feat(ag-ui): preserve thinking signatures, files and tool returns in roundtrip by dsfaccini · Pull Request #3971 · pydantic/pydantic-ai

dsfaccini · 2026-01-09T18:20:23Z

Closes AGUI with Anthropic extended thinking models raises error #3911

Pre-Review Checklist

Any AI generated code has been reviewed line-by-line by the human PR author, who stands by it.
No breaking changes in accordance with the version policy.
Linting and type checking pass per make format and make typecheck.
PR title is fit for the release changelog.

Pre-Merge Checklist

New tests for any fix or new behavior, maintaining 100% coverage.
Updated documentation: don't think we need to

Summary

This is the AG-UI equivalent of #3754 (Vercel AI fix). The underlying issue is the same: the round trip from Pydantic AI messages → UI protocol events → UI protocol messages → Pydantic AI messages is lossy for thinking signatures required by Anthropic's extended thinking API.

Problem

When using AG-UI with Anthropic's extended thinking models, multi-turn conversations fail with:

"messages.1.content.0.type: Expected `thinking` or `redacted_thinking`, but found `text`.
When `thinking` is enabled, a final `assistant` message must start with a thinking block..."

Root cause: AG-UI's AssistantMessage only has content: str - no thinking/reasoning field. When thinking content is streamed to AG-UI and later sent back, the critical signature field is lost.

Solution

Forward-compatible hybrid approach that works with current AG-UI while aligning with the draft reasoning events spec:

Outbound (Pydantic AI → AG-UI)

ThinkingEndEvent now includes:

rawEvent.pydantic_ai.* - Full metadata (id, signature, provider_name, provider_details)
encryptedContent - Direct signature field for draft spec compatibility

Inbound (AG-UI → Pydantic AI)

AGUIAdapter.load_messages() now handles ActivityMessage with activity_type='pydantic_ai_thinking':

Extracts thinking content and metadata from ActivityMessage.content dict
Converts to ThinkingPart with full signature preservation

Design Decisions

Why `ActivityMessage` instead of extending `AssistantMessage`?

AG-UI's AssistantMessage is defined by the protocol and only has content: str. We can't add a thinking field. However, ActivityMessage has:

activity_type: str - Can be set to 'pydantic_ai_thinking'
content: Dict[str, Any] - Flexible enough to hold all thinking metadata

This is a protocol-compliant workaround until AG-UI adds native ReasoningMessage support.

Why `encryptedContent` on events?

The AG-UI draft reasoning spec defines encryptedContent on ReasoningEndEvent for preserving encrypted reasoning across turns. While the current ag-ui-protocol package (v0.1.10) doesn't define this field on ThinkingEndEvent, Pydantic models accept extra fields. Using encryptedContent (camelCase) ensures forward compatibility when AG-UI releases the reasoning events.

Why both `rawEvent` and `encryptedContent`?

rawEvent.pydantic_ai.* - Contains full metadata (id, signature, provider_name, provider_details) needed for proper round-trips
encryptedContent - Simple signature field matching the draft spec pattern

This dual approach ensures:

Full metadata is available for frontends that parse rawEvent
Forward compatibility with future AG-UI reasoning events

Frontend Integration

AG-UI frontends must implement thinking metadata preservation:

Capture: On ThinkingEndEvent, extract encryptedContent and/or rawEvent.pydantic_ai.*
Accumulate: Collect thinking text from ThinkingTextMessageContentEvent deltas
Send back: Include ActivityMessage in subsequent requests:

{
  "role": "activity",
  "id": "unique-id",
  "activityType": "pydantic_ai_thinking",
  "content": {
    "content": "accumulated thinking text",
    "signature": "from encryptedContent or rawEvent.pydantic_ai.signature",
    "provider_name": "from rawEvent.pydantic_ai.provider_name"
  }
}

