localcode 0.5.2__tar.gz → 0.5.3__tar.gz

{localcode-0.5.2/src/localcode.egg-info → localcode-0.5.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.5.2
+Version: 0.5.3
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0
@@ -39,6 +39,11 @@ Requires-Dist: pywhispercpp>=1.2.0; extra == "voice"
 Requires-Dist: sounddevice>=0.4.6; extra == "voice"
 Requires-Dist: numpy>=1.26.0; extra == "voice"
 Requires-Dist: piper-tts>=1.2.0; extra == "voice"
+Provides-Extra: dev
+Requires-Dist: pytest; extra == "dev"
+Requires-Dist: pytest-timeout; extra == "dev"
+Requires-Dist: ruff; extra == "dev"
+Requires-Dist: build; extra == "dev"
 Dynamic: license-file
 
 <p align="center">

{localcode-0.5.2 → localcode-0.5.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "localcode"
-version = "0.5.2"
+version = "0.5.3"
 description = "High-performance AI coding on consumer hardware."
 readme = "README.md"
 requires-python = ">=3.10"
@@ -52,6 +52,13 @@ voice = [
   "numpy>=1.26.0",
   "piper-tts>=1.2.0",         # natural neural TTS — default speaker
 ]
+# What the pre-PR checks in CONTRIBUTING.md need.
+dev = [
+  "pytest",
+  "pytest-timeout",
+  "ruff",
+  "build",
+]
 
 
 [project.urls]
@@ -77,6 +84,7 @@ where = ["src"]
     "bin/localcode-ui",
     "ui/plugin/*.ts",
     "ui/FORK_COMMIT",
+    "ui/UPSTREAM_COMMIT",
     "skills/*.md",
 ]
 
@@ -98,3 +106,8 @@ markers = [
 [tool.ruff]
 line-length = 100
 target-version = "py311"
+
+# The gate CI enforces: syntax errors, broken comparisons, undefined names.
+# Pinned here so a bare `ruff check src tests` checks the same thing locally.
+[tool.ruff.lint]
+select = ["E9", "F63", "F7", "F82"]

{localcode-0.5.2 → localcode-0.5.3}/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__ = "0.5.2"
+    __version__ = "0.5.3"

localcode-0.5.3/src/localcode/bin/localcode-ui

@@ -0,0 +1,4 @@
+[diffend] Oversized file quarantined before diffing.
+name: localcode-0.5.3/src/localcode/bin/localcode-ui
+size: 102460976 bytes
+sha256: df29f021a70948fdc0b0a5960858ea3e83a72d18d803f9ea17d01004cf1ce319

{localcode-0.5.2 → localcode-0.5.3}/src/localcode/config.py

@@ -129,7 +129,7 @@ class RuntimeConfig:
     llama_cpp_expert_offload: bool = False # offload MoE experts to CPU (-ot exps=CPU)
     llama_cpp_draft_model: str = ""       # path to draft GGUF for speculative decoding
     llama_cpp_lookup_cache: bool = False   # prompt lookup decoding (2-4x on code edits)
-    kv_cache_type_k: str = "q8_0"          # K cache type: q8_0, q4_0, f16, turbo2, turbo3, turbo4
+    kv_cache_type_k: str = "q8_0"          # K cache type: q8_0, q4_0, f16, turbo3, turbo4 (turbo2 is V-only; as K it runs as q8_0)
     kv_cache_type_v: str = "q8_0"          # V cache type: q8_0 default; q4_0, f16, turbo2, turbo3, turbo4 available
     llama_cpp_cache_reuse: int = 256       # --cache-reuse N: reuse KV chunks across partial prefix matches (0 = off). Recovers prefix-cache hits after mid-context edits/compaction shift the tail; the stable system-prompt prefix is already reused automatically per slot.
     llama_cpp_binary: str = ""             # custom llama-server path (e.g. TurboQuant fork)

