fix: reasoning item streaming in mux

[TejasGhatte]

Summary

Fixes a bug where reasoning models (e.g. DeepSeek Reasoner) streaming through the Chat → Responses → Anthropic SSE pipeline produced orphan thinking blocks — content_block_delta events referencing an index that no prior content_block_start had opened. Strict clients like Claude Code reject this with "Content block not found", abort the stream, and report a client disconnect. Additionally, fixes the Agent SDK's non-streaming parser failing with "Missing required field in assistant message: 'signature'" when converted (non-Anthropic) reasoning blocks omit the signature field.

Changes

• Reasoning item lifecycle: Reasoning content now opens a dedicated output_item.added (type=reasoning) before emitting any reasoning deltas, and closes it with output_item.done before the text or tool-call item opens. This ensures content_block_start type=thinking always precedes any thinking_delta in the Anthropic SSE layer, and blocks never overlap.
• Output index ordering: Reasoning is assigned output index 0, with the text item receiving a higher index. Previously both reasoning and text deltas were hardcoded to index 0, causing collisions.
• Reasoning item state tracking: Added ReasoningItemAdded, ReasoningItemClosed, ReasoningOutputIndex, ReasoningBuffer, and TextOutputIndex fields to ChatToResponsesStreamState so the reasoning block's lifecycle is tracked independently from the text block.
• Signature defaulting: convertBifrostReasoningToAnthropicThinking now defaults Signature to an empty string pointer instead of nil for converted (non-Anthropic) reasoning blocks, satisfying the Agent SDK's required-field check.
• Removed empty delta workaround: The previous workaround that emitted an empty text delta for reasoning-only responses is removed; reasoning models now emit proper reasoning items instead.
• Tests: Added TestToBifrostResponsesStreamResponse_ReasoningOpensThinkingBlock (schema layer) and TestReasoningStream_NoOrphanThinkingBlock / TestConvertBifrostReasoning_SignaturePresent (Anthropic provider layer) to assert correct block ordering, ItemID propagation, and signature presence.

Type of change

• Bug fix
• Feature
• Refactor
• Documentation
• Chore/CI

Affected areas

• Core (Go)
• Transports (HTTP)
• Providers/Integrations
• Plugins
• UI (React)
• Docs

How to test

go test ./core/schemas/... -run TestToBifrostResponsesStreamResponse_ReasoningOpensThinkingBlock -v
go test ./core/providers/anthropic/... -run TestReasoningStream_NoOrphanThinkingBlock -v
go test ./core/providers/anthropic/... -run TestConvertBifrostReasoning_SignaturePresent -v
go test ./...

Expected: all three new tests pass, and no existing tests regress. When streaming a reasoning model (e.g. DeepSeek Reasoner) through the Anthropic SSE endpoint, Claude Code no longer disconnects with "Content block not found".

Screenshots/Recordings

N/A

Breaking changes

• Yes
• No

Related issues

Security considerations

No security implications. Changes are limited to stream event ordering and field defaulting in the reasoning/thinking block conversion path.

Checklist

• I read docs/contributing/README.md and followed the guidelines
• I added/updated tests where appropriate
• I updated documentation where needed
• I verified builds succeed (Go and UI)
• I verified the CI pipeline passes locally if applicable

[coderabbitai]

Warning

Review limit reached

You’ve reached a temporary PR review limit under our Fair Usage Limits Policy.

Your recent review volume is higher than typical usage, so adaptive limits are currently applied.

Next review available in: 42 minutes

Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available.
You're only billed for reviews past your plan's rate limits ($0.25/file).

How can I continue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews.

How do review limits work?

CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability.

For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: cf52ad6f-7f53-4bef-8216-a9762284021d

📥 Commits

Reviewing files that changed from the base of the PR and between 36391e1 and 31d9c6a.

📒 Files selected for processing (4)

• core/providers/anthropic/reasoningstream_test.go
• core/providers/anthropic/responses.go
• core/schemas/mux.go
• core/schemas/mux_test.go

📝 Walkthrough

Walkthrough

The change gives reasoning its own Responses output item, enforces ordering and closure before text or tool output, places reasoning first in terminal responses, and ensures Anthropic thinking blocks always include signatures. Tests cover stream event integrity, ordering, indices, and signature presence.