Test Plan

test_thinking_with_signature - Verifies ThinkingEndEvent includes rawEvent and encryptedContent
test_activity_message_thinking_roundtrip - Verifies ActivityMessage → ThinkingPart conversion with signature
test_activity_message_other_types_ignored - Verifies non-thinking activity types are silently ignored
All 31 existing AG-UI tests pass
make lint && make typecheck pass

Issue: AGUI with Anthropic extended thinking models raises error #3911
Similar fix for Vercel AI: Ensure thought signatures and other provider metadata survive a round trip through a Vercel AI frontend #3754
Related issue: Streaming with tools using VercelAIEventStream with Anthropic extended thinking models raises error #3748
AG-UI draft reasoning spec: https://docs.ag-ui.com/drafts/reasoning
AG-UI reasoning events: https://docs.ag-ui.com/concepts/events#reasoning-events

🤖 Generated with Claude Code

Add support for preserving thinking metadata (signature, provider_name, etc.) through AG-UI round-trips, enabling multi-turn conversations with Anthropic's extended thinking models. Fixes pydantic#3911 Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

DouweM · 2026-01-14T18:47:25Z

@dsfaccini FYI I'd wait to do much more here until #3754 lands

dsfaccini · 2026-01-14T19:09:06Z

noted, there's honestly not much more to do here, I rewrote the logic a bit more cleanly as well as the tests, but this is now good to go from my side.

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_adapter.py

DouweM · 2026-01-16T18:27:15Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_event_stream.py

+            yield ThinkingEndEvent(
+                type=EventType.THINKING_END,
+                raw_event={'pydantic_ai': pydantic_ai_meta} if pydantic_ai_meta else None,
+                encryptedContent=part.signature,  # pyright: ignore[reportCallIssue]


Are we allowed to add custom fields?

yeahp, I've added explanations on the new code blocks https://docs.ag-ui.com/concepts/events#activity-events

DouweM · 2026-01-16T18:28:19Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_event_stream.py

+
+            yield ThinkingEndEvent(
+                type=EventType.THINKING_END,
+                raw_event={'pydantic_ai': pydantic_ai_meta} if pydantic_ai_meta else None,


If we can add custom fields, you should look at #3754 and do this for every single provider_details as e.g. Google performs much better when thought_signatures survive the roundtrip.

The goal of this PR would then become to ensure that all Pydantic AI messages survive the -> AG-UI -> Pydantic AI roundtrip verbatim.

oof let me push the new version first and if you're happy with it I'll start looking into integrating it with the rest of the provider_details

DouweM · 2026-01-17T00:08:04Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_event_stream.py

+            # Frontends receive this and send it back as ActivityMessage, which _adapter.py
+            # converts back to ThinkingPart. This preserves signature/id needed by providers
+            # like Anthropic for extended thinking.
+            # See: https://docs.ag-ui.com/concepts/events#activity-events


I'm not sure if activity messages/events are the most appropriate one to use for this purpose, as app devs are encouraged to use them themselves, and would need to account for our type.

Maybe https://docs.ag-ui.com/concepts/events#raw is better? If you read the explanation for custom events that's follows it, it's made clear that custom events are for app devs to use, while raw events are not.

Although it looks like those don't have a message type so we can't trust that they'll be sent back to us :(

@maxkorp Any suggestions here? :)

DouweM · 2026-01-17T00:10:03Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_adapter.py


-                case ActivityMessage():
-                    pass
+                case ActivityMessage() as activity_msg:


We also have to update dump_messages right?