{localcode-0.5.2 → localcode-0.5.3}/src/localcode/model_families.py

@@ -251,6 +251,9 @@ def infer_family_from_profile(profile_id: str) -> ModelFamily:
     # Hermes <tool_call> JSON. Same adapter.
     if "kolibri" in low:
         return ModelFamily.QWEN
+    # OpenJev is a Qwen 3.5 fine-tune and keeps its chat template.
+    if "openjev" in low:
+        return ModelFamily.QWEN
     if low.startswith("llama") or "llama" in low:
         return ModelFamily.LLAMA
     if low.startswith("deepseek") or "deepseek" in low:

{localcode-0.5.2 → localcode-0.5.3}/src/localcode/models_catalog.py

@@ -450,6 +450,35 @@ CHOICES: list[ModelChoice] = [
         # tokens, empty reasoning channel). That is the default policy
         # ("chat_template"), the same one the Gemma 4 and Qwen entries use.
     ),
+    ModelChoice(
+        key="openjev",
+        # Pinned to the revision and digest that were verified live.
+        revision="10840f375658dea7afc5ff4711127bca8218b560",
+        sha256="38b512277edaeec6cd251d146bff6d97bc328fa9db19755401d6e9ab1ccbcba5",
+        size_bytes=18973872288,
+        name="OpenJev 27B (Q4, decision model)",
+        hf_repo="ggml-org/OpenJev-GGUF",
+        filename="OpenJev-Q4_K_M.gguf",
+        size_gb=19.0,
+        active_params="27B dense (Qwen 3.5 base, tuned for typed decisions)",
+        architecture="qwen35",
+        license="CC BY-NC 4.0 (non-commercial use only)",
+        humaneval_pass_at_1=None,
+        notes=(
+            "A DECISION model, not a coding model. It answers typed questions "
+            "about a state (text, JSON, a web page, a screenshot) with a choice, "
+            "a yes/no probability or a score, in one forward pass per question, "
+            "through the bundled server's /v1/systemone endpoint; `localcode api` "
+            "prints the endpoint and key. It also answers on the chat endpoint, "
+            "but it was tuned for decisions and is not benchmarked for coding. "
+            "The weights are CC BY-NC 4.0: research and non-commercial use only. "
+            "Needs ~32 GB unified memory. Pair with the mmproj for screenshot "
+            "input. Never auto-recommended."
+        ),
+        mmproj_filename="mmproj-OpenJev-Q8_0.gguf",
+        mmproj_size_gb=0.6,
+        mmproj_hf_filename="mmproj-OpenJev-Q8_0.gguf",
+    ),
 ]
 
 
@@ -488,6 +517,14 @@ def _system_ram_gb() -> int:
 # it is not the coding-agent experience a first-run user should land on by
 # default. Owner can override by emptying this set.
 _NO_AUTO_RECOMMEND_ARCHS = {"diffusion_gemma", "kolibri1"}
+# Entries excluded by key, for models that share an architecture with a
+# recommended one: OpenJev is a qwen35 like Qwen 3.8, but it is a decision
+# model and must never be a first-run coding default.
+_NO_AUTO_RECOMMEND_KEYS = {"openjev"}
+
+
+def _auto_recommendable(c: "ModelChoice") -> bool:
+    return c.architecture not in _NO_AUTO_RECOMMEND_ARCHS and c.key not in _NO_AUTO_RECOMMEND_KEYS
 
 # Capability order for auto-recommend, best → worst for coding-agent use. This
 # is deliberately NOT raw file size: the big MoEs measure ~95% HumanEval here
@@ -526,14 +563,14 @@ def recommend(ram_gb: int | None = None) -> ModelChoice:
     budget = ram_gb * 0.55
     candidates = [
         c for c in CHOICES
-        if c.architecture not in _NO_AUTO_RECOMMEND_ARCHS and c.size_gb <= budget
+        if _auto_recommendable(c) and c.size_gb <= budget
     ]
     if candidates:
         # Most capable that fits; tie-break toward the larger (better-quant) file.
         return min(candidates, key=lambda c: (_capability_rank(c), -c.size_gb))
     # Nothing fits the budget — smallest production-ready model so the user still
     # gets something runnable rather than an impossible recommendation.
-    prod = [c for c in CHOICES if c.architecture not in _NO_AUTO_RECOMMEND_ARCHS]
+    prod = [c for c in CHOICES if _auto_recommendable(c)]
     return min(prod or CHOICES, key=lambda c: c.size_gb)
 
