A2A client renders thought parts together with final user-facing response

🔴 Required Information

Describe the Bug:
In A2A mode, the local ADK client displays both internal thought text and the final user-facing answer in one combined response.
The response payload contains multiple text parts, including one with metadata.adk_thought: true, but the client renders both instead of filtering thought content.

Steps to Reproduce:

Install google-adk and run an ADK server with A2A enabled.
Run a local ADK client against that A2A server.
Send a normal user request that produces a multi-part response.
Inspect the returned message/task parts: one part has metadata.adk_thought: true, another part is the final answer.
Observe that the client output shows both parts concatenated in user-visible output.

Expected Behavior:
Only non-thought/final content should be shown to the user.
Parts marked with adk_thought: true should be hidden from user-facing rendering.

Observed Behavior:
Thought/internal text is rendered to the user together with final answer text.

Environment Details:

ADK Library Version (pip show google-adk): <fill>
Desktop OS: Linux
Python Version (python -V): <fill>

Model Information:

Are you using LiteLLM: <Yes/No>
Which model is being used: <fill>

🟡 Optional Information

Regression:
Unknown / not yet verified.

Logs:

Sanitized payload shape observed:
parts[0]: kind=text, metadata.adk_thought=true, text=<internal reasoning>
parts[1]: kind=text, text=<final user answer>
Rendered output combines both parts.

Additional Context:
This appears to be a client-side rendering/filtering issue for thought parts in A2A responses.

Minimal Reproduction Code:

# Pseudocode (sanitized)
response = client.send_message("test request")
for i, p in enumerate(response.parts):
    print(i, p.kind, getattr(p, "metadata", None), getattr(p, "text", None))
# Observed: part with adk_thought=true is still rendered in final UI output.

How often has this issue occurred?:
Always (100%)

fix: filter thought parts from A2A client user-facing response

I've submitted a fix in #4686.

The root cause is in RemoteA2aAgent._handle_a2a_response(): while intermediate (working/submitted) events correctly mark all parts as thought=True, completed responses containing both thought and non-thought parts yield them all to consumers without filtering.

The fix adds thought part filtering before returning the event — only non-thought parts are kept for completed responses, while intermediate events remain untouched.

@EliavShbt

@EliavShbt Thanks for reporting!
This looks like a missing filter step in _handle_a2a_response (remote_a2a_agent.py) ,when a completed task artifact is processed, convert_a2a_task_to_event() returns all parts including thought-flagged ones without stripping them before surfacing to the consumer.

As a workaround, filtering after the conversion with event.content.parts = [p for p in event.content.parts if not p.thought] should stop the thought text from leaking into the user facing response, could you try that and let us know if it clears things up on your end?

There's also a community PR (#4686) already targeting this exact gap worth validating the workaround against your setup before going that route separately.
Thank you.

fix: filter thought parts from A2A client user-facing response

@EliavShbt
Thanks for reporting! This looks like a missing filter step in _handle_a2a_response (remote_a2a_agent.py) ,when a completed task artifact is processed, convert_a2a_task_to_event() returns all parts including thought-flagged ones without stripping them before surfacing to the consumer.

As a workaround, filtering after the conversion with event.content.parts = [p for p in event.content.parts if not p.thought] should stop the thought text from leaking into the user facing response, could you try that and let us know if it clears things up on your end?

There's also a community PR (#4686) already targeting this exact gap worth validating the workaround against your setup before going that route separately.
Thank you.

Thanks for the suggestion — I tested the workaround and it doesn't fix my case, and I also traced the root cause.

The filter [p for p in event.content.parts if not p.thought] has no effect because p.thought is always None on those parts. The thought flag is present on the wire (customMetadata → a2a:response → artifacts → parts → metadata.adk_thought: true), but it never gets restored onto the GenAI Part during deserialization.

Root cause: convert_a2a_part_to_genai_part() in part_converter.py discards the metadata on inbound conversion:

# Current — drops adk_thought metadata:
if isinstance(part, a2a_types.TextPart):
    return genai_types.Part(text=part.text)

The outbound direction (convert_genai_part_to_a2a_part) correctly serializes part.thought into adk_thought, but the inverse never reads it back. So the round-trip is broken and all parts arrive with thought=None.

Fix needed in part_converter.py:

if isinstance(part, a2a_types.TextPart):
    genai_part = genai_types.Part(text=part.text)
    if part.metadata and part.metadata.get(_get_adk_metadata_key('thought')):
        genai_part.thought = True
    return genai_part

With this, thought is correctly restored from wire metadata, and the filter in PR #4686 then works as intended. Without it, #4686 only catches parts tagged by the submitted/working state logic — completed artifact parts with adk_thought in metadata still leak through.

Minimal payload demonstrating the issue:

{
  "content": {
    "parts": [
      { "text": "<internal reasoning text>" },
      { "text": "<final user-facing answer>" }
    ],
    "role": "model"
  },
  "customMetadata": {
    "a2a:response": {
      "artifacts": [{
        "parts": [
          { "kind": "text", "metadata": { "adk_thought": true }, "text": "<internal reasoning text>" },
          { "kind": "text", "text": "<final user-facing answer>" }
        ]
      }]
    }
  }
}

fix: filter thought parts from A2A client user-facing response

@EliavShbt

@EliavShbt your diagnosis is correct.

convert_a2a_part_to_genai_part() in part_converter.py was discarding the adk_thought metadata on inbound conversion, so p.thought was always None and the filter in #4686 had no effect on completed artifact parts.

This has already been fixed in main via commit e59929e1 ("fix: Propagate thought from A2A TextPart metadata to GenAI Part", 2026-02-25), which restores thought=True from wire metadata during deserialization — after which the filtering in #4686 works end-to-end.

The fix isn't in a released version yet, but you can pip-install directly from main or apply both changes locally to unblock yourself.

@surajksharma07

Thanks @surajksharma07 for confirming the root cause in convert_a2a_part_to_genai_part(). My PR #4686 addresses the filtering side in _handle_a2a_response, and the round-trip fix for adk_thought metadata preservation would complement it. Happy to coordinate if you're working on the part_converter.py side.

test: add tests for adk_thought metadata round-trip in part_converter

fix: filter thought parts from A2A client user-facing response

@seanzhou1023

@seanzhou1023 Please have a look into it once.

fix: filter thought parts from A2A client user-facing response

test: add tests for adk_thought metadata round-trip in part_converter

@EliavShbt

@EliavShbt — the root cause you traced (convert_a2a_part_to_genai_part discarding adk_thought on inbound conversion) was fixed in v1.27.0, so upgrading to v1.27.0 or later should fully resolve the thought-part leakage you were seeing.

PR #4686 from @weiguangli-io adds complementary filtering in _handle_a2a_response and is still under review for additional coverage.

Since we haven't heard back in a while, going ahead and closing this out — feel free to reopen if you're still hitting it after upgrading!

fix: filter thought parts from A2A client user-facing response

Navigation Menu

Repository navigation

A2A client renders thought parts together with final user-facing response #4676

🔴 Required Information

🟡 Optional Information

7 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

A2A client renders thought parts together with final user-facing response #4676

Description

🔴 Required Information

🟡 Optional Information

Activity

weiguangli-io commented on Mar 3, 2026

klateefa commented on Mar 4, 2026

EliavShbt commented on Mar 5, 2026

surajksharma07 commented on Mar 6, 2026

weiguangli-io commented on Mar 9, 2026

7 remaining items

surajksharma07 commented on Mar 26, 2026

surajksharma07 commented on May 27, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions