tool_use/tool_result pairing breaks causing 400 error with Claude Code / OpenClaw

[yinshanlake]

Bug Description

When using Agent Maestro as a proxy for Claude Code (or OpenClaw), the Anthropic API returns a 400 error about mismatched tool_use_id in tool_result blocks after extended tool-using conversations. This makes Agent Maestro unusable for any non-trivial agent workflow.

Error message:

messages.0.content.0: unexpected `tool_use_id` found in `tool_result` blocks: <id>. 
Each tool_result block must have a corresponding tool_use block in the previous message.

Root Cause Analysis

The Problem

The Anthropic Messages API enforces a strict invariant: every tool_result content block in a user message must reference a tool_use_id from a tool_use block in the immediately preceding assistant message. This is a hard constraint — violating it produces a 400 error.

When Agent Maestro converts Anthropic messages → VS Code Language Model API format → sends to the model, the VS Code LM API (Copilot backend) internally processes and may truncate/compact the conversation history. However, Agent Maestro does NOT actually re-validate the original Anthropic message array before forwarding it. The error actually comes from the VS Code LM API's internal handling — it passes the messages to its own backend which apparently enforces similar constraints.

Why the Current Workaround Is Insufficient

In anthropicRoutes.ts (lines 481-486), there's an existing workaround:

const isContextWindowExceeded =
  errorMessage.includes("unexpected `tool_use_id` found in `tool_result` blocks") &&
  maxInputTokens > 0 &&
  inputTokens > maxInputTokens;

This workaround only triggers when inputTokens > maxInputTokens, meaning it only handles the case where the context window is exceeded. However, the pairing error frequently occurs within the context window limit — the VS Code LM API backend can reject tool_use/tool_result pairings for various reasons unrelated to context length.

When the workaround DOES trigger, it returns a fake model_context_window_exceeded stop reason with inflated token counts to force the client (Claude Code/OpenClaw) to auto-compact. This is a reasonable strategy, but since it rarely triggers, most errors just bubble up as 500 internal server errors.

Message Flow

OpenClaw/Claude Code → Agent Maestro (localhost:23333/api/anthropic/v1/messages)
    → convertAnthropicMessagesToVSCode() [src/server/utils/anthropic.ts]
    → client.sendRequest() [VS Code Language Model API]
    → VS Code LM API backend (Copilot)
    → Error: "unexpected tool_use_id found in tool_result blocks"
    → Caught in catch block of anthropicRoutes.ts
    → Workaround condition NOT met (inputTokens <= maxInputTokens)
    → Returns 500 error to OpenClaw/Claude Code
    → OpenClaw completely broken

Proposed Fix

Add a message sanitization step before sending to the VS Code LM API that validates and repairs tool_use/tool_result pairing:

1. Before calling client.sendRequest(), scan the Anthropic messages array
2. For each user message containing tool_result blocks, verify that the immediately preceding assistant message contains matching tool_use blocks with the same IDs
3. For orphaned tool_result blocks (no matching tool_use in the previous assistant message):
   • Either inject a synthetic tool_use block into the preceding assistant message
   • Or remove the orphaned tool_result block and convert its content to a plain text block
4. For orphaned tool_use blocks in assistant messages (no matching tool_result in the next user message):
   • Either inject a synthetic tool_result block
   • Or remove the orphaned tool_use block

Additionally, the catch block workaround should be broadened: when the "unexpected tool_use_id" error occurs, it should ALWAYS return the model_context_window_exceeded fake response (not just when inputTokens > maxInputTokens), so the client always gets a chance to compact and retry.

Reproduction Steps

1. Install Agent Maestro v2.8.1 in VS Code
2. Configure OpenClaw or Claude Code to use http://localhost:23333/api/anthropic as the API base URL
3. Start a conversation that uses tools (e.g., file reading, bash commands)
4. After several rounds of tool use (typically 5-10+ tool calls), the error appears
5. Once the error occurs, the session becomes completely unusable — every subsequent request also fails

Environment

• Agent Maestro: v2.8.1
• VS Code: 1.100+
• Client: OpenClaw 2026.2.24 / Claude Code
• Model: claude-opus-4-20250514 (via VS Code Copilot)

Impact

This bug makes Agent Maestro completely unusable for any agent workflow that involves tool use (which is the primary use case for Claude Code / OpenClaw). After a few tool calls, the session breaks and cannot recover without restarting the entire conversation.