fix(chat): keep reused text blocks in order and drop empty reasoning spans

[kwakayama]

Follow-up to #3438. After that landed I probed the rest of the streaming assembly for the same defect class. Two findings, one latent and one live.

handleTextStart had the identical overwrite bug

It replaced its map entry unconditionally, exactly as handleReasoningStart did before #3438. A content id reused across steps blanks the earlier answer and moves the later one below everything that streamed in between:

stream: text(STEP1) → tool → text(STEP2)
before: ["tool-calc", "text:STEP2"]              ← STEP1 destroyed
after:  ["text:STEP1", "tool-calc", "text:STEP2"]

Reasoning ids collide on every run because providers index them (reasoning-0). Text ids usually carry a random suffix (text-BG3xwKvhoiJGrrdZ), which is the only reason this hadn't surfaced. It is reachable today: src/provider/local/model-runtime-adapter.ts:145 builds text-${Date.now()} per doStream call, so two steps completing inside one millisecond reuse the id.

Worth flagging because it would have been hard to spot from the UI: getAnswerPartsForRendering already hides pre-tool text once a post-tool answer exists, so the screen looks correct while the persisted message silently loses content.

Same rule as reasoning now applies. A start on a block that is still open is a replayed start and keeps its text and its order; a start after the block closed opens a new block and retires the old one to closedTextBlocks.

Empty reasoning spans rendered an empty "Thought process"

A reasoning-start / reasoning-end pair with no delta produced {type: "reasoning", text: ""}, and nothing guarded against it — not parts-builder, not groupPartsInOrder, not the Reasoning component, which renders the disclosure trigger regardless of content. Providers open and close the span on any step that did no thinking, so this is common.

This predates #3438 (a single empty span already produced it), but #3438 increased exposure: empty spans sharing an id used to collapse into one overwritten slot and now each is retained.

Completed empty spans are dropped. Two deliberate carve-outs, both covered by tests:

• An open empty span stays, so the thinking affordance still shows while a span streams.
• A redacted or signed span stays even with empty text, because it carries meaning without visible content.

Verification

Probes against the built handler, five steps with both ids fully reused:

L: reason:R1, text:T1, reason:R2, text:T2, reason:R3, text:T3, reason:R4, text:T4, reason:R5, text:T5

Nothing lost, nothing reordered, no duplicates.

Suites: agent + chat + react 1390 passed / 0 failed. deno task typecheck clean, deno fmt --check and deno lint clean.

Not fixed here

Reusing a tool call id across steps destroys the earlier call the same way. Left alone deliberately: correlating a result to a call requires the id to be unique, so a duplicate id is already ambiguous at the protocol level rather than something the client should paper over.

The AG-UI encoder still emits duplicate reasoning messageIds across steps. use-chat is now robust to it, but other consumers of that wire are not. Fixing it there is a wire-contract change and wants its own PR.

Summary by CodeRabbit

• Bug Fixes
  • Improved streaming responses when content identifiers are reused, preserving previously generated text and its ordering.
  • Prevented empty reasoning segments from appearing in responses.
  • Preserved meaningful reasoning content, including redaction metadata, during streaming.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Repo admins can enable using credits for code reviews in their settings.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: e76dbf05-ece0-4066-a64b-54e86f53af40

📥 Commits

Reviewing files that changed from the base of the PR and between 8f223ed and 8715b27.

📒 Files selected for processing (3)

• src/agent/react/use-chat/streaming/handler.test.ts
• src/agent/react/use-chat/streaming/handler.ts
• src/agent/react/use-chat/streaming/parts-builder.ts

---

📝 Walkthrough

Walkthrough

The streaming handler now preserves completed text blocks when providers reuse IDs. The parts builder includes those blocks in order and omits empty completed reasoning spans while retaining meaningful and redacted reasoning data. Tests cover these cases.

Changes

Streaming block preservation

Layer / File(s)	Summary
Text block lifecycle src/agent/react/use-chat/streaming/handler.ts	The handler tracks closed text blocks, preserves duplicate starts, resets them for new messages, and passes them through both streaming paths.
Transcript part assembly src/agent/react/use-chat/streaming/parts-builder.ts	The builder combines closed and active text blocks and omits completed reasoning spans without text, signatures, or redacted data.
Streaming regression coverage src/agent/react/use-chat/streaming/handler.test.ts	Tests cover reused text IDs, restarted text blocks, empty reasoning spans, and redacted reasoning data.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

• veryfront/veryfront-code#3438: Both changes update streaming block preservation and ordering.

Suggested reviewers: kojiwakayama

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes both primary fixes: preserving reused text block order and dropping empty reasoning spans.
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.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/text-block-order-and-empty-reasoning

---

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