44
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0

{localcode-0.3.42 → localcode-0.3.44}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "localcode"
-version = "0.3.42"
+version = "0.3.44"
 description = "High-performance AI coding on consumer hardware."
 readme = "README.md"
 requires-python = ">=3.10"

{localcode-0.3.42 → localcode-0.3.44}/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__ = "0.3.42"
+    __version__ = "0.3.44"

{localcode-0.3.42 → localcode-0.3.44}/src/localcode/agent/loop.py

@@ -119,6 +119,25 @@ from .app_tasks import (
     looks_like_partial_handoff,
 )
 
+
+def ephemeral_context_role(family: object) -> str:
+    """Role for the trailing per-round context block (ledger + fs-state + todo).
+
+    Gemma keeps ``system`` — as a user turn it re-greeted every round. But
+    Qwen 3.x's chat template raises ``System message must be at the beginning``
+    on ANY non-leading system message, which llama-server surfaces as HTTP 500
+    on every round once this block is non-empty (verified live: identical
+    content returns 500 as ``system``, 200 as ``user``). Qwen-family models
+    therefore receive it as a ``user`` turn.
+
+    `family` is the reasoning-caps family string (e.g. ``qwen35``,
+    ``qwen35moe``, ``gemma4-iswa``); anything falsy or non-qwen keeps the
+    historical ``system`` role.
+    """
+    fam = str(family or "").lower()
+    return "user" if fam.startswith("qwen") else "system"
+
+
 def run_agent_loop(
     app: "LocalCodeApp",
     user_text: str,
@@ -904,13 +923,24 @@ def run_agent_loop(
                 _fs_state = build_filesystem_state(changed_files)
             except Exception:
                 _fs_state = ""
-            # Inject ledger + fs-state + todo as ONE trailing SYSTEM message (not
-            # role:user — as a user turn the model re-greeted every round). Ephemeral.
+            # Inject ledger + fs-state + todo as ONE trailing ephemeral message.
+            # Role matters and is model-dependent:
+            #   - Gemma keeps role:system — as a user turn it re-greeted every round.
+            #   - Qwen 3.x's chat template raises "System message must be at the
+            #     beginning" on ANY non-leading system message → HTTP 500 every
+            #     round once this block is non-empty (verified live: identical
+            #     content returns 500 as `system`, 200 as `user`). Those families
+            #     must receive it as a user turn instead.
+            # TODO: replace the family gate with a real per-template capability
+            # probe if more families turn out to reject trailing system messages.
             try:
                 _ctx_block = "\n\n".join(b for b in (_ledger, _fs_state, _todo_note) if b)
                 if _ctx_block:
+                    _ctx_role = ephemeral_context_role(
+                        getattr(_reasoning_caps, "family", "")
+                    )
                     model_messages = list(model_messages) + [
-                        {"role": "system", "content": _ctx_block}
+                        {"role": _ctx_role, "content": _ctx_block}
                     ]
             except Exception:
                 pass

{localcode-0.3.42 → localcode-0.3.44}/src/localcode/tui/screens/setup.py

@@ -130,34 +130,57 @@ class SetupScreen(Screen):
         # screen-resize hook on this class still compiles.
         return
 
-    def _render_steps(self) -> str:
-        # Substitute the actually-selected model's size into the "Download model"
-        # label so the user sees the real number, not a hardcoded ~10GB.
-        # If a partial download exists on disk (from a prior aborted attempt),
-        # surface that as "Resume download" so the user knows we're picking up
-        # where they left off — not starting over.
+    def _model_state(self) -> tuple[str, float | None]:
+        """Disk state of the chosen model: ('present'|'partial'|'absent', size_gb).
+
+        - present: the GGUF is fully on disk (≥99% of catalog size) — this
+          launch only *loads* it, no download happens.
+        - partial: a `.incomplete` sidecar or an undersized final exists — a
+          download will *resume*.
+        - absent: nothing on disk — a fresh download is needed.
+        size_gb is None only when there's no selectable model.
+        """
         try:
             from ...models_catalog import current as current_choice
+            # Reuse the canonical completeness check so we agree with the rest
+            # of the app on what "downloaded" means. It compares against DECIMAL
+            # GB (size_gb * 1000**3); an earlier inline check here used binary
+            # GiB (1024**3), which over-counts by ~7% and made every fully
+            # present model read as "partial" → the bogus "Resume model /
+            # one-time download" caption on a model already on disk.
+            from ...bootstrap import _is_complete_download
             cfg = getattr(self.app, "config", None)
             chosen = current_choice(cfg) if cfg is not None else None
-            if chosen is not None:
-                verb = "Download"
-                # huggingface_hub writes `.incomplete` sidecars; our urllib
-                # fallback pre-allocates the final filename. Detect both.
-                partial = chosen.local_path
-                incomplete = partial.with_name(partial.name + ".incomplete")
-                has_partial = (
-                    (partial.exists() and partial.stat().st_size > 0
-                     and partial.stat().st_size < int(chosen.size_gb * 1024 ** 3 * 0.99))
-                    or incomplete.exists()
-                )
-                if has_partial:
-                    verb = "Resume"
-                model_label = f"{verb} model (~{chosen.size_gb:.1f} GB)"
-            else:
-                model_label = "Download model"
+            if chosen is None:
+                return ("absent", None)
+            final = chosen.local_path
+            if _is_complete_download(final, chosen):
+                return ("present", chosen.size_gb)
+            # huggingface_hub writes `.incomplete` sidecars; our urllib
+            # fallback pre-allocates the final filename. Either → a resumable
+            # partial. Otherwise nothing is on disk yet.
+            incomplete = final.with_name(final.name + ".incomplete")
+            has_partial = (
+                (final.exists() and final.stat().st_size > 0) or incomplete.exists()
+            )
+            return ("partial" if has_partial else "absent", chosen.size_gb)
         except Exception:
+            return ("absent", None)
+
+    def _render_steps(self) -> str:
+        # Substitute the actually-selected model's size into the model step
+        # label, and pick a verb that matches what's actually happening: a
+        # fully-present model is only *loaded* (not downloaded), a partial is
+        # *resumed*, and a missing one is *downloaded*.
+        state, size_gb = self._model_state()
+        if size_gb is None:
             model_label = "Download model"
+        elif state == "present":
+            model_label = f"Model ready (~{size_gb:.1f} GB)"
+        elif state == "partial":
+            model_label = f"Resume download (~{size_gb:.1f} GB)"
+        else:
+            model_label = f"Download model (~{size_gb:.1f} GB)"
 
