localcode 0.2.15__tar.gz → 0.2.16__tar.gz

{localcode-0.2.15/src/localcode.egg-info → localcode-0.2.16}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.2.15
+Version: 0.2.16
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0
@@ -54,7 +54,7 @@ Dynamic: license-file
 <p align="center">
   <img src="https://img.shields.io/pypi/v/localcode?style=flat-square&color=7c4dff" alt="PyPI">
   <img src="https://img.shields.io/badge/license-Apache_2.0-4caf50?style=flat-square" alt="License">
-  <img src="https://img.shields.io/badge/python-3.11+-3776ab?style=flat-square" alt="Python">
+  <img src="https://img.shields.io/badge/python-3.9+-3776ab?style=flat-square" alt="Python">
   <img src="https://img.shields.io/badge/platform-Apple%20Silicon-999999?style=flat-square" alt="Platform">
 </p>
 
@@ -108,7 +108,7 @@ We are building for a world of truly democratized AI — where everyone has acce
 
 - **Mac with Apple Silicon**
 - **16 GB RAM** minimum
-- **Python 3.11+**
+- **Python 3.9+**
 - **~12 GB free disk** (10 GB model + server)
 
 ### Tested hardware
@@ -122,18 +122,35 @@ LocalCode is early software. Hardware support is expected to broaden, but only t
 | 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 |
 
+## Models
+
+On launch, LocalCode recommends the best model for **your Mac's RAM** — there's no fixed default. You can pick any of these (or a different quant) in the model picker.
+
+| Model | Size (quant) | Active params | Min RAM | Architecture |
+| --- | ---: | --- | ---: | --- |
+| Gemma 4 12B | 7.4 GB (Q4) | 12B (dense) | 16 GB | gemma4-iswa |
+| 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 |
+| 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 3.6 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).
+
 ## How LocalCode works
 
 LocalCode runs a custom [llama.cpp](https://github.com/ggerganov/llama.cpp) fork with **TurboQuant KV cache compression** — a technique from Google's ICLR 2026 paper that we patched into llama.cpp for Apple Silicon. This compresses the KV cache 3.8× — fitting 32K context in 355 MiB on a 16 GB MacBook.
 
-The default model (**Gemma 4 26B-A4B**) is a Mixture-of-Experts architecture — 25.2 B total parameters but only 3.8 B active per token. That's what makes ~27 tok/s possible on a laptop.
+LocalCode picks a model based on **your Mac's RAM** — there's no fixed default. It scales from Gemma 4 12B on 16 GB up to Qwen 3.6 35B-A3B on 64 GB+. The recommended models are Mixture-of-Experts — only ~3.8 B parameters active per token — which is what makes ~27 tok/s possible on a laptop.
 
 Under the hood:
 
 - **TurboQuant KV cache** — asymmetric q8\_0-K + turbo4-V quantization, 3.8× compression vs. f16
 - **Multi-region mmap patch** — fixes a Metal OOM crash where llama.cpp's loader spanned the entire GGUF file into one Metal buffer
 - **GPU memory unlock** — auto-prompts to raise `iogpu.wired_limit_mb` for full Metal offload
-- **Agent loop** — phased CREATE/EDIT/CHAT/RUN/SEARCH routing with task state, evidence-driven completion, and recovery modes for small-model failure patterns
+- **Agent loop** — goal-typed routing (build / edit / run / chat) with task state, evidence-driven completion, and recovery modes for small-model failure patterns
 
 ## Sponsors
 

{localcode-0.2.15 → localcode-0.2.16}/README.md

@@ -9,7 +9,7 @@
 <p align="center">
   <img src="https://img.shields.io/pypi/v/localcode?style=flat-square&color=7c4dff" alt="PyPI">
   <img src="https://img.shields.io/badge/license-Apache_2.0-4caf50?style=flat-square" alt="License">
-  <img src="https://img.shields.io/badge/python-3.11+-3776ab?style=flat-square" alt="Python">
+  <img src="https://img.shields.io/badge/python-3.9+-3776ab?style=flat-square" alt="Python">
   <img src="https://img.shields.io/badge/platform-Apple%20Silicon-999999?style=flat-square" alt="Platform">
 </p>
 