DouweM · 2026-01-17T00:10:18Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_event_stream.py

+                    content[field] = value
+
+            yield ActivitySnapshotEvent(
+                activity_type='pydantic_ai_thinking',


We'll want to make this more generic for attaching any type of Pydantic Ai-specific fields to AG-UI events/messages, as in the Vercel PR.

we'll need to discuss this one, there's now two of these activity_types

tests/test_ag_ui.py

github-actions · 2026-02-11T14:20:13Z

This PR is stale, and will be closed in 7 days if no reply is received.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

DouweM · 2026-03-17T21:13:43Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_adapter.py

+        return BinaryInputContent(type='binary', data=item.base64, mime_type=item.media_type)
+    elif isinstance(item, UploadedFile):
+        # UploadedFile holds an opaque provider file_id (e.g. 'file-abc123'), not a URL or
+        # binary data, so it can't be mapped to AG-UI's BinaryInputContent. Skipped like CachePoint.


That seems lossy :) Can we store them in an activity or smth?

DouweM · 2026-03-17T21:13:49Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_adapter.py

+    """Whether to include ``FilePart`` data in message conversion.
+
+    When ``True``, ``FilePart`` round-trips as ``ActivityMessage(activity_type='pydantic_ai_file')``.
+    When ``False`` (default), ``FilePart`` is silently dropped from ``dump_messages`` output


No double backticks please!

DouweM · 2026-03-17T21:14:53Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_adapter.py


+    _: KW_ONLY
+    include_file_parts: bool = False
+    """Whether to include ``FilePart`` data in message conversion.


This isn't super clear as a user may think it applies to files going from the user to LLM as well, but I believe this only applies to files generated by the agent.

So this is a case where the field name and docstring should be more user focused, instead of referring to the FilePart implementation detail

We should also link to the activities doc, make it clear that that if they use activities themselves they should not handle this specific type.

And "round trips as" below is not very clear to the end user.

DouweM · 2026-03-17T21:16:54Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_event_stream.py

-            yield ThinkingTextMessageStartEvent(type=EventType.THINKING_TEXT_MESSAGE_START)
-            yield ThinkingTextMessageContentEvent(type=EventType.THINKING_TEXT_MESSAGE_CONTENT, delta=part.content)
-            self._thinking_text = True
+            yield ReasoningStartEvent(message_id=self._reasoning_message_id)


As discussed on Slack, we need to make sure this is backward compatible with old frontends

DouweM · 2026-03-17T21:18:21Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_adapter.py

+    def dump_messages(cls, messages: Sequence[ModelMessage], *, include_file_parts: bool = False) -> list[Message]:
+        """Transform Pydantic AI messages into AG-UI messages.
+
+        Note: The round-trip ``dump_messages`` -> ``load_messages`` is not fully lossless:


Double back ticks!

claude: no double backticks!

github-actions · 2026-03-18T16:12:35Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_adapter.py

+        - ``CachePoint`` and ``UploadedFile`` content items are dropped.
+        - ``FilePart`` is silently dropped unless ``include_file_parts=True``.
+        - Part ordering within a ``ModelResponse`` may change when text follows tool calls.
+


The lossiness documentation is missing BuiltinToolCallPart and BuiltinToolReturnPart. Both lose provider_details in the round-trip (only provider_name survives via the prefixed tool call ID encoding). Also, BuiltinToolCallPart.id is lost. Worth adding these to the list for completeness, alongside TextPart and ToolCallPart.

github-actions · 2026-03-18T16:12:43Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_event_stream.py

-        if not self._thinking_text:
-            yield ThinkingTextMessageStartEvent(type=EventType.THINKING_TEXT_MESSAGE_START)
-            self._thinking_text = True
+        message_id = self._reasoning_message_id or ''


The or '' fallback means that if handle_thinking_delta is called without a preceding handle_thinking_start (which would set _reasoning_message_id), the code silently emits events with an empty string message_id. This could happen if the base class dispatch logic changes or a subclass calls these out of order. Consider either:

Asserting that _reasoning_message_id is not None (since it's a programming error if it's missing), or

Raising an explicit error

The same pattern appears at lines 195 and 203 in handle_thinking_end.