 
@@ -785,6 +822,23 @@ MODEL_GROUPS: list[ModelGroup] = [
             "Text-only."
         ),
     ),
+    ModelGroup(
+        key="openjev",
+        display_name="OpenJev 27B (decision model)",
+        maker="OpenJev",
+        hf_repo="ggml-org/OpenJev-GGUF",
+        family="qwen",
+        architecture="qwen35",
+        license="CC BY-NC 4.0 (non-commercial use only)",
+        notes=(
+            "Decision model, not a coding model: typed choice / yes-no / score "
+            "answers about text, web pages and screenshots through /v1/systemone. "
+            "Q4_K_M, Q8_0 and BF16 quants. Non-commercial licence."
+        ),
+        mmproj_filename="mmproj-OpenJev-Q8_0.gguf",
+        mmproj_size_gb=0.6,
+        mmproj_hf_filename="mmproj-OpenJev-Q8_0.gguf",
+    ),
 ]
 
 

{localcode-0.5.2 → localcode-0.5.3}/src/localcode/runtime.py

@@ -466,6 +466,12 @@ class LocalCodeRuntimeGateway:
         # KV cache compression (asymmetric: q8_0 K + turbo4 V recommended)
         ctk = self.config.kv_cache_type_k
         ctv = self.config.kv_cache_type_v
+        # turbo2 is a VALUE type only. Two bits per key leaves each key about a
+        # third off its true direction, which attention cannot absorb: measured
+        # perplexity roughly doubles, identically on the CPU reference and on
+        # Metal (llama-cpp-turboquant/PATCHES.md). Keys fall back to q8_0.
+        if (ctk or "").lower() == "turbo2":
+            ctk = "q8_0"
         if ctk and ctk != "f16":
             cmd.extend(["--cache-type-k", ctk])
         if ctv and ctv != "f16":

localcode-0.5.3/src/localcode/ui/UPSTREAM_COMMIT

@@ -0,0 +1 @@
+907b3bc518fa48e90e8ec24dd327d13eee71c36c

{localcode-0.5.2 → localcode-0.5.3}/src/localcode/ui/auth.py

@@ -92,3 +92,14 @@ def request_allowed(headers, expected_token: str, port: int, *, want_json: bool
         if ctype != "application/json":
             return "Content-Type must be application/json"
     return None
+
+
+def redact_command(command: list[str]) -> list[str]:
+    """Keep the server key out of diagnostic logs."""
+    redacted = list(command)
+    for i, value in enumerate(redacted):
+        if value == "--api-key" and i + 1 < len(redacted):
+            redacted[i + 1] = "[redacted]"
+        elif value.startswith("--api-key="):
+            redacted[i] = "--api-key=[redacted]"
+    return redacted

{localcode-0.5.2 → localcode-0.5.3}/src/localcode/ui/plugin/localcode.ts

@@ -635,7 +635,7 @@ const LocalcodePlugin: Plugin = async ({ client, directory }) => {
           if (output.system[i].includes("__pending__")) output.system[i] = output.system[i].replaceAll("__pending__", loaded);
         }
       }
-      if (input.agent && input.agent !== "build") return;
+      const isBuild = !input.agent || input.agent === "build";
       // The system prompt must be IDENTICAL for every request of a session: it
       // is the cached prefix. This rule used to be added only once
       // workspaceActive flipped (after the first tool call of each turn), so the
@@ -643,7 +643,14 @@ const LocalcodePlugin: Plugin = async ({ client, directory }) => {
       // once per turn (15-23 s at 30k tokens). It is inert in plain chat (the
       // gates check workspaceActive in code), so always add it. The open-todo
       // list is not put here either: it rides on the user turn (chat.message).
-      output.system.push(PLANNING_RULE);
+      if (isBuild) output.system.push(PLANNING_RULE);
+      // The runtime maps every entry to a separate system message. OpenJev's
+      // template accepts exactly one, at the beginning; a second one raises
+      // "System message must be at the beginning" before inference. Preserve
+      // every instruction in order in a single prefix for every agent.
+      if (output.system.length > 1) {
+        output.system.splice(0, output.system.length, output.system.join("\n\n"));
+      }
     },
 