@@ -63,7 +63,7 @@ We are building for a world of truly democratized AI — where everyone has acce
 
 - **Mac with Apple Silicon**
 - **16 GB RAM** minimum
-- **Python 3.11+**
+- **Python 3.9+**
 - **~12 GB free disk** (10 GB model + server)
 
 ### Tested hardware
@@ -77,18 +77,35 @@ LocalCode is early software. Hardware support is expected to broaden, but only t
 | 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 |
 
+## Models
+
+On launch, LocalCode recommends the best model for **your Mac's RAM** — there's no fixed default. You can pick any of these (or a different quant) in the model picker.
+
+| Model | Size (quant) | Active params | Min RAM | Architecture |
+| --- | ---: | --- | ---: | --- |
+| Gemma 4 12B | 7.4 GB (Q4) | 12B (dense) | 16 GB | gemma4-iswa |
+| 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 |
+| 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 3.6 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).
+
 ## How LocalCode works
 
 LocalCode runs a custom [llama.cpp](https://github.com/ggerganov/llama.cpp) fork with **TurboQuant KV cache compression** — a technique from Google's ICLR 2026 paper that we patched into llama.cpp for Apple Silicon. This compresses the KV cache 3.8× — fitting 32K context in 355 MiB on a 16 GB MacBook.
 
-The default model (**Gemma 4 26B-A4B**) is a Mixture-of-Experts architecture — 25.2 B total parameters but only 3.8 B active per token. That's what makes ~27 tok/s possible on a laptop.
+LocalCode picks a model based on **your Mac's RAM** — there's no fixed default. It scales from Gemma 4 12B on 16 GB up to Qwen 3.6 35B-A3B on 64 GB+. The recommended models are Mixture-of-Experts — only ~3.8 B parameters active per token — which is what makes ~27 tok/s possible on a laptop.
 
 Under the hood:
 
 - **TurboQuant KV cache** — asymmetric q8\_0-K + turbo4-V quantization, 3.8× compression vs. f16
 - **Multi-region mmap patch** — fixes a Metal OOM crash where llama.cpp's loader spanned the entire GGUF file into one Metal buffer
 - **GPU memory unlock** — auto-prompts to raise `iogpu.wired_limit_mb` for full Metal offload
-- **Agent loop** — phased CREATE/EDIT/CHAT/RUN/SEARCH routing with task state, evidence-driven completion, and recovery modes for small-model failure patterns
+- **Agent loop** — goal-typed routing (build / edit / run / chat) with task state, evidence-driven completion, and recovery modes for small-model failure patterns
 
 ## Sponsors
 

{localcode-0.2.15 → localcode-0.2.16}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "localcode"
-version = "0.2.15"
+version = "0.2.16"
 description = "High-performance AI coding on consumer hardware."
 readme = "README.md"
 requires-python = ">=3.9"

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/__init__.py

@@ -1,3 +1,3 @@
 __all__ = ["__version__"]
 
-__version__ = "0.2.15"
+__version__ = "0.2.16"

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/agent/__init__.py

@@ -16,7 +16,7 @@ re-exported here so external callers can continue to import from
 Background: before T0.1, agent.py was a 1792-line monolith. The split
 broke it into focused modules (each ≤ ~720 LoC), with __init__.py
 reduced to this re-export surface — well under the 400-LoC cap the
-plan sets for god modules (see eval/OPTIMIZATION_PLAN.md § T0).
+plan sets for god modules (see dev/eval/OPTIMIZATION_PLAN.md § T0).
 
