simplicio-cli 0.5.13__tar.gz → 0.5.14__tar.gz

{simplicio_cli-0.5.13/simplicio_cli.egg-info → simplicio_cli-0.5.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: simplicio-cli
-Version: 0.5.13
+Version: 0.5.14
 Summary: Portable task-to-code pipeline that works with any LLM. Turn a one-line task into a verified code change — diff + test + verify loop. +55 pts on a 156-check benchmark, 21% faster, ~same tokens.
 Author-email: Wesley Simplicio <wesleybob4@gmail.com>
 License: MIT
@@ -30,8 +30,8 @@ Requires-Dist: sentence-transformers>=5.1.2
 Requires-Dist: numpy>=2.0.2
 Requires-Dist: anthropic>=0.105.2
 Requires-Dist: openai>=2.38.0
-Requires-Dist: simplicio-mapper>=0.6.10
-Requires-Dist: simplicio-prompt>=1.12.3
+Requires-Dist: simplicio-mapper>=0.7.1
+Requires-Dist: simplicio-prompt>=1.13.1
 Requires-Dist: httpx>=0.28.1
 Requires-Dist: orjson>=3.11.5
 Requires-Dist: diskcache>=5.6.3
@@ -561,7 +561,7 @@ user prompt. UserPromptSubmit is the right pre-hook for routing decisions.
 | GLM (z.ai) | `glm-4.6` | `https://api.z.ai/api/paas/v4` |
 | DeepSeek | `deepseek-chat` | `https://api.deepseek.com` |
 | OpenAI | `gpt-4.1` | `https://api.openai.com/v1` |
-| Local (Ollama) | `llama3` | `http://localhost:11434/v1` |
+| Local (Ollama) | `openbmb/minicpm5:latest` | `http://localhost:11434/v1` |
 | Local (in-process) | `local-llama/default` | *(leave unset)* |
 | Anthropic native | `claude-opus-4-7` | *(leave unset)* |
 
@@ -573,32 +573,31 @@ your `base_url` — so **any** OpenAI-like provider works without code changes.
 simplicio smoke      # prints provider config + one test call
 ```
 
-### Path 4 — offline-first local model (zero key, zero HTTP)
+### Path 4 — local Ollama primary with GGUF fallback
 
-simplicio ships an **in-process** backend powered by
-[`llama-cpp-python`](https://github.com/abetlen/llama-cpp-python). When **no
-provider is configured** (`SIMPLICIO_MODEL` *and* `SIMPLICIO_BASE_URL` both
-unset), it runs **Qwen3.5-2B Q6_K GGUF** directly. If that GGUF is not
-available yet, it falls back to the previous **Qwen2.5-Coder 1.5B Q8_0** and
-**Q6_K_L** executor files — small, code-specialized, fast on CPU, no API key,
-no Ollama, no HTTP overhead. The 6-layer contract is what keeps tiny local
-models usable for execution work.
+When **no provider is configured** (`SIMPLICIO_MODEL` and
+`SIMPLICIO_BASE_URL` both unset), simplicio uses local Ollama with
+`openbmb/minicpm5:latest`. If that call fails, it falls back to the in-process
+[`llama-cpp-python`](https://github.com/abetlen/llama-cpp-python) backend with
+`Qwen_Qwen3.5-2B-Q6_K.gguf`.
 
 ```bash
 pip install 'simplicio-cli[local]'          # pulls llama-cpp-python + huggingface-hub
 
 simplicio task "add input validation to createUser" \
-  --target src/users.ts --local              # forces the local model
+  --target src/users.ts --local              # forces local Ollama primary
 
-# the GGUF is fetched once from the Hugging Face Hub, then cached + reused
+# the fallback GGUF is fetched once from the Hugging Face Hub, then reused
 ```
 
 Explicit routes (override the default model/weights):
 