     "chat.message": async (input, output) => {

{localcode-0.5.2 → localcode-0.5.3}/src/localcode/ui/server_cmd.py

@@ -41,6 +41,20 @@ def server_command(gguf: str, port: int, alias: str | None = None) -> list[str]:
         choice = by_filename(Path(gguf).name)
         if choice is not None:
             control = choice.reasoning_control or "server"
+        # The capability registry has the last word. DiffusionGemma's catalog
+        # entry carries the default policy, but the model has no thinking
+        # toggle: with --reasoning off it ends generation at canvas position 0,
+        # so short answers come back empty and long ones lose their first
+        # words (found by the launcher-driven model gate, 2026-10-04).
+        from localcode.reasoning_capabilities import ReasoningControl, reasoning_capabilities
+        capabilities = reasoning_capabilities(Path(gguf).name)
+        if capabilities.control is ReasoningControl.NONE:
+            control = "none"
+        if capabilities.family == "diffusion":
+            # Auto extraction can put the entire post-tool answer in
+            # reasoning_content, leaving no visible text. Keep generated text
+            # in content without disabling this model's generation path.
+            cmd += ["--reasoning-format", "none"]
     except Exception:  # noqa: BLE001, S110
         pass
     mode = os.environ.get("LOCALCODE_INTERNAL_THINKING_MODE") or cfg.runtime.internal_thinking_mode or "off"

{localcode-0.5.2 → localcode-0.5.3}/src/localcode/ui/supervisor.py

@@ -203,7 +203,8 @@ class Supervisor:
         cmd[0] = self.server_bin
         self.ctx = self.ctx_total = int(cmd[cmd.index("--ctx-size") + 1])
         self.slots = int(cmd[cmd.index("--parallel") + 1]) if "--parallel" in cmd else 1
-        self.log.write(f"\n=== {time.ctime()} {' '.join(cmd)}\n".encode())
+        from localcode.ui.auth import redact_command
+        self.log.write(f"\n=== {time.ctime()} {' '.join(redact_command(cmd))}\n".encode())
         with self._budget_lock:
             from localcode.ui.context_budget import TimingTable
             self._timing = TimingTable()

{localcode-0.5.2 → localcode-0.5.3/src/localcode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.5.2
+Version: 0.5.3
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0
@@ -39,6 +39,11 @@ Requires-Dist: pywhispercpp>=1.2.0; extra == "voice"
 Requires-Dist: sounddevice>=0.4.6; extra == "voice"
 Requires-Dist: numpy>=1.26.0; extra == "voice"
 Requires-Dist: piper-tts>=1.2.0; extra == "voice"
+Provides-Extra: dev
+Requires-Dist: pytest; extra == "dev"
+Requires-Dist: pytest-timeout; extra == "dev"
+Requires-Dist: ruff; extra == "dev"
+Requires-Dist: build; extra == "dev"
 Dynamic: license-file
 
 <p align="center">

{localcode-0.5.2 → localcode-0.5.3}/src/localcode.egg-info/SOURCES.txt

@@ -1190,6 +1190,7 @@ src/localcode/skills/plan-task.md
 src/localcode/skills/review.md
 src/localcode/skills/run-tests.md
 src/localcode/ui/FORK_COMMIT
+src/localcode/ui/UPSTREAM_COMMIT
 src/localcode/ui/__init__.py
 src/localcode/ui/api.py
 src/localcode/ui/auth.py
@@ -1210,7 +1211,9 @@ tests/test_comprehensive_diffusion.py
 tests/test_comprehensive_download.py
 tests/test_comprehensive_install.py
 tests/test_comprehensive_machines.py
+tests/test_comprehensive_model_gate.py
 tests/test_comprehensive_models_matrix.py
+tests/test_comprehensive_upstream_gates.py
 tests/test_comprehensive_voice.py
 tests/test_comprehensive_whisper.py
 tests/test_config.py