         lines = ["[bold]Setup[/]\n"]
         for i, (key, label) in enumerate(self.STEPS):
@@ -321,6 +344,22 @@ class SetupScreen(Screen):
             steps = self._render_steps()
             self.query_one("#setup-status", Static).update(status)
             self.query_one("#setup-steps", Static).update(steps)
+            # Only show the download-framing captions when a download is
+            # actually going to happen. A fully-present model is just being
+            # loaded into memory — "one-time download" / "resumes where it
+            # left off" would be a lie there.
+            if self._model_state()[0] == "present":
+                self.query_one("#setup-onetime-note", Static).update(
+                    "[dim italic]loading into memory · first launch after boot is slower[/]"
+                )
+                self.query_one("#setup-quit-hint", Static).update("")
+            else:
+                self.query_one("#setup-onetime-note", Static).update(
+                    "[dim italic]one-time download · cached for future launches[/]"
+                )
+                self.query_one("#setup-quit-hint", Static).update(
+                    "[dim]Esc to quit — the download resumes where it left off next launch.[/]"
+                )
             self.refresh()
         except Exception:
             pass

{localcode-0.3.42 → localcode-0.3.44/src/localcode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.42
+Version: 0.3.44
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0

{localcode-0.3.42 → localcode-0.3.44}/src/localcode.egg-info/SOURCES.txt

@@ -958,6 +958,7 @@ tests/test_delete_command.py
 tests/test_download.py
 tests/test_edit_guard_and_fuzzy.py
 tests/test_embeddings_retrieval.py
+tests/test_ephemeral_context_role.py
 tests/test_evidence_and_lifecycle_hooks.py
 tests/test_execution_policies.py
 tests/test_fresh_install.py

localcode-0.3.44/tests/test_ephemeral_context_role.py

@@ -0,0 +1,36 @@
+"""Regression: the trailing per-round context block (ledger + fs-state + todo)
+must NOT be sent as a `system` message to Qwen-family models.
+
+Qwen 3.x's chat template raises "System message must be at the beginning" on
+any non-leading system message; llama-server surfaces that as HTTP 500 on every
+round once the block is non-empty, crashing the session (0.3.42 shipped Qwen
+3.8 with this latent bug — a single-tool task never populated the block, so it
+was missed until real multi-turn use). Gemma keeps `system` (a user turn made
+it re-greet every round).
+"""
+
+import pytest
+
+from localcode.agent.loop import ephemeral_context_role
+
+
+@pytest.mark.parametrize(
+    "family, expected",
+    [
+        ("qwen35", "user"),        # Qwen 3.8 27B (dense hybrid)
+        ("qwen35moe", "user"),     # Qwen 3.6 35B-A3B (MoE) — same template lineage
+        ("QWEN35", "user"),        # case-insensitive
+        ("gemma4-iswa", "system"), # Gemma keeps system (re-greets as user)
+        ("cohere2_moe", "system"),
+        ("", "system"),            # unknown / falsy → historical default
+        (None, "system"),
+    ],
+)
+def test_ephemeral_context_role(family, expected):
+    assert ephemeral_context_role(family) == expected
+
+
+def test_qwen_never_gets_a_trailing_system_message():
+    """The whole point: for any Qwen family string, the role is never 'system'."""
+    for fam in ("qwen35", "qwen35moe", "qwen3", "qwen-anything"):
+        assert ephemeral_context_role(fam) != "system"