 ```bash
-SIMPLICIO_MODEL=local-llama/default                                  # Qwen3.5-2B Q6_K primary, Qwen2.5 fallbacks
-SIMPLICIO_MODEL=local-llama/bartowski/Qwen_Qwen3.5-2B-GGUF::Qwen3.5-2B-Q6_K.gguf
-SIMPLICIO_MODEL=local-llama/bartowski/Qwen2.5-Coder-7B-Instruct-GGUF::Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf
+SIMPLICIO_MODEL=openbmb/minicpm5:latest
+SIMPLICIO_BASE_URL=http://localhost:11434/v1
+SIMPLICIO_API_KEY=ollama
+SIMPLICIO_MODEL=local-llama/default                                  # Qwen_Qwen3.5-2B-Q6_K.gguf fallback
+SIMPLICIO_MODEL=local-llama/bartowski/Qwen_Qwen3.5-2B-GGUF::Qwen_Qwen3.5-2B-Q6_K.gguf
 SIMPLICIO_MODEL=local-llama//models/my-model.gguf                    # direct local path
 SIMPLICIO_LOCAL_MODEL_PATH=/models/my-model.gguf                     # always wins
 ```

{simplicio_cli-0.5.13 → simplicio_cli-0.5.14}/README.md

@@ -516,7 +516,7 @@ user prompt. UserPromptSubmit is the right pre-hook for routing decisions.
 | GLM (z.ai) | `glm-4.6` | `https://api.z.ai/api/paas/v4` |
 | DeepSeek | `deepseek-chat` | `https://api.deepseek.com` |
 | OpenAI | `gpt-4.1` | `https://api.openai.com/v1` |
-| Local (Ollama) | `llama3` | `http://localhost:11434/v1` |
+| Local (Ollama) | `openbmb/minicpm5:latest` | `http://localhost:11434/v1` |
 | Local (in-process) | `local-llama/default` | *(leave unset)* |
 | Anthropic native | `claude-opus-4-7` | *(leave unset)* |
 
@@ -528,32 +528,31 @@ your `base_url` — so **any** OpenAI-like provider works without code changes.
 simplicio smoke      # prints provider config + one test call
 ```
 
-### Path 4 — offline-first local model (zero key, zero HTTP)
+### Path 4 — local Ollama primary with GGUF fallback
 
-simplicio ships an **in-process** backend powered by
-[`llama-cpp-python`](https://github.com/abetlen/llama-cpp-python). When **no
-provider is configured** (`SIMPLICIO_MODEL` *and* `SIMPLICIO_BASE_URL` both
-unset), it runs **Qwen3.5-2B Q6_K GGUF** directly. If that GGUF is not
-available yet, it falls back to the previous **Qwen2.5-Coder 1.5B Q8_0** and
-**Q6_K_L** executor files — small, code-specialized, fast on CPU, no API key,
-no Ollama, no HTTP overhead. The 6-layer contract is what keeps tiny local
-models usable for execution work.
+When **no provider is configured** (`SIMPLICIO_MODEL` and
+`SIMPLICIO_BASE_URL` both unset), simplicio uses local Ollama with
+`openbmb/minicpm5:latest`. If that call fails, it falls back to the in-process
+[`llama-cpp-python`](https://github.com/abetlen/llama-cpp-python) backend with
+`Qwen_Qwen3.5-2B-Q6_K.gguf`.
 
 ```bash
 pip install 'simplicio-cli[local]'          # pulls llama-cpp-python + huggingface-hub
 
 simplicio task "add input validation to createUser" \
-  --target src/users.ts --local              # forces the local model
+  --target src/users.ts --local              # forces local Ollama primary
 
-# the GGUF is fetched once from the Hugging Face Hub, then cached + reused
+# the fallback GGUF is fetched once from the Hugging Face Hub, then reused
 ```
 
 Explicit routes (override the default model/weights):
 
 ```bash
-SIMPLICIO_MODEL=local-llama/default                                  # Qwen3.5-2B Q6_K primary, Qwen2.5 fallbacks
-SIMPLICIO_MODEL=local-llama/bartowski/Qwen_Qwen3.5-2B-GGUF::Qwen3.5-2B-Q6_K.gguf
-SIMPLICIO_MODEL=local-llama/bartowski/Qwen2.5-Coder-7B-Instruct-GGUF::Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf
+SIMPLICIO_MODEL=openbmb/minicpm5:latest
+SIMPLICIO_BASE_URL=http://localhost:11434/v1
+SIMPLICIO_API_KEY=ollama
+SIMPLICIO_MODEL=local-llama/default                                  # Qwen_Qwen3.5-2B-Q6_K.gguf fallback
+SIMPLICIO_MODEL=local-llama/bartowski/Qwen_Qwen3.5-2B-GGUF::Qwen_Qwen3.5-2B-Q6_K.gguf
 SIMPLICIO_MODEL=local-llama//models/my-model.gguf                    # direct local path
 SIMPLICIO_LOCAL_MODEL_PATH=/models/my-model.gguf                     # always wins
 ```