 Unused-import warnings in this file are expected — every imported
 name is intentionally re-exported. The `# noqa: F401` comments
@@ -140,7 +140,7 @@ from .helpers import (  # noqa: F401
 
 # ── Prompt templates + project-instructions loader ────────────────────────
 # Moved to agent/prompts.py during the T0.1-b split. Re-exported here so
-# external callers (eval/prompt_variants.py, tests/promptfoo, app.py,
+# external callers (dev/eval/prompt_variants.py, tests/promptfoo, app.py,
 # tests/test_context_pipeline_e2e.py) that do
 #   `from localcode.agent import SYSTEM_PROMPT`
 # keep working unchanged. See agent/prompts.py for the commented
@@ -160,7 +160,6 @@ from .prompts import (  # noqa: F401 — re-exports for back-compat
 # and the dispatcher (for _execute_tool). Plan-mode gating still lives
 # here because it's cross-tool policy, not tool-specific logic.
 
-from ..tools import ALL_SCHEMAS as TOOL_SCHEMAS  # for agent-loop tool registry injection
 
 
 # ── Result Management ────────────────────────────────────────────────────

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/agent/helpers.py

@@ -53,16 +53,6 @@ __all__: list[str] = []
 _SHRINK_GUARD_SOURCE_EXTS = {".py", ".js", ".jsx", ".ts", ".tsx", ".css", ".html"}
 
 
-def _canonical_creation_root(app: "LocalCodeApp") -> Path | None:
-    task = getattr(getattr(app, "session", None), "current_task", None)
-    if getattr(task, "task_kind", "") != "new_app":
-        return None
-    slug = str(getattr(task, "task_slug", "") or "").strip()
-    if not slug:
-        return None
-    return (app.repo_root / slug).resolve()
-
-
 def _path_within(path: Path, root: Path) -> bool:
     try:
         path.resolve().relative_to(root)

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/agent/sections.py

@@ -14,14 +14,14 @@ Problems with the monolith:
     session, which means every new session re-evaluates the whole
     prompt. Our benchmark shows prompt_ms of ~48 ms warm; the
     theoretical floor with a stable prefix cache is ~5-10 ms.
-  • No way to A/B sections individually. `eval/prompt_variants.py`
+  • No way to A/B sections individually. `dev/eval/prompt_variants.py`
     only supports coarse "strip rule N" mutations. You can't
     answer "what if we drop the build-app rules for debugging
     tasks?" because there's no boundary between them.
   • No way to gate sections on trigger phrases. Every session ships
     rule 10 (open browser) whether or not a web app is in scope.
 
-This module is **Phase 1** of T0.11 (see eval/OPTIMIZATION_PLAN.md):
+This module is **Phase 1** of T0.11 (see dev/eval/OPTIMIZATION_PLAN.md):
 introduce the section-registry concept with behaviour-neutral
 defaults. The composer produces the same output as the old
 `SYSTEM_PROMPT.format(...)` for the default all-sections-on case.
@@ -55,11 +55,9 @@ work is an add-only operation.
 from __future__ import annotations
 
 from dataclasses import dataclass
-from pathlib import Path
 from typing import Callable
 
 from .prompts import SYSTEM_PROMPT
-from .prompts import _load_project_instructions
 
 
 __all__ = [
@@ -241,10 +239,3 @@ def compose_system_prompt(
         parts.insert(-1, CACHE_BOUNDARY_MARKER)
 
     return "".join(parts)
-
-
-def load_project_instructions(repo_root: Path) -> str:
-    """Convenience re-export so callers outside agent/ that don't
-    want to depend on `agent.prompts` internals can build a
-    SectionContext purely from this module."""
-    return _load_project_instructions(repo_root)

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/app.py

@@ -4,7 +4,6 @@ from concurrent.futures import ThreadPoolExecutor, as_completed
 from dataclasses import replace
 import logging
 from pathlib import Path
-import re
 
 from rich.console import Console
 from rich.panel import Panel
@@ -41,7 +40,7 @@ from .autonomy import AutonomyLevel, apply_autonomy_to_permissions, get_policy
 from .hooks import HookRunner
 from .snapshots import SnapshotStore, create_snapshot
 from .turn_diff import TurnDiffTracker, print_turn_diff
-from .agent.goal import GoalState, infer_goal_state
+from .agent.goal import infer_goal_state
 from .performance import detect_machine_profile, benchmark_report, apply_preset, should_promote_legacy_default_to_laptop_26b
 