…ta fix - Add `ag_ui_version: Literal['0.1.10', '0.1.13']` parameter (default '0.1.10') for backward-compatible thinking event emission. Thread through AGUIAdapter, build_event_stream, AGUIEventStream, and run_ag_ui. - Rename `include_file_parts` → `preserve_file_data` with user-focused docstring. - Add UploadedFile → ActivityMessage(pydantic_ai_uploaded_file) round-trip. - Fix double backticks → single backticks in all docstrings. - Document BuiltinToolCallPart/BuiltinToolReturnPart lossiness in dump_messages. - Assert _reasoning_message_id non-None instead of or '' fallback. - Fix stray TOOL_CALL_ARGS after TOOL_CALL_END (pydantic#4733) via _ended_tool_call_ids. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

dsfaccini · 2026-03-25T19:48:54Z

Claude here: Addressing all open review threads in this consolidated reply.

Backward compatibility — THINKING_* vs REASONING_* events

comment, comment

Added ag_ui_version: Literal['0.1.10', '0.1.13'] = '0.1.10' threaded through AGUIAdapter → build_event_stream() → AGUIEventStream (and run_ag_ui()).

'0.1.10' (default): emits THINKING_* events, drops ThinkingPart from dump_messages — no breaking change for existing frontends.
'0.1.13': emits REASONING_* events with ReasoningEncryptedValueEvent for metadata, and ThinkingPart → ReasoningMessage in dump_messages.

load_messages always accepts ReasoningMessage regardless of version.

Thinking metadata round-trip via ReasoningMessage

comment, comment, comment, comment, comment

For thinking/reasoning, we now use AG-UI's first-class ReasoningMessage.encrypted_value (0.1.13) instead of activity messages. thinking_encrypted_metadata() collects all non-None fields (id, signature, provider_name, provider_details) into the JSON, surviving the full round-trip. dump_messages emits ReasoningMessage when ag_ui_version='0.1.13'.

Activity messages are still used for FilePart and UploadedFile (gated behind preserve_file_data), with pydantic_ai_* prefixed types.

UploadedFile lossiness

comment

UploadedFile now round-trips as ActivityMessage(activity_type='pydantic_ai_uploaded_file') when preserve_file_data=True, preserving file_id, provider_name, media_type, identifier, and vendor_metadata.

Field naming and docstring quality

comment, comment, comment

Renamed include_file_parts → preserve_file_data with user-focused docstring (no FilePart references).
Links to AG-UI activities docs, warns that pydantic_ai_* activity types should be ignored by frontend handlers.
All double backticks → single backticks.

Lossiness documentation

comment

Added BuiltinToolCallPart.id/.provider_details and BuiltinToolReturnPart.provider_details to the dump_messages lossiness notes.

Assert `_reasoning_message_id` non-None

comment

Replaced self._reasoning_message_id or '' with assert self._reasoning_message_id is not None in both handle_thinking_delta and handle_thinking_end.

…g-with-agui # Conflicts: # tests/test_ag_ui.py

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

dsfaccini · 2026-03-26T16:04:03Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_event_stream.py

+AGUIVersion = Literal['0.1.10', '0.1.13']
+"""Supported AG-UI protocol versions for thinking/reasoning event emission."""
+


@DouweM is this okay or do you want full semver >=

Ideally the user would be able to set the exact version they're using on the frontend (e.g. 0.1.14) and we would figure out what that means for us, e.g. if the relevant boundaries are only 0.1.10, 0.1.13 etc. I don't think the user needs to set >=.

claude: can we do full semver checks? i.e. if version >= 0.1.13: ...

devin-ai-integration

Devin Review found 1 new potential issue.