{simplicio_cli-0.5.13 → simplicio_cli-0.5.14}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "simplicio-cli"
-version = "0.5.13"
+version = "0.5.14"
 description = "Portable task-to-code pipeline that works with any LLM. Turn a one-line task into a verified code change — diff + test + verify loop. +55 pts on a 156-check benchmark, 21% faster, ~same tokens."
 readme = "README.md"
 license = { text = "MIT" }
@@ -44,8 +44,8 @@ dependencies = [
   "numpy>=2.0.2",
   "anthropic>=0.105.2",
   "openai>=2.38.0",
-  "simplicio-mapper>=0.6.10",
-  "simplicio-prompt>=1.12.3",
+  "simplicio-mapper>=0.7.1",
+  "simplicio-prompt>=1.13.1",
   "httpx>=0.28.1",
   "orjson>=3.11.5",
   "diskcache>=5.6.3",
@@ -54,8 +54,8 @@ dependencies = [
 
 [project.optional-dependencies]
 bench = ["fpdf2>=2.8.4"]
-# Offline-first in-process inference (Path 4). Pulls the llama.cpp Python
-# bindings plus huggingface-hub to fetch the default Qwen3.5-2B Q6_K GGUF.
+# Offline fallback in-process inference. Pulls the llama.cpp Python bindings
+# plus huggingface-hub to fetch Qwen_Qwen3.5-2B-Q6_K.gguf.
 local = ["llama-cpp-python>=0.3.23", "huggingface-hub>=1.8.0"]
 
 [project.urls]

simplicio_cli-0.5.14/simplicio/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.5.14"

{simplicio_cli-0.5.13 → simplicio_cli-0.5.14}/simplicio/cli.py

@@ -91,8 +91,8 @@ def _add_task_args(p: argparse.ArgumentParser, *, target_required: bool) -> None
     p.add_argument(
         "--local",
         action="store_true",
-        help="force the in-process local model (Qwen3.5-2B Q6_K GGUF, "
-        "no API key); overrides SIMPLICIO_MODEL/SIMPLICIO_BASE_URL",
+        help="force local Ollama (openbmb/minicpm5:latest, no API key) "
+        "with Qwen3.5 GGUF fallback; overrides SIMPLICIO_MODEL/SIMPLICIO_BASE_URL",
     )
 
 
@@ -121,10 +121,11 @@ def _add_run_args(p: argparse.ArgumentParser) -> None:
 
 def _force_local_if_requested(a: argparse.Namespace) -> None:
     if getattr(a, "local", False):
-        # Force Path 4: pin the local model and drop any HTTP endpoint so the
-        # in-process llama backend wins regardless of the ambient config.
-        os.environ["SIMPLICIO_MODEL"] = "local-llama/default"
-        os.environ.pop("SIMPLICIO_BASE_URL", None)
+        # Force Path 4: local Ollama primary. The provider layer falls back to
+        # the local Qwen GGUF if the Ollama call fails.
+        os.environ["SIMPLICIO_MODEL"] = "openbmb/minicpm5:latest"
+        os.environ["SIMPLICIO_BASE_URL"] = "http://localhost:11434/v1"
+        os.environ.setdefault("SIMPLICIO_API_KEY", "ollama")
 
 
 def _run_task_command(a: argparse.Namespace) -> int:

{simplicio_cli-0.5.13 → simplicio_cli-0.5.14}/simplicio/local_models.py

@@ -1,12 +1,11 @@
 """local_models.py — hardware-tier → local model recommendation + Ollama plumbing.
 
-Encodes the hardware map documented in bench/SCRATCH_MODE_RFC.md:
+Encodes the local LLM standard:
 
-  cpu-tiny      → qwen2.5-coder:3b           (~2 GB Q4)
-  cpu-small     → qwen2.5-coder:7b           (~5 GB Q4)
-  gpu-mid       → qwen2.5-coder:14b          (~9 GB Q4)
-  gpu-large     → unsloth Qwen3-Coder-30B-A3B-Instruct Q4_K_M (~17 GB)
-  gpu-xlarge    → unsloth Qwen3-Coder-Next Q4 (~26 GB)
+  all tiers      → openbmb/minicpm5:latest via local Ollama (~0.7 GB)
+
+The GGUF fallback is handled by simplicio.providers as
+Qwen_Qwen3.5-2B-Q6_K.gguf when Ollama is unavailable.
 