 
@@ -62,67 +61,6 @@ _FOLLOWUP_CHAT_ONLY = {
 _CODING_TASK_KINDS = {"new_app", "existing_app_edit", "run_or_launch"}
 
 
-def _looks_like_task_followup(user_text: str, current_task: object | None) -> bool:
-    """Return true when a short correction should continue the task.
-
-    This is deliberately conservative: it only attaches to a coding task
-    and only for short correction/addition wording. The actual feature
-    remains model/tool-driven; LocalCode does not classify domain terms.
-    """
-    if current_task is None:
-        return False
-    if getattr(current_task, "task_kind", "") not in _CODING_TASK_KINDS:
-        return False
-    text = (user_text or "").strip().lower()
-    if not text:
-        return False
-    first = next(iter(re.findall(r"[a-z0-9]+", text)), "")
-    if text in _FOLLOWUP_CHAT_ONLY:
-        return False
-    if len(text) <= 80 and "?" not in text and first not in {"what", "why", "how", "where", "when", "who"}:
-        return True
-    return any(
-        phrase in text
-        for phrase in (
-            "i meant",
-            "you forgot",
-            "missing",
-            "add ",
-            "also add",
-            "it should",
-            "should have",
-            "doesn't have",
-            "didn't add",
-        )
-    )
-
-
-def _continue_goal_for_task_followup(goal_state: GoalState, user_text: str, current_task: object) -> GoalState:
-    criteria = list(getattr(current_task, "success_criteria", []) or goal_state.success_criteria)
-    criteria.extend([
-        "Follow-up correction is implemented in the existing task context",
-        "Assistant does not stop with a permission question when the requested implementation is local and feasible",
-    ])
-    current_kind = getattr(current_task, "task_kind", goal_state.task_kind)
-    if current_kind in {"new_app", "existing_app_edit"}:
-        goal_type = "edit_existing"
-        task_kind = "existing_app_edit"
-    else:
-        goal_type = getattr(current_task, "goal_type", goal_state.goal_type)
-        task_kind = current_kind
-    # Preserve task identity; this turn is a correction on existing work,
-    # not a fresh task name.
-    return replace(
-        goal_state,
-        goal_type=goal_type,
-        task_kind=task_kind,
-        task_slug=getattr(current_task, "task_slug", goal_state.task_slug),
-        goal_summary=f"{getattr(current_task, 'goal_summary', '')}\nFollow-up: {user_text}".strip()[:240],
-        success_criteria=tuple(dict.fromkeys(criteria)),
-        allows_blocking_question=False,
-    )
-
-
 def _canonical_project_dir_has_files(repo_root: Path, slug: str) -> bool:
     """True when a creation task's canonical project directory already exists.
 

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/autonomy.py

@@ -101,18 +101,3 @@ def apply_autonomy_to_permissions(permissions, policy: AutonomyPolicy) -> None:
 
     if policy.level == AutonomyLevel.FULL_AUTO:
         permissions.approve_all()
-
-
-def format_autonomy_status(level: AutonomyLevel) -> str:
-    """Format current autonomy level for display."""
-    policy = POLICIES[level]
-    icons = {True: "✓", False: "✗"}
-    return (
-        f"Autonomy: {level.value}\n"
-        f"  {policy.description}\n"
-        f"  reads:    {icons[policy.auto_approve_reads]}\n"
-        f"  writes:   {icons[policy.auto_approve_writes]}\n"
-        f"  bash:     {icons[policy.auto_approve_bash]}\n"
-        f"  installs: {icons[policy.auto_approve_installs]}\n"
-        f"  git:      {icons[policy.auto_approve_git]}"
-    )

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/bootstrap.py

@@ -539,13 +539,16 @@ def download_model(
 ) -> tuple[bool, str]:
     """Download a GGUF model selected from `models_catalog.CHOICES`.
 
