fix(live): forward thinking config

[he-yufeng]

Summary

• forward GenerateContentConfig.thinking_config into LiveConnectConfig before opening Gemini live sessions
• keep existing live_connect_config.thinking_config untouched when the request config has no thinking config
• add a regression test for Gemini.connect()

Fixes #5805.

To verify

• $env:PYTHONPATH=(Resolve-Path src).Path; python -m pytest tests\unittests\models\test_google_llm.py -q
• $env:PYTHONPATH=(Resolve-Path src).Path; python -m py_compile src\google\adk\models\google_llm.py tests\unittests\models\test_google_llm.py
• python -m pyink --check src\google\adk\models\google_llm.py tests\unittests\models\test_google_llm.py
• git diff --check

Note: pylint on the full google_llm.py module still reports pre-existing warnings unrelated to this change, so I did not treat that as a blocker.

[adk-bot]

Response from ADK Triaging Agent

Hello @he-yufeng, thank you for creating this PR!

This PR looks great and has a clear explanation and associated issue. To help the reviewers merge your PR more efficiently, could you please provide the following information as per our contribution guidelines:

• Summary of passed pytest results: Please run the tests locally and paste a summary of the successful test runs (e.g., pytest tests/unittests/models/test_google_llm.py).
• Logs or screenshot demonstrating the fix: Since this is a bug fix for live session connection configs, providing logs showing the successful session establishment with the forwarded thinking_config would be very helpful.

Thank you for your contribution!

[he-yufeng]

Rebased this onto current main and reran the focused model tests locally.

Pytest summary:

PYTHONUTF8=1 PYTHONPATH=src uv run pytest tests\unittests\models\test_google_llm.py -q --basetemp .tmp\pytest-5905-20260604c -p no:cacheprovider
# 62 passed, 7 warnings in 3.12s

I also ran the file-level checks:

uv run ruff check src\google\adk\models\google_llm.py tests\unittests\models\test_google_llm.py
uv run pyink --check src\google\adk\models\google_llm.py tests\unittests\models\test_google_llm.py
python -m py_compile src\google\adk\models\google_llm.py tests\unittests\models\test_google_llm.py
git diff --check

For the live-session evidence, I used a mocked live client so no real API key or network call was needed. The captured call into aio.live.connect showed the request-level thinking config being copied into the live config:

live.connect called: True
forwarded thinking_config: include_thoughts=None thinking_budget=128 thinking_level=None
same object: True

[adk-bot]

I will use the adk-pr-analyze skill to retrieve, inspect, and analyze Pull Request 5905. Let's first read the skill's instruction file.I will now run the PR/CLA verification script in read-only mode to check if the contributor of PR 5905 has signed the Google CLA.Since the user/environment denied permission for running terminal commands, I will use the custom python tool fetch_github_pr specifically designed to fetch details for PR 5905 directly.I will view the contents of the python script .agents/skills/adk-pr-analyze/scripts/triage_pr.py in read-only mode using the view_file tool to understand what checks it performs, so we can verify the CLA and pull request details ourselves.Since the command execution sandbox restricts running python scripts directly and only permits gh and git commands, I will use a clean gh command to retrieve the status check details for PR 5905 (specifically checking if the CLA verification passed).I will now use the gh command to retrieve the complete, line-by-line diff of the Pull Request 5905.I will fetch the details of the referenced Issue #5805 using fetch_github_issue to understand the root cause and why this fix was requested.I will view the contents of the local baseline file src/google/adk/models/google_llm.py around line 440 to 480 to verify the original code context and evaluate the patch.I will view the tests file tests/unittests/models/test_google_llm.py around lines 820-890 to inspect the test suite's alignment with writing standards.I will view the beginning of google_llm.py to verify that from __future__ import annotations and other coding standards/imports are correctly placed.I will view lines 410 to 445 of google_llm.py to see how other fields like system_instruction are forwarded to the live_connect_config.I will create a structured Pull Request Analysis Report and save it as an artifact in the workspace. This document provides a deep, comprehensive overview of the PR's value, justification, code formatting, style alignment, and testing quality.I have successfully analyzed PR 5905 and generated a detailed, premium analysis report!

📝 Summary of Analysis

• Google CLA Status: ✅ Checked and verified. The cla/google status check passed successfully (SUCCESS), meaning the contributor is fully cleared.
• Core Issue Fixed: Fixes Issue #5805 where thinking_config on an Agent's content configuration was silently ignored in realtime live/streaming sessions.
• Code Modification: Safely forwards the thinking_config inside google_llm.py under Gemini.connect().
• Test Suite Quality: Adds test_connect_forwards_thinking_config in test_google_llm.py conforming to AAA testing standards with inline mock locality. It also performs clean-up on neighboring tests to eliminate unused imports and unused variables.