 Hard rule (issue #32 follow-up):
 - NEVER auto-pull a model that does not fit the detected tier.
@@ -35,29 +34,25 @@ class ModelSpec:
     notes: str = ""
 
 
+DEFAULT_LOCAL_OLLAMA_ID = "openbmb/minicpm5:latest"
+DEFAULT_LOCAL_OLLAMA_LABEL = "MiniCPM5 local (Ollama)"
+DEFAULT_LOCAL_OLLAMA_SIZE_GB = 0.7
+DEFAULT_LOCAL_OLLAMA_NOTES = (
+    "canonical local primary; falls back to Qwen_Qwen3.5-2B-Q6_K.gguf in providers"
+)
+
+
 # Order matters: list is consulted in the rare case the user asks for the
 # next-step-up model. The default lookup is by exact tier.
 RECOMMENDATIONS: dict[str, ModelSpec] = {
-    "cpu-tiny":   ModelSpec("cpu-tiny",   "qwen2.5-coder:3b",   2.0,
-                            "Qwen2.5-Coder 3B (Q4)",
-                            "minimal but usable; 12-18 tok/s on CPU"),
-    "cpu-small":  ModelSpec("cpu-small",  "qwen2.5-coder:7b",   5.0,
-                            "Qwen2.5-Coder 7B (Q4)",
-                            "good balance; ~25 tok/s on M1/M2"),
-    "gpu-mid":    ModelSpec("gpu-mid",    "qwen2.5-coder:14b",  9.0,
-                            "Qwen2.5-Coder 14B (Q4)",
-                            "real-work proxy; 30-50 tok/s on 16 GB VRAM"),
-    "gpu-large":  ModelSpec("gpu-large",
-                            "hf.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF:Q4_K_M",
-                            17.5, "Qwen3-Coder 30B A3B (Q4_K_M, unsloth)",
-                            "MoE 3B active params; matches our bench cloud results"),
-    "gpu-xlarge": ModelSpec("gpu-xlarge",
-                            "hf.co/unsloth/Qwen3-Coder-Next-GGUF:Q4_K_M",
-                            26.0, "Qwen3-Coder Next (Q4)",
-                            "best local coder 2026; needs 32+ GB unified or VRAM"),
-    "unknown":    ModelSpec("unknown",    "qwen2.5-coder:7b",   5.0,
-                            "Qwen2.5-Coder 7B (Q4) [fallback]",
-                            "safe default when detection inconclusive"),
+    tier: ModelSpec(
+        tier,
+        DEFAULT_LOCAL_OLLAMA_ID,
+        DEFAULT_LOCAL_OLLAMA_SIZE_GB,
+        DEFAULT_LOCAL_OLLAMA_LABEL,
+        DEFAULT_LOCAL_OLLAMA_NOTES,
+    )
+    for tier in ("cpu-tiny", "cpu-small", "gpu-mid", "gpu-large", "gpu-xlarge", "unknown")
 }
 
 
@@ -71,7 +66,7 @@ def ollama_present() -> bool:
 def ollama_list_installed() -> list[str]:
     """Return the list of Ollama models currently installed.
 
-    Each entry is the full tag the user can target (e.g. "qwen2.5-coder:7b").
+    Each entry is the full tag the user can target (e.g. "openbmb/minicpm5:latest").
     Empty list on failure — never raises.
     """
     if not ollama_present():
@@ -97,7 +92,7 @@ def ollama_list_installed() -> list[str]:
 
 
 def is_installed(ollama_id: str) -> bool:
-    """Looser match: handle `hf.co/...:Q4_K_M` vs `qwen2.5-coder:7b` styles."""
+    """Looser match hook kept for future aliases; currently exact tag match."""
     installed = ollama_list_installed()
     return ollama_id in installed
 

{simplicio_cli-0.5.13 → simplicio_cli-0.5.14}/simplicio/providers.py

@@ -21,13 +21,18 @@ Four modes, picked by SIMPLICIO_MODEL prefix (or by absence of config):
      given SIMPLICIO_HOOK_GUARD=1 so the inner CLI does not re-trigger the
      simplicio UserPromptSubmit hook (recursion guard).
 