-    If `choice` is None, defaults to the first entry in CHOICES (Gemma 4 26B).
+    If `choice` is None, falls back to `recommend()` — the RAM-appropriate model
+    for this machine, not a hardcoded default.
     Returns (success, path_or_error_message).
     """
-    from .models_catalog import CHOICES
+    from .models_catalog import recommend
 
     if choice is None:
-        choice = CHOICES[0]
+        # No explicit choice → the RAM-appropriate model for THIS machine,
+        # not a hardcoded default.
+        choice = recommend()
 
     # Already present?
     existing = get_model_path(choice.filename)
@@ -1247,18 +1250,6 @@ def _ensure_ollama() -> bool:
     return result.returncode == 0
 
 
-def ensure_model_downloaded(model_tag: str, on_progress: Callable[[str], None] | None = None) -> tuple[bool, str]:
-    """Ensure the model is downloaded via Ollama. Installs Ollama if needed."""
-    if not _ensure_ollama():
-        return False, "Could not install Ollama. Install manually: https://ollama.com/download"
-    # Start Ollama service if not running
-    if platform.system().lower() == "darwin":
-        subprocess.Popen(["ollama", "serve"], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
-                        start_new_session=True)
-        import time; time.sleep(2)
-    return pull_model(model_tag, on_progress=on_progress)
-
-
 def build_turboquant(on_progress: Callable[[str], None] | None = None) -> tuple[bool, str]:
     """Build the TurboQuant llama.cpp fork from source with Metal support."""
     source = _find_turboquant_source()

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/formatting.py

@@ -56,15 +56,6 @@ def truncate_with_tail(
     return "\n".join(head_block + [ellipsis] + tail_block)
 
 
-def truncate_one_line(text: str, max_chars: int = 80) -> str:
-    """Truncate to a single line for inline display (e.g. tool-call
-    args summary). Replaces newlines with spaces, hard-caps length."""
-    s = (text or "").replace("\n", " ").strip()
-    if len(s) <= max_chars:
-        return s
-    return s[: max_chars - 1] + "…"
-
-
 def split_stdout_stderr(text: str) -> tuple[str, str]:
     """Heuristic split of bash output into (stdout, stderr).
 

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/model_families.py

@@ -31,7 +31,7 @@ Scope we DO NOT cover here:
     override yet).
   - Prompt rendering differences — the SYSTEM_PROMPT content itself
     is still one string, and word-level per-family variants would live
-    in eval/prompt_variants.py, not here.
+    in dev/eval/prompt_variants.py, not here.
 """
 from __future__ import annotations
 

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/models_catalog.py

@@ -319,25 +319,57 @@ def _system_ram_gb() -> int:
     return 16
 
 
+# Architectures we never AUTO-recommend (users can still pick them explicitly in
+# the model picker): diffusion needs a separate one-shot runner — not the
+# standard llama-server — and cohere2_moe is unvalidated on this stack.
+_NO_AUTO_RECOMMEND_ARCHS = {"diffusion_gemma", "cohere2_moe"}
+
+# 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
+# (even at low bit) and must outrank the 12B dense, and a higher-bit quant of a
+# family wins. A 12B-BF16 must never be recommended over a 26B/35B MoE just
+# because its file happens to be larger.
+_RECOMMEND_ORDER = [
+    "qwen-q8",         # 35B-A3B MoE, near-lossless Q8
+    "gemma-q8",        # 26B-A4B MoE, near-lossless Q8
+    "qwen",            # 35B-A3B MoE Q2 — 94.7% HumanEval
+    "gemma",           # 26B-A4B MoE Q3 — 95.1% HumanEval
+    "gemma-12b-bf16",  # 12B dense, full precision
+    "gemma-12b",       # 12B dense Q4
+]
+
+
+def _capability_rank(choice) -> int:
+    """Lower = more capable. Curated models first; any other production model
+    sorts after them — so a newly-added quant is still recommendable, but never
+    outranks a curated MoE."""
+    order = {k: i for i, k in enumerate(_RECOMMEND_ORDER)}
+    return order.get(choice.key, len(_RECOMMEND_ORDER) + 1)
+
+
 def recommend(ram_gb: int | None = None) -> ModelChoice:
-    """Pick the catalog entry best suited to the host's RAM.
+    """Pick the best model for THIS machine's RAM — capability-ranked, never an
+    experimental architecture, and never a hardcoded default.
 