Changes

Reasoning stream handling

Layer / File(s)	Summary
Responses reasoning lifecycle core/schemas/mux.go, core/schemas/mux_test.go	Reasoning and text output indices now have separate lifecycle tracking; reasoning opens before deltas, closes before text or tools, and is emitted first in terminal output. Tests validate event ordering, item IDs, and indices.
Anthropic thinking conversion core/providers/anthropic/responses.go, core/providers/anthropic/reasoningstream_test.go	Converted and synthesized thinking blocks now receive non-nil signatures. Tests validate Anthropic SSE block integrity and signature presence.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
  participant ChatStream
  participant ToBifrostResponsesStreamResponse
  participant AnthropicSSEConversion
  ChatStream->>ToBifrostResponsesStreamResponse: reasoning delta
  ToBifrostResponsesStreamResponse->>AnthropicSSEConversion: reasoning output item and delta
  ChatStream->>ToBifrostResponsesStreamResponse: text or tool delta
  ToBifrostResponsesStreamResponse->>AnthropicSSEConversion: close reasoning item
  ToBifrostResponsesStreamResponse->>AnthropicSSEConversion: emit text or tool output

Loading

Possibly related issues

• Issue 5149: The changes address Anthropic reasoning-stream ordering, closure, buffering, and signature handling.

Suggested reviewers: akshaydeo, roroghost17

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly matches the main change: reasoning item streaming fixes in the mux path.
Description check	✅ Passed	The PR description follows the template and covers summary, changes, testing, risks, and checklist, with only minor gaps in related issues/checklist completion.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch 07-16-fix_reasoning_item_streaming_in_mux

---

_{Comment @coderabbitai help to get the list of available commands.}

[TejasGhatte]

• feat: configurable keep alive duration in network config #5328
• fix: reasoning item streaming in mux #5286 👈 (View in Graphite)
• fix: handle line by line errors in streaming #5262
• dev

This stack of pull requests is managed by Graphite. Learn more about stacking.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Confidence Score: 5/5

No additional blocking issue qualifies for this follow-up review.

• The updated stream opens and closes the initial reasoning item before later output items.
• The new reasoning state is reset on both pool acquisition and release.
• No distinct unresolved production failure was found beyond the previously reported cases.

Important Files Changed

Filename	Overview
core/schemas/mux.go	Adds independent reasoning-item state, lifecycle events, buffering, and output-index handling.
core/providers/anthropic/responses.go	Ensures converted Anthropic thinking blocks include a signature field.
core/providers/anthropic/reasoningstream_test.go	Adds end-to-end checks for valid Anthropic thinking-block ordering and signature presence.
core/schemas/mux_test.go	Adds schema-level coverage for reasoning item lifecycle and output-index ordering.

_{Reviews (3): Last reviewed commit: "fix: reasoning item streaming in mux" | Re-trigger Greptile}

[coderabbitai]

🧹 Nitpick comments (2)

core/schemas/mux.go (1)

2225-2245: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Reasoning output items never carry a Status, unlike text/tool items.

The reasoning doneItem built in closeReasoningItem (Lines 1727-1735) and the terminal reasoning message built here both omit Status, whereas the equivalent text (Status: &statusFinal) and tool-call done items always set it. If downstream consumers of the Responses stream key off status (e.g. "completed") to know a reasoning item finished, this asymmetry could be surfaced as another spec-compliance gap for strict clients — the same category of issue this PR is otherwise fixing.