-4. In-process local inference via llama-cpp-python (offline-first, zero key)
+4. Local Ollama default (offline-first, zero key)
+     SIMPLICIO_MODEL=(unset)
+     SIMPLICIO_BASE_URL=(unset)
+     -> http://localhost:11434/v1 with openbmb/minicpm5:latest
+     If Ollama is unavailable or the default call fails, simplicio falls back
+     to Path 5.
+
+5. In-process local inference via llama-cpp-python (GGUF fallback, zero key)
      SIMPLICIO_MODEL=local-llama/<repo>::<file.gguf>   -> explicit HF GGUF
-     SIMPLICIO_MODEL=local-llama/default               -> bundled default
+     SIMPLICIO_MODEL=local-llama/default               -> fallback Qwen GGUF
      SIMPLICIO_MODEL=local-llama//abs/path/model.gguf  -> direct local path
-     This is also the DEFAULT when neither SIMPLICIO_MODEL nor
-     SIMPLICIO_BASE_URL is set: simplicio runs Qwen3.5-2B Q6_K GGUF on CPU with
-     no HTTP overhead. Legacy Qwen2.5-Coder GGUFs remain fallback executors. The
+     The
      GGUF is reused from ~/.simplicio/models/executor when present, otherwise
      fetched once from the Hugging Face Hub. Requires the `local` extra:
      pip install 'simplicio-cli[local]'.
@@ -69,21 +74,17 @@ def _inline_feedback(prompt, feedback):
 
 
 # --------------------------------------------------------------------------- #
-# Path 4: in-process local inference (llama-cpp-python). Offline-first default.
+# Path 4: local Ollama primary + Path 5: in-process GGUF fallback.
 # --------------------------------------------------------------------------- #
 
-# Qwen3.5-2B Q6_K is the preferred local executor. The previous Qwen2.5-Coder
-# files stay in the automatic fallback chain so existing offline setups keep
-# working if the new GGUF is not present yet.
+DEFAULT_OLLAMA_MODEL = "openbmb/minicpm5:latest"
+DEFAULT_OLLAMA_BASE_URL = "http://localhost:11434/v1"
+DEFAULT_OLLAMA_API_KEY = "ollama"
+
+# Qwen3.5-2B Q6_K is the local GGUF fallback for machines where Ollama is not
+# reachable or the MiniCPM5 call fails.
 LOCAL_DEFAULT_REPO = "bartowski/Qwen_Qwen3.5-2B-GGUF"
-LOCAL_DEFAULT_FILE = "Qwen3.5-2B-Q6_K.gguf"
-LOCAL_FALLBACK_REPO = "bartowski/Qwen2.5-Coder-1.5B-Instruct-GGUF"
-LOCAL_FALLBACK_FILE = "Qwen2.5-Coder-1.5B-Instruct-Q8_0.gguf"
-LOCAL_SECONDARY_FALLBACK_FILE = "Qwen2.5-Coder-1.5B-Instruct-Q6_K_L.gguf"
-LOCAL_FALLBACK_SPECS = (
-    (LOCAL_FALLBACK_REPO, LOCAL_FALLBACK_FILE),
-    (LOCAL_FALLBACK_REPO, LOCAL_SECONDARY_FALLBACK_FILE),
-)
+LOCAL_DEFAULT_FILE = "Qwen_Qwen3.5-2B-Q6_K.gguf"
 LOCAL_EXECUTOR_DIR = "~/.simplicio/models/executor"
 LOCAL_MODEL_PREFIX = "local-llama/"
 
@@ -95,19 +96,29 @@ _LOCAL_LLAMA_CACHE = {}
 def _is_local(model, base):
     """True when generate() should route to the in-process llama backend.
 
-    Either an explicit `local-llama/` model, or the offline-first default:
-    nothing configured at all (no model, no OpenAI-compatible base_url).
+    Only explicit `local-llama/` models use the in-process backend. The empty
+    config default is the local Ollama primary and is handled separately.
     """
     if model and model.startswith(LOCAL_MODEL_PREFIX):
         return True
-    return not model and not base
+    return False
+
+
+def _is_default_ollama(model, base):
+    """True when the local default should use Ollama first."""
+    if not model and not base:
+        return True
+    return (
+        model == DEFAULT_OLLAMA_MODEL
+        and (not base or base.rstrip("/") == DEFAULT_OLLAMA_BASE_URL.rstrip("/"))
+    )
 
 
 def _local_spec(model):
     """Resolve (repo, file, path) for a local-llama model id.
 