-    Rule of thumb: model weights + KV + OS headroom should fit in ~70% of
-    unified memory. We bias conservatively — being slightly under-quantized
-    is worse than crashing on load. Falls back to the first entry if nothing
-    matches (shouldn't happen with the current catalog).
+    Weights must fit in ~55% of unified memory (leaves room for KV cache,
+    activations, OS). Among the production-ready models that fit, return the
+    most capable (see ``_RECOMMEND_ORDER``) so the recommendation scales with
+    the user's hardware instead of defaulting to any one model.
     """
     if ram_gb is None:
         ram_gb = _system_ram_gb()
-    # Sort candidates by size_gb; pick the biggest whose weights fit
-    # in ~55% of RAM (leaves room for KV cache, activations, OS).
     budget = ram_gb * 0.55
-    fits = [c for c in CHOICES if c.size_gb <= budget]
-    if fits:
-        return max(fits, key=lambda c: c.size_gb)
-    # Nothing fits the budget — return the smallest entry so at least
-    # something is suggested rather than a recommendation the user can't run.
-    return min(CHOICES, key=lambda c: c.size_gb)
+    candidates = [
+        c for c in CHOICES
+        if c.architecture not in _NO_AUTO_RECOMMEND_ARCHS 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]
+    return min(prod or CHOICES, key=lambda c: c.size_gb)
 
 
 def by_key(key: str) -> ModelChoice | None:
@@ -404,39 +436,6 @@ def current(config) -> ModelChoice | None:
     return by_filename(name)
 
 
-def format_choice_long(c: ModelChoice, *, downloaded: bool, current_marker: bool = False) -> str:
-    """Multi-line formatted description for the picker UI."""
-    marker = " (current)" if current_marker else ""
-    status = "downloaded" if downloaded else f"will download {c.size_gb:.1f} GB"
-    if c.humaneval_pass_at_1 is not None:
-        bench = f"{c.humaneval_pass_at_1*100:.1f}% HumanEval pass@1 (measured on this stack)"
-    else:
-        bench = "no benchmark (untested on this stack)"
-    # Warn if this model is too big for the current Mac's RAM budget.
-    # Short inline warning; the picker screen also dims models with a
-    # bad fit so users can see at a glance.
-    try:
-        from .health import estimate_fit
-        fits, reason = estimate_fit(c.size_gb)
-    except Exception:
-        fits, reason = True, ""
-    fit_line = "" if fits else f"  ⚠ fit:      {reason}\n"
-    return (
-        f"[{c.key}] {c.name}{marker}\n"
-        f"  source:    {c.hf_repo}\n"
-        f"  url:       {c.hf_url}\n"
-        f"  filename:  {c.filename}\n"
-        f"  size:      {c.size_gb:.1f} GB ({status})\n"
-        f"{fit_line}"
-        f"  active:    {c.active_params}\n"
-        f"  arch:      {c.architecture}\n"
-        f"  license:   {c.license}\n"
-        f"  benchmark: {bench}\n"
-        f"  download path: {c.local_path}\n"
-        f"  note:      {c.notes}"
-    )
-
-
 # ---------------------------------------------------------------------------
 # Curated MODEL-GROUP layer (ADDITIVE — built on top of the catalog above).
 #

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/patching.py

@@ -1,9 +1,6 @@
 from __future__ import annotations
 
 import re
-import subprocess
-import tempfile
-from pathlib import Path
 from dataclasses import dataclass
 
 
@@ -24,13 +21,6 @@ class DiffFile:
     hunks: list[DiffHunk]
 
