Repository navigation
thinking_config is silently dropped with use_interactions_api=True #7924
Description
🔴 Required Information
Describe the Bug:
With Gemini(use_interactions_api=True), generate_content_config.thinking_config.thinking_level is never sent. build_generation_config doesn't map it to the interactions generation_config.thinking_level, and logs no warning.
Steps to Reproduce:
- Create an
LlmAgentwithGemini(use_interactions_api=True)andthinking_config=types.ThinkingConfig(thinking_level=types.ThinkingLevel.HIGH). - Run it with DEBUG logging on
google_adk.google.adk.models.interactions_utils. - Check
Generation Configin the "Interactions API Request" log.
Expected Behavior:
Generation Config: {"thinking_level": "high"}
Observed Behavior:
Generation Config: {}. The model thinks at its default level.
Environment Details:
- ADK Library Version (pip show google-adk): 2.13.0 (also
main@ 8cd3b12) - Desktop OS: Linux
- Python Version (python -V): 3.13
Model Information:
- Are you using LiteLLM: No
- Which model is being used: gemini-3.8-flash (Vertex AI)
🟡 Optional Information
Regression:
No. It has never been sent.
Logs:
Generation Config:
{}
Additional Context:
The interactions API accepts generation_config.thinking_level (minimal | low | medium | high); sending "high" directly gave about 2.5× the thought tokens. Similar: #7870 (output_schema), #7376 (temperature/top_p).
Minimal Reproduction Code:
from google.adk.models import interactions_utils
from google.genai import types
config = types.GenerateContentConfig(
thinking_config=types.ThinkingConfig(thinking_level=types.ThinkingLevel.HIGH)
)
print(interactions_utils.build_generation_config(config)) # {}How often has this issue occurred?:
- Always (100%)
Activity
surajksharma07 commentedon Nov 2, 2026
Thanks @jhackshaw, nice catch and the token numbers really help. Reproduced this on 2.13.0 and on current main: build_generation_config returns {} for thinking_level=HIGH even though the interactions GenerationConfigParam has the field.
Mapping the level to lowercase and skipping UNSPECIFIED, like #7926 does, gave {"thinking_level": "high"} and test_interactions_utils.py stayed green. One small thing: thinking_budget (and include_thoughts) still get dropped without any log, so a warning for the budget like the sampling params have might be worth adding.
Could you check that on your side and confirm it behaves well across the levels on a couple of models, including the 400 for an unsupported level? Once that’s tested #7926 should be good to go for review.
jhackshaw commentedon Nov 2, 2026
Thanks for reproducing it! I tested this live on Vertex AI (
global, google-genai 2.34.0). Each request went through ADK's ownGeminimodel, once withuse_interactions_api=Trueand once with generateContent as a baseline. The prompt was a counting puzzle and each cell is the median thought tokens of 3 runs.With this PR
Thinking level is unsupported: THINKING_LEVEL_MINIMALthinking_level MINIMAL is not supported by this modelOn
main(one run each), the level never arrives:Unsupported level. An unsupported level surfaces as a
BadRequestError400 from ADK, with the API's message matching what generateContent does for the same request.thinking_budget/include_thoughts. I checked both against the Interactions schema in every supported google-genai release (2.30 through 2.34):thinking_budgethas no equivalent. It now logs a one-time warning, like the sampling params do. The warning points atthinking_level, or at turning offuse_interactions_api. Tested with budgets of 0, -1, and 1024.include_thoughtsdoes have an equivalent:generation_config.thinking_summaries("auto"/"none"). So the PR now mapsTrueto"auto"andFalseto"none"instead of warning. Unset sends nothing, which keeps the API's default. I verified it live: the raw API returns summaries only with"auto", and the default sends none, which matches generateContent.surajksharma07 commentedon Nov 2, 2026
Thanks @jhackshaw, this is really thorough, and the per-model numbers make it clear the level never landed on main. Went through the updated diff (315e77b): the level mapping, the one-time thinking_budget warning and include_thoughts -> thinking_summaries all look right. The budget test is also safe across the 0/-1/1024 runs because the autouse fixture clears _WARNED_SAMPLING_PARAMS.
One gap: with thinking_summaries="auto" the summaries only reach ADK users in SSE mode. On the non-streaming path the ThoughtStep branch in interactions_utils.py (~L751) keeps only the signature and drops step.summary, so include_thoughts=True still shows no thought text there. Could you map the summary to a Part(text=..., thought=True) next to the signature part, the way _handle_thought_summary does for streaming, or call it out as a follow-up?
Also the PR description still only talks about thinking_level, so it’d be good to add the thinking_summaries mapping and the budget warning. This isn’t in any release yet (still reproduces on 2.13.0 and current main), so I’ll flag #7926 for maintainer review once that’s sorted.
jhackshaw commentedon Nov 2, 2026
Thanks @surajksharma07, nice catch. I updated to handle the non-streaming path and the PR description as well!
fix: send the thinking level on the interactions API