     Forms after the `local-llama/` prefix:
-      "" / "default" / "auto"   -> bundled Qwen3.5-2B Q6_K default
+      "" / "default" / "auto"   -> Qwen3.5-2B Q6_K GGUF fallback
       "<repo>::<file.gguf>"     -> explicit HF repo + filename
       "/abs/path/model.gguf"    -> direct local path (no download)
       "<repo>"                  -> HF repo + default/SIMPLICIO_LOCAL_MODEL_FILE
@@ -137,10 +148,6 @@ def _local_executor_dir() -> Path:
 
 def _local_candidates(repo, fname):
     """Return candidate (repo, GGUF filename) pairs in preference order."""
-    if os.environ.get("SIMPLICIO_LOCAL_MODEL_FILE"):
-        return [(repo, fname)]
-    if repo == LOCAL_DEFAULT_REPO and fname == LOCAL_DEFAULT_FILE:
-        return [(repo, fname), *LOCAL_FALLBACK_SPECS]
     return [(repo, fname)]
 
 
@@ -296,11 +303,93 @@ def _charge_if_budgeted(model, prompt, out):
     charge_provider_call(model, prompt, out or "")
 
 
+def _generate_local_cached(prompt, feedback, model, max_tokens, cache_full_prompt):
+    from ._cache import CacheEntry, cache, make_key
+
+    eff_model = model or (LOCAL_MODEL_PREFIX + "default")
+    # Fold the resolved weights into the cache key: two different GGUFs can
+    # both route as `local-llama/default` (via SIMPLICIO_LOCAL_MODEL_PATH /
+    # _REPO / _FILE), and must NOT share cached completions.
+    repo, fname, path = _local_spec(eff_model)
+    weights = path or f"{repo}/{fname}"
+    key = make_key(
+        "local-llama",
+        eff_model,
+        prompt,
+        feedback=feedback,
+        max_tokens=max_tokens,
+        weights=weights,
+    )
+    cached = cache().get(key)
+    if cached is not None:
+        return cached.completion
+    out = _local_generate(prompt, feedback, eff_model, max_tokens)
+    _charge_if_budgeted(eff_model, cache_full_prompt, out)
+    cache().put(key, CacheEntry(out, provider_id="local-llama", model=eff_model))
+    return out
+
+
+def _openai_compatible_generate(model, base, key, prompt, feedback, max_tokens):
+    from openai import OpenAI
+
+    cli = OpenAI(base_url=base, api_key=key)
+    r = cli.chat.completions.create(
+        model=model, max_tokens=max_tokens, messages=_msgs(prompt, feedback)
+    )
+    return r.choices[0].message.content
+
+
+def _generate_default_ollama_with_fallback(prompt, feedback, max_tokens, cache_full_prompt):
+    from ._cache import CacheEntry, cache, make_key
+
+    provider_id = f"openai-compatible:{DEFAULT_OLLAMA_BASE_URL}"
+    key = make_key(
+        provider_id,
+        DEFAULT_OLLAMA_MODEL,
+        prompt,
+        feedback=feedback,
+        max_tokens=max_tokens,
+    )
+    cached = cache().get(key)
+    if cached is not None:
+        return cached.completion
+    try:
+        out = _openai_compatible_generate(
+            DEFAULT_OLLAMA_MODEL,
+            DEFAULT_OLLAMA_BASE_URL,
+            DEFAULT_OLLAMA_API_KEY,
+            prompt,
+            feedback,
+            max_tokens,
+        )
+    except Exception as exc:  # noqa: BLE001 - default local path falls back to GGUF
+        try:
+            return _generate_local_cached(
+                prompt,
+                feedback,
+                LOCAL_MODEL_PREFIX + "default",
+                max_tokens,
+                cache_full_prompt,
+            )
+        except SystemExit as fallback_exc:
+            raise SystemExit(
+                "simplicio: default local Ollama model "
+                f"{DEFAULT_OLLAMA_MODEL} failed ({exc}); fallback GGUF "
+                f"{LOCAL_DEFAULT_FILE} also failed ({fallback_exc})"
+            )
+    _charge_if_budgeted(DEFAULT_OLLAMA_MODEL, cache_full_prompt, out)
+    cache().put(
+        key,
+        CacheEntry(out, provider_id=provider_id, model=DEFAULT_OLLAMA_MODEL),
+    )
+    return out
+
+
 def generate(prompt, feedback=None, max_tokens=4000, template_version=None):
     # Cache lookup BEFORE provider config. Key uses just SIMPLICIO_MODEL
     # (no credential check) so a hit returns without requiring an API key
     # to be set in the environment.
-    from ._cache import cache, make_key, CacheEntry
+    from ._cache import cache, make_key
     model_name = os.environ.get("SIMPLICIO_MODEL", "").strip()
     cache_full_prompt = _inline_feedback(prompt, feedback)
     cache_key = make_key(
@@ -315,35 +404,21 @@ def generate(prompt, feedback=None, max_tokens=4000, template_version=None):
     c = _cfg()
     model = c["model"]
 