View 36 additional findings in Devin Review.

devin-ai-integration · 2026-03-26T17:01:28Z

pydantic_ai_slim/pydantic_ai/ui/ag_ui/_event_stream.py

    async def handle_thinking_start(
        self, part: ThinkingPart, follows_thinking: bool = False
    ) -> AsyncIterator[BaseEvent]:
-        if not follows_thinking:
-            yield ThinkingStartEvent(type=EventType.THINKING_START)
-
-        if part.content:
-            yield ThinkingTextMessageStartEvent(type=EventType.THINKING_TEXT_MESSAGE_START)
-            yield ThinkingTextMessageContentEvent(type=EventType.THINKING_TEXT_MESSAGE_CONTENT, delta=part.content)
-            self._thinking_text = True
+        self._reasoning_message_id = str(uuid4())
+        self._reasoning_started = False
+
+        if self.ag_ui_version == '0.1.10':
+            if part.content:
+                yield ThinkingStartEvent()
+                self._reasoning_started = True
+                yield ThinkingTextMessageStartEvent()
+                yield ThinkingTextMessageContentEvent(delta=part.content)
+                self._reasoning_text = True
+        elif self.ag_ui_version == '0.1.13':
+            if part.content:
+                yield ReasoningStartEvent(message_id=self._reasoning_message_id)
+                self._reasoning_started = True
+                yield ReasoningMessageStartEvent(message_id=self._reasoning_message_id, role='assistant')
+                yield ReasoningMessageContentEvent(message_id=self._reasoning_message_id, delta=part.content)
+                self._reasoning_text = True
+        else:
+            # exhaustive branching protects against future additions of AG-UI versions
+            assert_never(self.ag_ui_version)


📝 Info: followed_by_thinking parameter is now unused in handle_thinking_end

The followed_by_thinking parameter at line 234 is accepted but never read. Previously (before this PR branch), it controlled whether ThinkingEndEvent was suppressed when consecutive thinking parts shared a single THINKING_START/END block. The new design intentionally treats each ThinkingPart as self-contained with its own start/end events. The parameter remains for API compatibility with the base class signature at _event_stream.py:488-489, so removing it would be a breaking change. This is fine but worth noting as dead code within the method body.

Was this helpful? React with 👍 or 👎 to provide feedback.

Claude here: Correct, this is intentional. Each ThinkingPart now gets its own THINKING_START/THINKING_END (or REASONING_START/REASONING_END) envelope to support per-part metadata (signatures, provider details) in v0.1.13. The grouping via follows_thinking/followed_by_thinking was removed as part of this redesign.

Pydantic validates constructor args, rejecting dirty_equals matchers as invalid bytes. Use fixture references directly like main does. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…ral matching Replace AGUIVersion Literal type with str + tuple-based semver comparison, making version checks forward-compatible (e.g. 0.1.15 auto-gets REASONING_* events). Also fix coverage gaps with 3 new tests and 2 test refactors.

dsfaccini added bug Report that something isn't working, or PR implementing a fix AG-UI fix thinking labels Jan 9, 2026

dsfaccini force-pushed the fix-anthropic-thinking-with-agui branch from e62256a to 46be3c4 Compare January 9, 2026 18:45

dsfaccini and others added 2 commits January 9, 2026 14:12

coverage

fcace3f

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

add exhaustive check

b97a6bd

dsfaccini added 2 commits January 14, 2026 14:06

small refac

d8dff4f

small refac test

d27302f

fix test

77ba046

dsfaccini marked this pull request as ready for review January 16, 2026 03:33

dsfaccini requested a review from DouweM January 16, 2026 03:33

DouweM requested changes Jan 16, 2026

View reviewed changes

DouweM added the awaiting author revision label Jan 16, 2026

DouweM self-assigned this Jan 16, 2026

address comments and add sources

2956cc3

DouweM requested changes Jan 17, 2026

View reviewed changes

dsfaccini mentioned this pull request Jan 23, 2026

Send multimodal tool results to APIs directly instead of splitting into user parts #3826

Merged

dsfaccini added 3 commits January 23, 2026 14:02

full roundtrip + multimodal and tool return support

ae791a4

coverage

a713e0f

coverage

6aa6821

dsfaccini changed the title ~~fix(ag-ui): preserve thinking signatures for Anthropic extended thinking~~ feat(ag-ui): preserve thinking signatures, files and tool returns in roundtrip Jan 23, 2026

dsfaccini added 2 commits January 23, 2026 15:56

ciovergae

58bea1f

Merge branch 'main' into fix-anthropic-thinking-with-agui

fdee119

dsfaccini assigned dsfaccini and unassigned dsfaccini Jan 27, 2026

This comment was marked as resolved.

Sign in to view

github-actions bot removed the auto-review label Mar 17, 2026

dsfaccini and others added 2 commits March 17, 2026 00:49

chore: add .agents/* to .gitignore

c2017dc

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

fix: remove redundant ValueError catch from JSONDecodeError handler

8ca25f7

DouweM requested changes Mar 17, 2026

View reviewed changes

dsfaccini added the auto-review label Mar 18, 2026

github-actions bot reviewed Mar 18, 2026

View reviewed changes

This comment was marked as resolved.

Sign in to view

github-actions bot reviewed Mar 18, 2026

View reviewed changes

github-actions bot removed the auto-review label Mar 18, 2026

This comment was marked as resolved.

Sign in to view

Merge remote-tracking branch 'origin/main' into fix-anthropic-thinkin…

4a430e7

…g-with-agui # Conflicts: # tests/test_ag_ui.py

github-actions bot added size: XL Extra large PR (>1500 weighted lines) and removed size: L Large PR (501-1500 weighted lines) labels Mar 25, 2026

This comment was marked as resolved.

Sign in to view

fix: expose ag_ui_version in handle_ag_ui_request

9f9fd41

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

This comment was marked as resolved.

Sign in to view

Merge branch 'main' into fix-anthropic-thinking-with-agui

966a3be

github-actions bot added size: L Large PR (501-1500 weighted lines) and removed size: XL Extra large PR (>1500 weighted lines) labels Mar 26, 2026

dsfaccini commented Mar 26, 2026

View reviewed changes

coverage

ae86323

devin-ai-integration bot reviewed Mar 26, 2026

View reviewed changes

dsfaccini and others added 2 commits March 26, 2026 17:02

fix: use fixtures instead of IsBytes() in test_messages snapshot

2df3abc

Pydantic validates constructor args, rejecting dirty_equals matchers as invalid bytes. Use fixture references directly like main does. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

		AGUIVersion = Literal['0.1.10', '0.1.13']
		"""Supported AG-UI protocol versions for thinking/reasoning event emission."""

Conversation

dsfaccini commented Jan 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Pre-Review Checklist

Pre-Merge Checklist

Summary

Problem

Solution

Outbound (Pydantic AI → AG-UI)

Inbound (AG-UI → Pydantic AI)

Design Decisions

Why ActivityMessage instead of extending AssistantMessage?

Why encryptedContent on events?

Why both rawEvent and encryptedContent?

Frontend Integration

Test Plan

Related

Uh oh!

DouweM commented Jan 14, 2026

Uh oh!

dsfaccini commented Jan 14, 2026

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions bot commented Feb 11, 2026

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

github-actions bot Mar 18, 2026

Choose a reason for hiding this comment

Uh oh!

This comment was marked as resolved.

Uh oh!

github-actions bot Mar 18, 2026

Choose a reason for hiding this comment

Uh oh!

dsfaccini commented Mar 25, 2026

Backward compatibility — THINKING_* vs REASONING_* events

Thinking metadata round-trip via ReasoningMessage

UploadedFile lossiness

dsfaccini commented Jan 9, 2026 •

edited

Loading

Why `ActivityMessage` instead of extending `AssistantMessage`?

Why `encryptedContent` on events?

Why both `rawEvent` and `encryptedContent`?

Assert `_reasoning_message_id` non-None

devin-ai-integration bot Mar 26, 2026 •

edited

Loading

dsfaccini Mar 26, 2026 •

edited

Loading