ui: fix stop/continue during an agentic loop#23356
Conversation
|
Just want to mention that it is not feasible to resume a partial tool call with the chat completions API. There are too many models/templates to transform partial JSON to their respective tool calling notation. The maintenance cost is not worth it IMO. If this is desired, then the continue logic needs to be re-imagined to one of the following:
The logic here looks fine from a prompt construction view. Worse case, the tool call would need to be regenerated from scratch but could still be seeded with the reasoning/content. |
|
We can improve it a little by seeding the regenerated turn with the reasoning/content already produced before the cut, so a continue lands in the CoT of the last tool call instead of regenerating it from scratch, but we can't do any better than that. |
|
It adds quite a bit of complexity, because seeding the CoT is only useful if the regenerated turn re enters the agentic loop so the tool call actually fires, otherwise the user just sees the tool call reappear without running, so regenerating from scratch stays way simpler. |
Address allozaur review: replace the kind string literals in the ContinueIntent union with a ContinueIntentKind enum (append_text, rerun_turn, next_turn), next to the other agentic enums. Producers, the chat store consumers and the unit test all reference the enum, no magic strings left.
3 participants
Overview
fix: continue button preserves tool_calls and routes through the agentic loop
The Continue button was rebuilding the final assistant message by hand as role plus content plus reasoning_content, dropping tool_calls and attachments from the persisted DatabaseMessage, which broke continue_final_message for any turn carrying tool_calls.
Additional information
A pure classifier in lib/utils/agentic.ts now walks the history around the target and returns append_text, rerun_turn, or next_turn, and continueAssistantMessage dispatches accordingly: classical resume for plain text, branch via regenerateMessageWithBranching when tool_calls have no results yet, or anchor a fresh agentic turn at the last tool result.
10 unit tests cover the classifier, the full vitest suite stays at 224/224, type check and prod build are clean.
Requirements