roadmodel 0.2.2__tar.gz → 0.2.3__tar.gz

{roadmodel-0.2.2 → roadmodel-0.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: roadmodel
-Version: 0.2.2
+Version: 0.2.3
 Summary: BYO-key CLI that recommends the right AI model, platform, and settings for a prompt.
 Project-URL: Homepage, https://roadmodel.ai
 Project-URL: Repository, https://github.com/nathanramoscfa/roadmodel

{roadmodel-0.2.2 → roadmodel-0.2.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "roadmodel"
-version = "0.2.2"
+version = "0.2.3"
 description = "BYO-key CLI that recommends the right AI model, platform, and settings for a prompt."
 readme = "README.md"
 requires-python = ">=3.11"

{roadmodel-0.2.2 → roadmodel-0.2.3}/src/roadmodel/__init__.py

@@ -1,2 +1,2 @@
 # src/roadmodel/__init__.py
-__version__ = "0.2.2"
+__version__ = "0.2.3"

{roadmodel-0.2.2 → roadmodel-0.2.3}/src/roadmodel/providers/__init__.py

@@ -13,4 +13,5 @@ class ProviderAdapter(Protocol):
         model: str | None = None,
         api_key: str,
         max_output_tokens: int | None = None,
+        thinking_budget: int | None = None,
     ) -> str: ...

{roadmodel-0.2.2 → roadmodel-0.2.3}/src/roadmodel/providers/anthropic.py

@@ -13,7 +13,16 @@ def recommend(
     model: str | None = None,
     api_key: str,
     max_output_tokens: int | None = None,
+    thinking_budget: int | None = None,
 ) -> str:
+    # thinking_budget is accepted for ProviderAdapter Protocol parity but
+    # intentionally NOT forwarded: it is a Gemini-specific knob for the
+    # recommender latency work (issue #132). Anthropic extended-thinking has
+    # different semantics (a `thinking` block with its own budget_tokens and
+    # minimums) and the recommender response shape does not tolerate small
+    # caps on Anthropic at all (PR #128). Anthropic reasoning control is
+    # Phase 5 paid-frontier scope.
+    _ = thinking_budget
     try:
         from anthropic import Anthropic, APIError
     except Exception as exc:  # pragma: no cover - dependency/runtime guard

{roadmodel-0.2.2 → roadmodel-0.2.3}/src/roadmodel/providers/google.py

@@ -15,6 +15,7 @@ def recommend(
     model: str | None = None,
     api_key: str,
     max_output_tokens: int | None = None,
+    thinking_budget: int | None = None,
 ) -> str:
     try:
         from google import genai
@@ -32,6 +33,13 @@ def recommend(
         config: Any = {"system_instruction": system}
         if max_output_tokens is not None:
             config["max_output_tokens"] = max_output_tokens
+        if thinking_budget is not None:
+            # Gemini 2.5+ Flash reasons by default, and that reasoning is
+            # decoded before the visible answer (and counts against
+            # max_output_tokens). thinking_budget caps it: 0 disables
+            # thinking entirely, a small value bounds it. `is not None` —
+            # not truthiness — because 0 is a meaningful value (thinking off).
+            config["thinking_config"] = {"thinking_budget": thinking_budget}
         response = client.models.generate_content(
             model=model or DEFAULT_MODEL,
             contents=prompt,

{roadmodel-0.2.2 → roadmodel-0.2.3}/src/roadmodel/providers/openai.py

@@ -31,7 +31,14 @@ def recommend(
     model: str | None = None,
     api_key: str,
     max_output_tokens: int | None = None,
+    thinking_budget: int | None = None,
 ) -> str:
+    # thinking_budget is accepted for ProviderAdapter Protocol parity but
+    # intentionally NOT forwarded: it is a Gemini-specific knob for the
+    # recommender latency work (issue #132). OpenAI reasoning control uses a
+    # different mechanism (`reasoning.effort` on reasoning models), out of
+    # scope for the free-tier recommender, which runs on Gemini Flash.
+    _ = thinking_budget
     try:
         from openai import APIError, OpenAI
     except Exception as exc:  # pragma: no cover - dependency/runtime guard

{roadmodel-0.2.2 → roadmodel-0.2.3}/src/roadmodel/recommend.py

@@ -124,6 +124,7 @@ def recommend(
     config: Config,
     *,
     max_output_tokens: int | None = None,
+    thinking_budget: int | None = None,
 ) -> dict[str, str]:
     user_context_text = user_context.read(config.user_context_path)
     system_prompt, user_prompt = build_prompt(prompt, user_context_text=user_context_text)
@@ -134,6 +135,7 @@ def recommend(
         model=config.model,
         api_key=config.api_key,
         max_output_tokens=max_output_tokens,
+        thinking_budget=thinking_budget,
     )
     return parse_response(raw_response)
 
@@ -169,9 +171,12 @@ def recommend_structured(
     output_tokens: int | None = None,
     max_mode: bool = False,
     max_output_tokens: int | None = None,
+    thinking_budget: int | None = None,
 ) -> dict[str, Any]:
     """Return roadmap-style structured output plus optional cost estimates."""
-    base = recommend(prompt, config, max_output_tokens=max_output_tokens)
+    base = recommend(
+        prompt, config, max_output_tokens=max_output_tokens, thinking_budget=thinking_budget
+    )
     payload: dict[str, Any] = {
         "model": base["model"],
         "platform": base["platform"],