♻️ Suggested fix

+		statusCompleted := "completed"
 		doneItem := &ResponsesMessage{
 			Type: &reasoningType,
 			Role: &role,
+			Status: &statusCompleted,
 			Content: &ResponsesMessageContent{

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@core/schemas/mux.go` around lines 2225 - 2245, Set the completed status on
reasoning output items in both closeReasoningItem and the terminal
reasoning-message construction shown here. Reuse the existing statusFinal symbol
and ensure each reasoning doneItem or ResponsesMessage includes Status alongside
its type, role, content, and optional ID, matching text and tool-call completion
behavior.

core/providers/anthropic/reasoningstream_test.go (1)

39-39: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Local p() pointer helper duplicates the existing schemas.Ptr().

schemas.Ptr is already used for this exact purpose elsewhere in this same PR (e.g. core/schemas/mux_test.go). Reusing it here avoids a redundant local helper.

♻️ Suggested fix

-	p := func(s string) *string { return &s }
 	chunks := []*schemas.BifrostChatResponse{
-		reasoningStreamChunk("c1", &schemas.ChatStreamResponseChoiceDelta{Role: p("assistant"), Reasoning: p("")}, nil),
+		reasoningStreamChunk("c1", &schemas.ChatStreamResponseChoiceDelta{Role: schemas.Ptr("assistant"), Reasoning: schemas.Ptr("")}, nil),

Based on learnings, in maximhq/bifrost prefer using bifrost.Ptr()/schemas.Ptr() to create pointers instead of ad-hoc address-of helpers, for consistency across the codebase.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@core/providers/anthropic/reasoningstream_test.go` at line 39, Remove the
local p pointer helper in the reasoning stream tests and replace its usages with
the existing schemas.Ptr helper, preserving the current pointer values and test
behavior.

Source: Learnings

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@core/providers/anthropic/reasoningstream_test.go`:
- Line 39: Remove the local p pointer helper in the reasoning stream tests and
replace its usages with the existing schemas.Ptr helper, preserving the current
pointer values and test behavior.

In `@core/schemas/mux.go`:
- Around line 2225-2245: Set the completed status on reasoning output items in
both closeReasoningItem and the terminal reasoning-message construction shown
here. Reuse the existing statusFinal symbol and ensure each reasoning doneItem
or ResponsesMessage includes Status alongside its type, role, content, and
optional ID, matching text and tool-call completion behavior.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: abd99a0a-1045-44df-b459-d26b34f471a9

📥 Commits

Reviewing files that changed from the base of the PR and between e7545b2 and 952c020.

📒 Files selected for processing (4)

• core/providers/anthropic/reasoningstream_test.go
• core/providers/anthropic/responses.go
• core/schemas/mux.go
• core/schemas/mux_test.go

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@core/providers/anthropic/reasoningstream_test.go`:
- Around line 147-164: Update the stream test around the resumed reasoning chunk
in the chunks loop to capture the events emitted specifically for that chunk
after text begins, and assert that it contains no thinking content-block start
or delta events. Keep the existing open-block validation, while ensuring both a
newly opened thinking block and a completely missing expected event sequence are
detected.
- Around line 127-129: Update the assertion for blocks[0].Signature in the
reasoning stream test to verify both that the pointer is non-nil and that its
dereferenced value is the empty string when the fixture has no signature.
Preserve the existing failure context while ensuring synthesized non-empty
values fail.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 4b8e17e9-2672-4fb8-b3de-a222fbe5cad4

📥 Commits

Reviewing files that changed from the base of the PR and between 952c020 and 36391e1.

📒 Files selected for processing (4)

• core/providers/anthropic/reasoningstream_test.go
• core/providers/anthropic/responses.go
• core/schemas/mux.go
• core/schemas/mux_test.go

🚧 Files skipped from review as they are similar to previous changes (3)

• core/schemas/mux_test.go
• core/providers/anthropic/responses.go
• core/schemas/mux.go

[Pratham-Mishra04]

Merge activity

• Jul 17, 11:14 AM UTC: A user started a stack merge that includes this pull request via Graphite.
• Jul 17, 11:19 AM UTC: Graphite rebased this pull request as part of a merge.
• Jul 17, 11:21 AM UTC: Graphite rebased this pull request as part of a merge.
• Jul 17, 11:23 AM UTC: Graphite rebased this pull request as part of a merge.
• Jul 17, 11:23 AM UTC: @Pratham-Mishra04 merged this pull request with Graphite.

The base branch was changed.