claude-opus-4-7

[lchoquel]

Summary by cubic

Adds Gemini 3.1 models and “latest” aliases for Google, promotes claude-4.7-opus to default premium, and correctly maps Google’s MINIMAL reasoning level. Also simplifies ReactFlow asset caching and improves error messages when a local model deck is stale.

• New Features

  • Added gemini-3.1-flash-lite and gemini-3.1-pro to Google backend configs and test profiles.
  • Introduced gemini-pro-latest, gemini-flash-latest, and gemini-flash-lite-latest aliases; switched LLM deck defaults to use them (e.g., best-gemini = gemini-pro-latest, small vision/creative now gemini-flash-latest, cheap retrieval uses gemini-flash-lite-latest).
  • Promoted claude-4.7-opus as best-claude and as default for premium, premium-vision, and premium-structured.
  • Mapped ReasoningEffort.MINIMAL to Google’s ThinkingLevel.MINIMAL; enum and default map updated; tests and docs updated.

• Refactors

  • Simplified ReactFlow standalone asset caching with functools.cache and improved ModelNotFoundError guidance for stale local decks (pipelex init inference hint).
  • Renamed gemini-3.0-flash-preview to gemini-3.0-flash; regenerated gateway model docs.

^{Written for commit 0891ac9. Summary will update on new commits.}

[greptile-apps]

Greptile Summary

• Adds Gemini 3.1 series models (gemini-3.1-flash-lite, gemini-3.1-pro) and three floating "latest" aliases (gemini-pro-latest, gemini-flash-latest, gemini-flash-lite-latest) to the Google backend config, and maps ReasoningEffort.MINIMAL to Google's new ThinkingLevel.MINIMAL (previously collapsed to LOW).
• Simplifies standalone_assets.py caching from manual global/nullable-module-var pattern to @functools.cache, and updates the LLM deck to use claude-4.7-opus as the best-claude and default-premium aliases.

Confidence Score: 5/5

Safe to merge — no P0/P1 issues; all changes are additive config and a clean refactor.

The only finding is a P2 clarification question about whether the "latest" alias entries in google.toml are gateway-only or also intended for direct Google API use. No logic bugs, data integrity issues, or breaking changes were found.

pipelex/kit/configs/inference/backends/google.toml — the three "latest" alias entries lack explicit model_id and may not resolve correctly against the direct Google API.

Important Files Changed

Filename	Overview
pipelex/plugins/google/google_config.py	Adds MINIMAL to GoogleThinkingLevel enum and maps it to genai_types.ThinkingLevel.MINIMAL in get_reasoning_level; match/case stays exhaustive and correct.
pipelex/graph/reactflow/standalone_assets.py	Replaces manual global-variable caching with @functools.cache; functionally equivalent and eliminates the PLW0603 noqa suppressions.
pipelex/kit/configs/inference/backends/google.toml	Adds gemini-3.1-flash-lite, gemini-3.1-pro, and three "latest" alias entries; aliases omit model_id (factory falls back to section name), which works for the gateway but may not be valid for direct Google API calls.
pipelex/kit/configs/inference/deck/1_llm_deck.toml	Updates best-claude alias and default-premium* aliases to claude-4.7-opus, adds cheap presets and retrieval tiers; no issues.
tests/unit/pipelex/plugins/google/test_google_reasoning.py	Adds MINIMAL to ADAPTIVE mode parametrize and verifies ThinkingLevel.MINIMAL is returned; MANUAL + NONE test correctly omits get_reasoning_budget mock since the disabled-gate short-circuits before that call.
tests/unit/pipelex/cogt/llm/test_llm_config_reasoning.py	_DEFAULT_GOOGLE_LEVEL_MAP updated to map "minimal" → "minimal"; tests for level-value validation and all-disabled scenario are correct.
pipelex/pipelex.toml	google_config.effort_to_level_map.minimal changed from "low" to "minimal", consistent with the new GoogleThinkingLevel.MINIMAL enum value.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[ReasoningEffort] --> B{GoogleLLMWorker
_build_thinking_config}
    B --> C{reasoning_effort set?}
    C -- yes --> D{ThinkingMode?}
    C -- no --> E{reasoning_budget set?}
    E -- yes --> F[_build_thinking_config_for_budget
ThinkingConfig budget]
    E -- no --> G[return None]
    D -- MANUAL --> H[get_reasoning_level
effort → level string]
    H --> I{level is None
disabled?}
    I -- yes --> J[ThinkingConfig budget=0]
    I -- no --> K[get_reasoning_budget
prompting_target + effort]
    K --> L[ThinkingConfig thinking_budget]
    D -- ADAPTIVE --> M[get_reasoning_level
effort → ThinkingLevel]
    M --> N{level is None?}
    N -- yes --> O[ThinkingConfig budget=0]
    N -- no --> P[ThinkingConfig thinking_level
MINIMAL / LOW / MEDIUM / HIGH]
    D -- NONE --> Q[raise LLMCapabilityError]

Loading

Prompt To Fix All With AI

This is a comment left during a code review.
Path: pipelex/kit/configs/inference/backends/google.toml
Line: 91-111

Comment:
**Latest-alias entries rely on name-fallback for `model_id`**

`gemini-pro-latest`, `gemini-flash-latest`, and `gemini-flash-lite-latest` have no explicit `model_id`, so `InferenceModelSpecFactory` will fall back to the section name (line 99: `blueprint.model_id or name`). For Pipelex Gateway this is fine — the gateway resolves these aliases. For a direct Google Gemini API backend, though, the string `"gemini-pro-latest"` (without a version number) is sent as-is to Google's API; if Google doesn't recognise it as a valid model alias the call will fail at runtime.

Consider either adding an explicit `model_id` (e.g. pointing to the current latest concrete model ID) or a comment clarifying these entries are gateway-only.

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "Merge branch 'dev' into release/v0.24.0" | Re-trigger Greptile}

[cubic-dev-ai]

1 issue found across 15 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="pipelex/kit/configs/inference/backends/google.toml">

<violation number="1" location="pipelex/kit/configs/inference/backends/google.toml:65">
P2: Renaming this preset drops the existing `gemini-3.0-flash-preview` handle and breaks backward compatibility. Keep the old name as an alias or retain the original preset key.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}