fix: persist all messages in multi-step tool calls

[seer-by-sentry]

This PR addresses an AI_APICallError where Anthropic rejected API calls due to tool_use IDs without corresponding tool_result blocks immediately after them.

Root Cause:
The onFinish handler in app/(chat)/api/chat/route.ts was incorrectly saving only a single assistant message from the appendResponseMessages output. When streamText used tools with maxSteps > 1, it generated multiple messages (e.g., assistant[tool_use], user[tool_result], assistant[text]). However, only the first assistant message (often the tool_use block) was persisted to the database, leading to a corrupted conversation history where tool_use blocks were not followed by tool_result blocks.

Solution:
The onFinish handler has been updated to correctly save all relevant messages generated by appendResponseMessages (excluding the initial user message). This ensures that tool_use and tool_result pairs, along with the final assistant response, are fully persisted in the database, maintaining a valid conversation history for the LLM API.

Fixes SUPERMEMORY-BACKEND-ZW3

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
infinite-chat	Error		Jun 3, 2026 11:47pm