[BUG] Gemini native provider: version-less aliases (gemini-flash-latest, gemini-flash-lite-latest, gemini-pro-latest) lose tool support — tools are silently dropped from requests

Description

The native Gemini provider derives supports_tools, is_gemini_2_0, and the automatic thinking_config from a version regex — re.search(r"gemini-(\d+(?:\.\d+)?)", model.lower()) in GeminiCompletion._normalize_gemini_fields (lib/crewai/src/crewai/llms/providers/gemini/completion.py:99). The version-less aliases that crewAI lists in its own GEMINI_MODELS table (gemini-flash-latest, gemini-flash-lite-latest, gemini-pro-latest; lib/crewai/src/crewai/llms/constants.py:310-312) contain no digit after gemini-, so the regex never matches and they are treated as pre-1.5 models.

Two user-visible consequences:

1. _prepare_generation_config only attaches tools if tools and self.supports_tools (line 490), so LLM(model="gemini/gemini-flash-latest").call(messages, tools=[...]) sends a request with no tools at all — the model can never call them.
2. supports_function_calling() returns False, so an Agent on one of these aliases silently downgrades from native function calling to the ReAct text protocol.

Structured output also takes the legacy response_schema path and thinking is not auto-enabled, unlike the numbered equivalents (e.g. gemini-2.5-flash).

Steps to Reproduce

1. pip install "crewai[google-genai]" (no API key needed).
2. Run:

from crewai.llms.providers.gemini.completion import GeminiCompletion
g = GeminiCompletion(model="gemini-flash-latest")
print(g.supports_tools, g.supports_function_calling())
cfg = g._prepare_generation_config(None, [{"type": "function", "function": {"name": "lookup", "description": "d", "parameters": {"type": "object", "properties": {}}}}], None)
print(getattr(cfg, "tools", None))

3. Compare with GeminiCompletion(model="gemini-2.5-flash").

Expected behavior

The -latest aliases resolve to current 2.5-generation models, so supports_tools / supports_function_calling() are True and the request config contains the tools, exactly as for gemini-2.5-flash.

Screenshots/Code snippets

model                                 supports_tools  supports_fc  tools_sent  thinking_auto
gemini-2.5-flash                      True            True         True        True
gemini-2.0-flash                      True            True         True        False
gemini-flash-latest                   False           False        False       False
gemini-flash-lite-latest              False           False        False       False
gemini-pro-latest                     False           False        False       False

Operating System

Windows 11 (reproduces anywhere; capability detection is offline)

Python Version

3.12 (verified on 3.13.12 as well)

crewAI Version

1.15.22 (also reproduced on main @ 3831e8b)

crewAI Tools Version

n/a (not required)

Virtual Environment

Venv (uv)

Evidence

Script output above: the three aliases report supports_tools=False and omit tools from the GenerateContentConfig, while gemini-2.5-flash includes them. llm.py:474 documents "latest" versions as intended-supported; constants.py:310-312 lists the aliases.

Possible Solution

When the version regex misses, consult an explicit alias map for the version-less aliases (e.g. {"gemini-flash-latest": 2.5, "gemini-flash-lite-latest": 2.5, "gemini-pro-latest": 2.5}) rather than loosening the regex — a "first number anywhere" regex would mis-detect gemini-gemma-2-27b-it (gemma has no function calling) as 2.0. Same class of fix as merged PR #4009 for Anthropic. Happy to open a PR with this plus regression tests.

Additional context

Also affects gemini-robotics-er-1.5-preview and the gemini-live-* ids in the table (out of scope for the minimal fix; can be handled separately). Verified no existing issue/PR covers this (searched supports_tools, alias names, thinking config, and all open PR file lists touching the provider on 2026-09-20). This issue was found and written with AI assistance; please apply the llm-generated label per the contribution policy.

Start in lib/crewai/src/crewai/llms/providers/gemini/completion.py at GeminiCompletion._normalize_gemini_fields and review the alias entries in lib/crewai/src/crewai/llms/constants.py. Reproduce the comparison with gemini-2.5-flash, then add regression coverage for the three -latest aliases. Done means their capability flags and automatic configuration match the expected current-generation behavior and tools remain in the generated config.