fix: prevent tool call loss from late-arriving names and duplicate finish chunks by drewd789 · Pull Request #2404 · QwenLM/qwen-code

drewd789 · 2026-03-15T21:34:21Z

Summary

Fixes three complementary bugs that cause tool calls to disappear during streaming:

Parser bug: Function name arriving after JSON completion causes index reassignment
Converter bug: Waiting until finish_reason to emit allows parser state changes to lose tool calls
Pipeline bug: Duplicate finish_reason chunks cause second chunk to overwrite first chunk's tool calls

Root Causes

Bug 1: Parser Index Reassignment (`streamingToolCallParser.ts`)

findMostRecentIncompleteIndex() and findNextAvailableIndex() considered tool calls "complete" based on JSON state alone (depth=0, non-empty buffer, has ID), without checking if meta.name exists. When name arrived in a later chunk, parser reassigned it to a new index, leaving JSON at the old index → getCompletedToolCalls() returned empty.

Bug 2: Late Emission (`converter.ts`)

Tool calls were only emitted at finish_reason, not during streaming. This allowed parser state changes between JSON completion and name arrival to cause tool call loss.

Bug 3: Finish Chunk Overwriting (`pipeline.ts`)

handleChunkMerging() unconditionally stored every finish chunk. Some providers send finish_reason twice (first with tool calls, second with usage metadata only). Second chunk overwrote the first → tool calls lost.

Changes

`streamingToolCallParser.ts`

Check meta.name when determining if tool call is complete in findMostRecentIncompleteIndex()
Check meta.name when determining if index is available in findNextAvailableIndex()
Prevents index reassignment when name arrives late

`converter.ts`

Emit tool calls immediately during streaming when result.complete && result.value && meta.name
Track emitted tool call IDs in emittedToolCallIds Set
Skip already-emitted IDs at finish_reason to prevent duplicates
Clear emittedToolCallIds in resetStreamingToolCalls()

`pipeline.ts`

Skip finish chunks without tool calls when pending finish already has tool calls
Preserves first finish chunk's tool calls from being overwritten

Testing

Unit tests added for all three fixes:

Parser tests:

should keep name at correct index when it arrives after JSON
should not reassign index when name arrives after JSON
should handle multiple tool calls with late names

Converter tests:

should emit tool call immediately when JSON completes during streaming
should not emit duplicate tool calls at finish_reason
should handle multiple finish_reason chunks without duplicates

Pipeline tests:

should skip finish chunk without tool calls when pending finish has tool calls

All tests pass, confirming the fixes work correctly.

Tests for all three tool call loss fixes: **streamingToolCallParser.ts:** - should keep name at correct index when it arrives after JSON - should not reassign index when name arrives after JSON - should handle multiple tool calls with late names **converter.ts:** - should emit tool call immediately when JSON completes during streaming - should not emit duplicate tool calls at finish_reason - should handle multiple finish_reason chunks without duplicates **pipeline.ts:** - should skip finish chunk without tool calls when pending finish has tool calls These tests verify the fixes for: 1. Parser index reassignment when name arrives late 2. Late emission allowing parser state changes to lose tool calls 3. Finish chunk overwriting when API sends finish_reason twice

…nish chunks Fixes three complementary bugs that cause tool calls to disappear during streaming: **streamingToolCallParser.ts:** - Check meta.name when determining if tool call is complete - Prevents index reassignment when name arrives after JSON completion **converter.ts:** - Emit tool calls immediately during streaming when complete + has name - Track emitted tool call IDs in emittedToolCallIds Set - Skip already-emitted IDs at finish_reason to prevent duplicates - Clear emittedToolCallIds in resetStreamingToolCalls() **pipeline.ts:** - Skip finish chunks without tool calls when pending finish already has them - Prevents second finish chunk from overwriting first chunk's tool calls Root causes: 1. Parser considered tool calls "complete" without checking for meta.name 2. Converter only emitted at finish_reason, allowing parser state changes to lose tool calls 3. Pipeline unconditionally stored every finish chunk, overwriting previous ones

wenshao · 2026-04-24T03:51:14Z

    openaiResponse: OpenAI.Chat.ChatCompletion,
  ): GenerateContentResponse {
-    const choice = openaiResponse.choices?.[0];
+    const choice = openaiResponse.choices[0];


