fix(api): normalize chat-completion tool calls into Responses events (#608)#735
Open
chronoai-shining wants to merge 2 commits into
Open
fix(api): normalize chat-completion tool calls into Responses events (#608)#735chronoai-shining wants to merge 2 commits into
chronoai-shining wants to merge 2 commits into
Conversation
…608) After #574, chat-completion providers were routed to /chat/completions with text deltas translated into Responses-API events — but the stream parser dropped delta.tool_calls. The playground tool-use loop only matches on Responses-API response.output_item.done with item.type=function_call, so a tool call from a DeepSeek-style provider never reached the loop. Models would emit execute_in_sandbox(...) and the JSON arrived as plain assistant text instead of triggering the sandbox — runtime-based and mixed skills appeared to respond but never actually ran. Fix in parseChatCompletionStream: - Per-index Map<number, ToolCallAccumulator> carries {id, name, arguments}. Each delta.tool_calls[] chunk merges into its buffer (id+name on the first chunk, arguments JSON string accumulates). - A turn flushes when ANY of: explicit finish_reason (tool_calls / stop / anything non-null), upstream [DONE], or stream EOF. A `flushed` guard makes flush idempotent so multiple end signals never produce duplicate events. - The synthesized event matches the shape chatService already parses via outputItemDoneEventSchema, so zero changes are needed in the playground loop — its existing pendingToolCall capture works for both upstream formats. Parallel tool calls within one turn are supported (one done event per index, emitted in index order). Missing `index` falls back to 0. Streams that close without [DONE] or finish_reason still flush at EOF so buffered calls are never lost. Coverage: 6 new tests in llm.test.ts cover chunked accumulation + finish_reason flush, EOF flush without [DONE], parallel tool calls, idempotent flush across finish_reason+[DONE], intermixed text+tool deltas (event order), and missing index fallback. 17/17 green in the file; full ornn-api suite no new regressions (18 unrelated pre-existing failures in validateSkillFrontmatter #649). Fixes #608
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
parseChatCompletionStreamnow accumulatesdelta.tool_callschunks per index and emits synthesizedresponse.output_item.done/function_callevents on flush. The playground tool-use loop already matches on that shape, so zero changes were needed inchatService.ts.finish_reason(tool_calls/stop/ non-null), upstream[DONE], or stream EOF. Flush is idempotent via aflushedguard.indexfalls back to 0.llm.test.ts, 17/17 in file green.Test plan
bun test src/clients/nyxid/llm.test.ts— 17/17 greenbun run typecheck:api— no new errors inllm.tsornn-apisuite — no new regressionsexecute_in_sandboxshould be called by the loop, not rendered as text. Combined with the session reuse from [Bug] [Playground Sandbox] Installed CLI state is not preserved across tool calls #531, multi-round runtime skills should now work end-to-end on chat-completion providers.Fixes #608