{localcode-0.5.2 → localcode-0.5.3}/src/localcode.egg-info/requires.txt

@@ -13,6 +13,12 @@ sounddevice>=0.4.6
 numpy>=1.26.0
 piper-tts>=1.2.0
 
+[dev]
+pytest
+pytest-timeout
+ruff
+build
+
 [voice]
 pywhispercpp>=1.2.0
 sounddevice>=0.4.6

{localcode-0.5.2 → localcode-0.5.3}/tests/test_comprehensive_binary.py

@@ -153,3 +153,14 @@ def test_bundled_binary_second_launch_is_fast(binary: Path):
     _version(binary)
     proc = _version(binary, timeout=30)
     assert proc.returncode == 0, proc.stderr[-400:]
+
+
+def test_turbo2_keys_fall_back_to_q8_0():
+    """turbo2 is a value-only cache type; as the key type the launcher uses q8_0."""
+    from localcode.config import RuntimeConfig
+    from localcode.runtime import LocalCodeRuntimeGateway
+
+    cfg = RuntimeConfig(kv_cache_type_k="turbo2", kv_cache_type_v="turbo2")
+    cmd = LocalCodeRuntimeGateway(cfg).llama_server_command("/path/model.gguf", 8081)
+    assert cmd[cmd.index("--cache-type-k") + 1] == "q8_0"
+    assert cmd[cmd.index("--cache-type-v") + 1] == "turbo2"

localcode-0.5.3/tests/test_comprehensive_model_gate.py

@@ -0,0 +1,77 @@
+"""Contract failures which the local promotion gate must reject."""
+import json
+import subprocess
+from unittest.mock import Mock
+
+import pytest
+
+from scripts import verify_models as gate
+
+
+def test_tool_loop_rejects_text_instead_of_a_call():
+    with pytest.raises(AssertionError, match="no tool call"):
+        gate.validate_tool({"content": "I would use get_weather"})
+
+
+@pytest.mark.parametrize("arguments", ['{"city":"London"}', '{}', 'not JSON'])
+def test_tool_loop_rejects_wrong_or_invalid_arguments(arguments):
+    with pytest.raises((AssertionError, json.JSONDecodeError)):
+        gate.validate_tool({"tool_calls": [{"id": "call_1", "function": {"name": "get_weather", "arguments": arguments}}]})
+
+
+def test_decision_gate_rejects_nonfinite_probabilities():
+    data = {"answers": {"route": {"choice": "billing", "probabilities": {"billing": float("nan"), "shipping": 0, "technical": 0}}, "angry": {"noul": 0.5}, "urgency": {"score": 1}}}
+    with pytest.raises(AssertionError):
+        gate.validate_decisions(data)
+
+
+def test_missing_required_models_fail_instead_of_skipping(monkeypatch):
+    monkeypatch.setattr(gate.catalog, "CHOICES", [])
+    with pytest.raises(SystemExit) as exc:
+        gate.main(["--strict"])
+    assert exc.value.code == 2
+
+
+def test_environment_restores_credentials(monkeypatch):
+    monkeypatch.setenv(gate.auth.KEY_ENV, "original")
+    with gate.environment({gate.auth.KEY_ENV: "temporary"}):
+        assert gate.auth.server_key() == "temporary"
+    assert gate.auth.server_key() == "original"
+
+
+def test_failed_startup_reaps_the_process(monkeypatch, tmp_path):
+    choice = Mock(key="gemma", local_path=tmp_path / "model.gguf")
+    proc = Mock()
+    proc.poll.side_effect = [1, None]
+    monkeypatch.setattr(gate, "server_command", lambda *a: ["llama-server", "--ctx-size", "4096"])
+    monkeypatch.setattr(gate.subprocess, "Popen", lambda *a, **kw: proc)
+    with pytest.raises(RuntimeError, match="exited during model load"):
+        with gate.Server(choice, tmp_path):
+            pass
+    proc.terminate.assert_called_once()
+    proc.wait.assert_called_once_with(timeout=15)
+
+
+def test_shutdown_kills_then_waits_after_timeout(tmp_path):
+    server = gate.Server(Mock(), tmp_path)
+    server.proc = Mock()
+    server.proc.poll.return_value = None
+    server.proc.wait.side_effect = [subprocess.TimeoutExpired("server", 15), 0]
+    server.__exit__(None, None, None)
+    server.proc.kill.assert_called_once()
+    assert server.proc.wait.call_count == 2
+
+
+@pytest.mark.parametrize("mutation", ["commit", "binaries", "coverage", "runtime", "turbo"])
+def test_invalid_receipts_cannot_publish_a_promotion(mutation):
+    rows = [{"key": key, "turbo4": False, "checks": ["auth", "template", "chat", "tool-loop", "systemone", "runtime"]} for key in gate.REQUIRED_KEYS]
+    rows.append({"key": "qwen", "turbo4": True, "checks": ["tool-loop"]})
+    receipt = {"commit": "candidate", "dirty": False, "strict": True, "success": True, "binaries": {"server": "hash"}, "models": rows}
+    gate.validate_receipt(receipt, "candidate", {"server": "hash"}, gate.REQUIRED_KEYS)
+    if mutation == "commit": receipt["commit"] = "old"
+    elif mutation == "binaries": receipt["binaries"] = {"server": "old"}
+    elif mutation == "coverage": receipt["models"] = rows[1:]
+    elif mutation == "runtime": next(r for r in rows if r["key"] == "openjev")["checks"].remove("runtime")
+    else: receipt["models"] = rows[:-1]
+    with pytest.raises(AssertionError):
+        gate.validate_receipt(receipt, "candidate", {"server": "hash"}, gate.REQUIRED_KEYS)