[Critical] convertOpenAIResponseToGemini() now unconditionally dereferences openaiResponse.choices[0]. The previous code handled an empty/missing first choice and returned an empty Gemini response, so an OpenAI-compatible response with choices: [] will now throw before usage metadata can be converted.

Suggested change

const choice = openaiResponse.choices[0];

const choice = openaiResponse.choices?.[0];

const response = new GenerateContentResponse();

if (!choice) {

response.candidates = [];

response.responseId = openaiResponse.id;

response.createTime = openaiResponse.created

? openaiResponse.created.toString()

: new Date().getTime().toString();

response.modelVersion = this.model;

response.promptFeedback = { safetyRatings: [] };

return response;

}

— gpt-5.5 via Qwen Code /review

wenshao · 2026-04-24T03:51:14Z

+        return false;
      }
+      collectedGeminiResponses.push(response);
+      setPendingFinish(response);


[Critical] This branch preserves the earlier finish chunk with tool calls, but it also drops response.usageMetadata from the later duplicate finish chunk. That breaks providers that send tool calls on the first finish chunk and token usage on a later finish/usage chunk.

Suggested change

setPendingFinish(response);

if (hasPendingFinish && pendingToolCallCount > 0) {

if (response.usageMetadata) {

lastCollected.usageMetadata = response.usageMetadata;

setPendingFinish(lastCollected);

}

return false;

}

— gpt-5.5 via Qwen Code /review

wenshao · 2026-04-24T03:51:14Z

+            if (meta.name) {
+              const id =
+                meta.id ||
+                `call_${Date.now()}_${Math.random().toString(36).substring(2, 9)}`;


[Critical] addChunk() can remap the provider-supplied index to a different internal parser index, but early emission still reads metadata with the original index. If a provider reuses index 0, this can pair the new arguments with stale metadata, skip the completed call, or emit it again at finish. Please have addChunk() return the resolved internal index or metadata, then use that same key for early emission and duplicate tracking.

— gpt-5.5 via Qwen Code /review

wenshao · 2026-04-24T03:51:14Z

+          // Skip if already emitted during streaming (prevents duplicates)
+          if (toolCall.id && this.emittedToolCallIds.has(toolCall.id)) {
+            continue;
+          }


[Critical] Duplicate suppression only tracks provider IDs, but tool-call IDs are optional in this path. For an id-less call emitted early, no stable emitted key is recorded; when finish_reason arrives, getCompletedToolCalls() can emit the same parser entry again with a different synthetic ID. Track emitted parser/internal indices in addition to provider IDs, or expose a stable parser key for each completed call.

— gpt-5.5 via Qwen Code /review

wenshao · 2026-04-24T03:51:14Z

-      if (!buffer.trim() || depth > 0 || !meta?.id) {
+      // Tool call is incomplete if: JSON open, buffer empty, missing ID, or missing name
+      // Any condition true → index available for reuse
+      if (!buffer.trim() || depth > 0 || !meta?.id || !meta.name) {


[Critical] Treating a slot with complete JSON and an ID as reusable just because the function name has not arrived yet can corrupt multi-tool streams. If call_1 sends complete arguments before its name, then call_2 starts on the same provider index, findNextAvailableIndex() can return the occupied slot and append/overwrite the first call’s state. Keep such slots occupied for collision allocation; late-name routing should be handled separately.