-    # Path 4: in-process local inference. Explicit `local-llama/` model, or the
-    # offline-first default when nothing is configured. No API key, no HTTP.
-    if _is_local(model, c["base"]):
-        eff_model = model or (LOCAL_MODEL_PREFIX + "default")
-        # Fold the resolved weights into the cache key: two different GGUFs can
-        # both route as `local-llama/default` (via SIMPLICIO_LOCAL_MODEL_PATH /
-        # _REPO / _FILE), and must NOT share cached completions.
-        repo, fname, path = _local_spec(eff_model)
-        weights = path or f"{repo}/{fname}"
-        key = make_key(
-            "local-llama",
-            eff_model,
-            prompt,
-            feedback=feedback,
-            max_tokens=max_tokens,
-            weights=weights,
+    # Path 4: no config means local Ollama first, then GGUF fallback.
+    if _is_default_ollama(model, c["base"]):
+        return _generate_default_ollama_with_fallback(
+            prompt, feedback, max_tokens, cache_full_prompt
         )
-        cached = cache().get(key)
-        if cached is not None:
-            return cached.completion
-        out = _local_generate(prompt, feedback, eff_model, max_tokens)
-        _charge_if_budgeted(eff_model, cache_full_prompt, out)
-        cache().put(key, CacheEntry(out, provider_id="local-llama", model=eff_model))
-        return out
+
+    # Path 5: in-process local inference via explicit `local-llama/` model.
+    if _is_local(model, c["base"]):
+        return _generate_local_cached(prompt, feedback, model, max_tokens, cache_full_prompt)
 
     if not model:
         raise SystemExit(
             "set SIMPLICIO_MODEL (e.g. anthropic/claude-opus-4, claude-cli/sonnet, "
-            "codex-cli/gpt-5, local-llama/default, glm-4.6, llama3, claude-opus-4-7)"
+            "codex-cli/gpt-5, openbmb/minicpm5:latest, local-llama/default, "
+            "glm-4.6, llama3, claude-opus-4-7)"
         )
     provider_id = _provider_id(model, c["base"])
     key = make_key(
@@ -394,13 +469,9 @@ def generate(prompt, feedback=None, max_tokens=4000, template_version=None):
         return out
 
     # Any OpenAI-compatible endpoint (OpenRouter, GLM, DeepSeek, local...)
-    from openai import OpenAI
-
-    cli = OpenAI(base_url=c["base"], api_key=c["key"])
-    r = cli.chat.completions.create(
-        model=model, max_tokens=max_tokens, messages=_msgs(prompt, feedback)
+    out = _openai_compatible_generate(
+        model, c["base"], c["key"], prompt, feedback, max_tokens
     )
-    out = r.choices[0].message.content
     _charge_if_budgeted(model, cache_full_prompt, out)
     cache().put(key, CacheEntry(out, provider_id=provider_id, model=model))
     return out
@@ -408,6 +479,12 @@ def generate(prompt, feedback=None, max_tokens=4000, template_version=None):
 
 def info():
     c = _cfg()
+    if _is_default_ollama(c["model"], c["base"]):
+        return (
+            f"model={DEFAULT_OLLAMA_MODEL} provider=ollama "
+            f"base={DEFAULT_OLLAMA_BASE_URL} fallback={LOCAL_DEFAULT_FILE} "
+            "key=not-needed"
+        )
     if _is_local(c["model"], c["base"]):
         eff_model = c["model"] or (LOCAL_MODEL_PREFIX + "default (auto)")
         repo, fname, path = _local_spec(c["model"] or "")