{localcode-0.5.2 → localcode-0.5.3}/tests/test_comprehensive_models_matrix.py

@@ -171,3 +171,22 @@ def test_every_catalog_reasoning_policy_resolves(choice):
     if choice.reasoning_control == "always":
         assert caps.control is ReasoningControl.ALWAYS
         assert not caps.supports_budget
+
+
+@pytest.mark.parametrize("choice", ALL_CHOICES, ids=_choice_ids)
+def test_launcher_only_silences_reasoning_where_a_switch_exists(choice, monkeypatch):
+    """The UI launcher may pass --reasoning off only to models that have the switch.
+
+    DiffusionGemma has none: with the flag it returns empty or head-truncated text.
+    """
+    from localcode.reasoning_capabilities import ReasoningControl, reasoning_capabilities
+    from localcode.ui.server_cmd import server_command
+
+    monkeypatch.delenv("LOCALCODE_INTERNAL_THINKING_MODE", raising=False)
+    cmd = server_command(f"/models/{choice.filename}", 8081, choice.key)
+    has_switch = reasoning_capabilities(choice.filename).control is ReasoningControl.CHAT_TEMPLATE
+    assert ("--reasoning" in cmd) == has_switch
+    if choice.architecture == "diffusion_gemma":
+        assert cmd[cmd.index("--reasoning-format") + 1] == "none"
+    else:
+        assert "--reasoning-format" not in cmd