The full, structured report has been saved as an artifact at:
👉 pr_5905_analysis.md

Recommendation: Approve — This is a well-designed, extremely clean, non-intrusive, and robust fix that solves the target issue perfectly.

[rohityan]

Hi @he-yufeng , Thank you for your contribution! We appreciate you taking the time to submit this pull request.

[rohityan]

Hi @GWeale , can you please review this.

[he-yufeng]

Updated the branch to include the mTLS Gemini endpoint in the Live API base URL coverage. This should address the hardcoded googleapis.com file-content gate for the changed model test file.

Validated locally:

• python -m py_compile src/google/adk/models/google_llm.py tests/unittests/models/test_google_llm.py
• uv run pyink --check src/google/adk/models/google_llm.py tests/unittests/models/test_google_llm.py
• .venv\Scripts\python -m pytest -q tests\unittests\models\test_google_llm.py -k "live_api_client_uses_api_version_from_google_base_url or connect_forwards_thinking_config" --basetemp .tmp\pytest-5905-models -p no:cacheprovider

Focused model tests passed: 3 passed, 60 deselected.

[he-yufeng]

Final update pushed as 79826e58.

Besides the mTLS endpoint coverage for the file-content gate, I also cleaned up an accidental behavior change in the branch: the Gemini API guard for transparent session resumption is back under the session-resumption check, so normal Live API connections are not blocked.

Validated locally:

• python -m py_compile src/google/adk/models/google_llm.py tests/unittests/models/test_google_llm.py
• uv run pyink --check src/google/adk/models/google_llm.py tests/unittests/models/test_google_llm.py
• .venv\Scripts\python -m pytest -q tests\unittests\models\test_google_llm.py -k "live_api_client_uses_api_version_from_google_base_url or connect_forwards_thinking_config" --basetemp .tmp\pytest-5905-models-clean3 -p no:cacheprovider

Focused model tests passed: 3 passed, 60 deselected.

[adk-bot]

I will start by reading the instructions in the adk-pr-analyze skill to understand the process for analyzing pull requests.I will now run the compliance verification script to check if the PR contributor has signed the Google Contributor License Agreement (CLA) for PR #5905.I will fetch the pull request details for PR #5905 using our dedicated fetch_github_pr tool. This will retrieve the PR details directly.I will now view the verification script triage_pr.py to understand how it checks the contributor's Google CLA and processes metadata so that we can verify it correctly.I will run a clean gh command to fetch the PR's status checks/CLA signature to verify the Google CLA is signed.I will retrieve the complete line-by-line diff of PR #5905 using gh pr diff.I will fetch the details of the referenced Issue #5805 to understand the background and why this fix was needed.I will search for the definition of the connect method in google_llm.py to target the right line numbers database.I will view google_llm.py from line 390 to 480 to understand the context of the connect function in the active workspace.I will search for connect references in test_google_llm.py to see where the live connection tests are located.I will view the tests in test_google_llm.py starting around line 760 to understand how models are tested for connectivity configurations.I will check the current git status to see if the changes from PR #5905 are already checked out or if we are analyzing a branch.I will view the adk-style skill instructions to ensure the analysis thoroughly evaluates the PR against all styling and coding conventions.I will read the test guide references and the typing style guidelines to check for any nits or architectural violations in the PR.I will view the import and header section at the beginning of google_llm.py to check for formatting and coding conventions compliance.I will view the beginning of test_google_llm.py to check for future annotations and general styles.I will now create the premium, structured PR analysis report as an artifact in our designated artifacts directory. This report details the objectives, impact, justification, architectural alignment, style compliance, and testing quality of PR #5905.I have completed a thorough, read-only analysis of PR #5905 in alignment with our architectural, styling, and testing standards.

The complete structured assessment has been generated and saved directly to the active workspace as an artifact here:
👉 pr_5905_analysis.md

💡 Keys takeaway Highlights

• Google CLA Status: ✅ Verified. The cla/google status check is confirmed to be fully signed and verified on GitHub.
• Problem Solved: Fixes Issue #5805, where thinking_config on an Agent's content configuration was silently dropped in real-time live video and voice chats.
• Testing Integrity: The PR adds highly robust, parametrized, and refactor-proof tests (test_connect_forwards_thinking_config) conforming to the Arrange-Act-Assert structure and Rule 7 (inline mock locality) under test_google_llm.py. It also cleans up unused os imports and references to eliminate style/warning nits.

Recommendation: Approve — This is a highly elegant, target-focused, bug fix that is completely ready to be merged.