 
-def extract_last_diff_block(text: str) -> str | None:
-    matches = DIFF_BLOCK_RE.findall(text)
-    if not matches:
-        return None
-    return matches[-1].strip()
-
-
 def parse_diff(diff_text: str) -> list[DiffFile]:
     files: list[DiffFile] = []
     current: DiffFile | None = None
@@ -73,18 +63,3 @@ def build_diff(selected_files: list[DiffFile]) -> str:
             blocks.append(hunk.header)
             blocks.extend(hunk.lines)
     return "\n".join(blocks) + ("\n" if blocks else "")
-
-
-def apply_diff(repo_root: Path, diff_text: str) -> tuple[bool, str]:
-    with tempfile.NamedTemporaryFile("w", suffix=".diff", delete=False) as handle:
-        handle.write(diff_text)
-        patch_path = Path(handle.name)
-    result = subprocess.run(
-        ["git", "apply", "--reject", "--whitespace=nowarn", str(patch_path)],
-        cwd=repo_root,
-        capture_output=True,
-        text=True,
-        check=False,
-    )
-    output = (result.stdout + "\n" + result.stderr).strip()
-    return result.returncode == 0, output or "Patch applied."

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/plans.py

@@ -104,14 +104,6 @@ def read_plan(slug: str) -> str | None:
         return None
 
 
-def write_plan(slug: str, content: str) -> Path:
-    """Persist the plan. Returns the path on disk so callers can surface it."""
-    ensure_plans_dir()
-    p = plan_path(slug)
-    p.write_text(content, encoding="utf-8")
-    return p
-
-
 def new_slug() -> str:
     """Generate a fresh slug, retrying up to 10 times if a file already
     exists (very unlikely but keeps us honest)."""

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/process_registry.py

@@ -151,26 +151,6 @@ def process_summary(repo_root: Path | str) -> str:
     return "\n".join(lines)
 
 
-def stop_all_records(repo_root: Path | str) -> str:
-    records = load_records(repo_root)
-    if not records:
-        return "No LocalCode-managed processes recorded."
-    stopped = 0
-    skipped = 0
-    for record in list(records):
-        if record.stopped_at:
-            skipped += 1
-            continue
-        if record.pid <= 0 or not _pid_alive(record.pid):
-            skipped += 1
-            continue
-        if stop_record(repo_root, record):
-            stopped += 1
-        else:
-            skipped += 1
-    return f"Stopped {stopped} LocalCode-managed process group(s). Skipped {skipped}."
-
-
 def _pid_alive(pid: int) -> bool:
     try:
         os.kill(pid, 0)

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/recommendations.py

@@ -1,6 +1,6 @@
 """Per-machine model launch recommendations — the output of the model-opt loop.
 
-The offline optimizer (eval/model_opt.py) sweeps quants/params, scores each
+The offline optimizer (dev/eval/model_opt.py) sweeps quants/params, scores each
 combo on the eval suite, and writes the winner here. At launch,
 ``runtime.llama_server_command`` consults ``load_overrides`` so the tuned
 params are applied automatically.

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/runtime.py

@@ -217,7 +217,7 @@ class StreamEvent(dict):
 def apply_param_overrides(cmd: list[str], env: dict | None = None) -> list[str]:
     """Rewrite llama-server flags from ``LOCALCODE_OVERRIDE_*`` env vars.
 
-    Lets the offline model-optimizer (eval/model_opt.py) sweep launch
+    Lets the offline model-optimizer (dev/eval/model_opt.py) sweep launch
     parameters — GPU layers, context size, threads, batch — WITHOUT editing
     config or the catalog. Only flags already present in ``cmd`` are
     rewritten; an unset env var leaves the command untouched, so the default

{localcode-0.2.15 → localcode-0.2.16}/src/localcode/snapshots.py

@@ -193,21 +193,3 @@ def create_snapshot(session_id: str, repo_root: str, messages: list[dict],
     store = SnapshotStore()
     store.save(snapshot)
     return snapshot
-
-
-def restore_snapshot(snapshot: GhostSnapshot) -> dict:
-    """Restore file system state from a snapshot.
-
-    Returns dict of files restored.
-    """
-    restored = {}
-    root = Path(snapshot.repo_root)
-    for rel_path, content in snapshot.file_snapshots.items():
-        full = root / rel_path
-        try:
-            full.parent.mkdir(parents=True, exist_ok=True)
-            full.write_text(content)
-            restored[rel_path] = "restored"
-        except Exception as e:
-            restored[rel_path] = f"error: {e}"
-    return restored