localcode-0.5.3/tests/test_comprehensive_upstream_gates.py

@@ -0,0 +1,47 @@
+"""Updater regressions: fork ancestry and the boundary around writer credentials."""
+import os
+import subprocess
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def git(path, *args):
+    return subprocess.check_output(["git", "-C", str(path), *args], text=True, stderr=subprocess.DEVNULL).strip()
+
+
+def test_opencode_refuses_a_fork_behind_upstream_before_changing_the_pin(tmp_path):
+    upstream = tmp_path / "upstream"
+    upstream.mkdir()
+    git(upstream, "init", "-b", "dev")
+    git(upstream, "config", "user.name", "Test")
+    git(upstream, "config", "user.email", "test@example.invalid")
+    (upstream / "a").write_text("one")
+    git(upstream, "add", "a")
+    git(upstream, "commit", "-m", "initial")
+    fork = tmp_path / "fork"
+    subprocess.run(["git", "clone", str(upstream), str(fork)], check=True, capture_output=True)
+    git(fork, "branch", "localcode")
+    (upstream / "a").write_text("two")
+    git(upstream, "commit", "-am", "new upstream")
+    pin = ROOT / "src/localcode/ui/FORK_COMMIT"
+    old = pin.read_bytes()
+    result = subprocess.run(["bash", str(ROOT / "scripts/bump_opencode.sh")], env={**os.environ, "LOCALCODE_UI_FORK_REPO": str(fork), "LOCALCODE_OPENCODE_UPSTREAM_REPO": str(upstream), "LOCALCODE_UI_BUILD_DIR": str(tmp_path / "work")}, capture_output=True, text=True)
+    assert result.returncode == 20
+    assert "has not absorbed upstream" in result.stderr
+    assert pin.read_bytes() == old
+
+
+def test_upstream_build_does_not_keep_checkout_credentials():
+    workflow = (ROOT / ".github/workflows/upstream-candidate.yml").read_text()
+    build = workflow.split("  propose:", 1)[0]
+    assert "permissions:\n  contents: read" in build
+    assert "persist-credentials: false" in build
+    assert "GH_TOKEN:" not in build
+    assert "--include='src/localcode/ui/FORK_COMMIT'" in workflow
+    assert "--include='.github" not in workflow
+
+
+def test_no_workflow_runs_on_the_users_mac():
+    for path in (ROOT / ".github/workflows").glob("*.yml"):
+        assert "runs-on: self-hosted" not in path.read_text()

{localcode-0.5.2 → localcode-0.5.3}/tests/test_models_catalog.py

@@ -84,6 +84,7 @@ _EXPECTED_VISION = {
     "diffusiongemma": False,
     "north-mini-code": False,
     "kolibri": False,
+    "openjev": True,
 }
 
 
@@ -278,3 +279,29 @@ def test_kolibri_catalog_entry_is_pinned_and_never_auto_recommended() -> None:
     g = group_for_filename("Kolibri-1-Q4_K_M.gguf")
     assert g is not None and g.architecture == "kolibri1"
     assert infer_family_from_profile("kolibri-1-q4") == ModelFamily.QWEN
+
+
+def test_openjev_is_pinned_listed_with_quants_and_never_auto_recommended() -> None:
+    from localcode.models_catalog import group_for_filename, recommend
+    from localcode.model_families import ModelFamily, infer_family_from_profile
+
+    choice = by_key("openjev")
+    assert choice is not None
+    assert choice.architecture == "qwen35"
+    assert choice.revision != "main" and len(choice.revision) == 40
+    assert choice.sha256 and len(choice.sha256) == 64
+    assert choice.size_bytes == 18973872288
+    assert choice.supports_vision
+    # the licence restriction must be visible wherever the entry is shown
+    assert "non-commercial" in choice.license.lower()
+    assert "decision" in choice.name.lower()
+
+    # a decision model shares qwen35 with Qwen 3.8, so it is excluded by key
+    for ram in (16, 32, 64, 96, 128, 192, 512):
+        assert recommend(ram).key != "openjev"
+
+    # the other quants in the repo resolve to the group
+    for quant in ("OpenJev-Q8_0.gguf", "OpenJev-BF16.gguf"):
+        g = group_for_filename(quant)
+        assert g is not None and g.key == "openjev"
+    assert infer_family_from_profile("openjev-q4") == ModelFamily.QWEN

localcode-0.5.3/tests/test_real_models.py

@@ -0,0 +1,22 @@
+"""Opt-in local inference using the shipping launcher and authenticated API.
+
+For update promotion use scripts/verify_models.py --strict; that fails if a
+required model is absent, covers turbo4 and the bundled OpenCode runtime, and
+writes a receipt. This pytest tier only exercises downloaded models.
+"""
+from pathlib import Path
+
+import pytest
+
+from scripts.verify_models import Server, catalog, verify
+
+pytestmark = pytest.mark.real_models
+CHOICES = [c for c in catalog.CHOICES if c.local_path.is_file()]
+
+
+@pytest.mark.parametrize("choice", CHOICES or [None], ids=lambda c: c.key if c else "no-models")
+def test_shipping_model_contract(choice, tmp_path: Path):
+    if choice is None:
+        pytest.skip("no models downloaded; promotion uses the strict CLI gate")
+    with Server(choice, tmp_path) as server:
+        verify(server, runtime=choice.key in {"qwen38", "openjev", "diffusiongemma"})