Suggested change

if (!buffer.trim() || depth > 0 || !meta?.id || !meta.name) {

if (!buffer.trim() || depth > 0 || !meta?.id) {

return this.nextAvailableIndex;

}

— gpt-5.5 via Qwen Code /review

drewd789 requested review from DennisYu07, DragonnZhang, LaZzyMan, Mingholy, gwinthis, pomelo-nwu and tanzhenxin as code owners March 15, 2026 21:34

drewd789 force-pushed the fix/tool-call-name-after-json branch from e7bf0c4 to 55dd9ac Compare March 15, 2026 23:59

github-actions Bot mentioned this pull request Mar 16, 2026

📊 AI CLI 工具社区动态日报 2026-03-16 gsscsd/big_model_radar#43

Open

drewd789 closed this Mar 16, 2026

drewd789 reopened this Mar 16, 2026

Mingholy added the scope/content-generation AI content generation label Mar 16, 2026

Mingholy mentioned this pull request Mar 16, 2026

fix(pipeline): handle duplicate finish_reason chunks from OpenRouter #2403

Merged

Mingholy self-assigned this Mar 16, 2026

github-actions Bot mentioned this pull request Mar 17, 2026

📊 AI CLI 工具社区动态日报 2026-03-17 gsscsd/big_model_radar#50

Open

tanzhenxin added the status/need-information More information is needed to resolve this issue. label Mar 18, 2026

drewd789 added 2 commits April 17, 2026 20:15

drewd789 force-pushed the fix/tool-call-name-after-json branch from 55dd9ac to 59048af Compare April 18, 2026 03:30

wenshao requested changes Apr 24, 2026

View reviewed changes

BingqingLyu mentioned this pull request Apr 27, 2026

fix: prevent tool call loss from late-arriving names and duplicate finish chunks BingqingLyu/qwen-code#51

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: prevent tool call loss from late-arriving names and duplicate finish chunks#2404

fix: prevent tool call loss from late-arriving names and duplicate finish chunks#2404
drewd789 wants to merge 2 commits into
QwenLM:mainfrom
drewd789:fix/tool-call-name-after-json

drewd789 commented Mar 15, 2026

Uh oh!

Mingholy commented Mar 16, 2026

Uh oh!

tanzhenxin commented Mar 18, 2026

Uh oh!

wenshao Apr 24, 2026

Uh oh!

wenshao Apr 24, 2026

Uh oh!

wenshao Apr 24, 2026

Uh oh!

wenshao Apr 24, 2026

Uh oh!

wenshao Apr 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

-    const choice = openaiResponse.choices[0];
+    const choice = openaiResponse.choices?.[0];
+    const response = new GenerateContentResponse();
+    if (!choice) {
+      response.candidates = [];
+      response.responseId = openaiResponse.id;
+      response.createTime = openaiResponse.created
+        ? openaiResponse.created.toString()
+        : new Date().getTime().toString();
+      response.modelVersion = this.model;
+      response.promptFeedback = { safetyRatings: [] };
+      return response;
+    }

-      setPendingFinish(response);
+      if (hasPendingFinish && pendingToolCallCount > 0) {
+        if (response.usageMetadata) {
+          lastCollected.usageMetadata = response.usageMetadata;
+          setPendingFinish(lastCollected);
+        }
+        return false;
+      }

Conversation

drewd789 commented Mar 15, 2026

Summary

Root Causes

Bug 1: Parser Index Reassignment (streamingToolCallParser.ts)

Bug 2: Late Emission (converter.ts)

Bug 3: Finish Chunk Overwriting (pipeline.ts)

Changes

streamingToolCallParser.ts

converter.ts

pipeline.ts

Testing

Related

Uh oh!

Mingholy commented Mar 16, 2026

Uh oh!

tanzhenxin commented Mar 18, 2026

Uh oh!

wenshao Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

wenshao Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

wenshao Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

wenshao Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

wenshao Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Bug 1: Parser Index Reassignment (`streamingToolCallParser.ts`)

Bug 2: Late Emission (`converter.ts`)

Bug 3: Finish Chunk Overwriting (`pipeline.ts`)

`streamingToolCallParser.ts`

`converter.ts`

`pipeline.ts`