{localcode-0.5.2 → localcode-0.5.3}/tests/test_server_flags_accepted.py

@@ -44,18 +44,20 @@ pytestmark = pytest.mark.skipif(
 
 
 @pytest.mark.parametrize("mode", ["turbo", "turbo-think", "context", "speed"])
-def test_every_emitted_flag_is_accepted_by_the_bundled_binary(mode: str) -> None:
-    from localcode.config import RuntimeConfig
-    from localcode.runtime import LocalCodeRuntimeGateway
+def test_every_emitted_flag_is_accepted_by_the_bundled_binary(mode, monkeypatch, tmp_path):
+    from localcode.config import AppConfig, RuntimeConfig, UIConfig
+    from localcode.models_catalog import CHOICES
+    from localcode.ui import server_cmd
 
     known = _help_flags()
     assert "--ctx-size" in known, "could not parse llama-server --help"
-
-    cfg = RuntimeConfig(provider="llama_cpp", base_url="http://127.0.0.1:8081", model="model.gguf")
-    cfg.laptop_26b_runtime_mode = mode
-    cmd = LocalCodeRuntimeGateway(cfg).llama_server_command("/path/model.gguf", 8081)
-    # flags the UI launcher appends on top of the gateway command (ui/server_cmd.py)
-    cmd += ["--host", "--api-key", "--alias", "--reasoning", "--reasoning-budget", "--mmproj", "--slot-save-path"]
-
-    unknown = sorted({f for f in _emitted(cmd) if f not in known})
-    assert not unknown, f"llama-server no longer accepts: {unknown}"
+    cfg = AppConfig(runtime=RuntimeConfig(provider="llama_cpp"), ui=UIConfig())
+    cfg.runtime.laptop_26b_runtime_mode = mode
+    monkeypatch.setattr(server_cmd, "load_config", lambda: cfg)
+    monkeypatch.setenv("LOCALCODE_SERVER_KEY", "test-key")
+    monkeypatch.setenv("LOCALCODE_AGENT_RUN_DIR", str(tmp_path / "run"))
+    monkeypatch.setattr(server_cmd, "mmproj_for", lambda gguf: tmp_path / "projector.gguf")
+    for choice in CHOICES:
+        cmd = server_cmd.server_command(str(tmp_path / choice.filename), 8081, choice.key)
+        unknown = sorted({f for f in _emitted(cmd) if f not in known})
+        assert not unknown, f"{choice.key}: llama-server no longer accepts: {unknown}"

{localcode-0.5.2 → localcode-0.5.3}/tests/test_ui_auth.py

@@ -163,3 +163,10 @@ def test_server_command_carries_the_session_key(monkeypatch, tmp_path):
     assert cmd[cmd.index("--host") + 1] == "127.0.0.1"
     monkeypatch.delenv(auth.KEY_ENV)
     assert "--api-key" not in server_cmd.server_command(str(tmp_path / "m.gguf"), 8123, "m")
+
+
+def test_server_key_is_redacted_from_diagnostics():
+    from localcode.ui.auth import redact_command
+    command = ["server", "--api-key", "secret", "--port", "8081", "--api-key=other"]
+    assert redact_command(command) == ["server", "--api-key", "[redacted]", "--port", "8081", "--api-key=[redacted]"]
+    assert command[2] == "secret"

localcode-0.5.2/src/localcode/bin/localcode-ui

@@ -1,4 +0,0 @@
-[diffend] Oversized file quarantined before diffing.
-name: localcode-0.5.2/src/localcode/bin/localcode-ui
-size: 102460976 bytes
-sha256: a134cd79d1ba7dfe34652b68bf2dc7bbcd93a4644bb350d7d3753b6f9accef14