[BUG] Claude requests fail with HTTP 400 when multiple tool calls run concurrently (tool message order mismatch)

[little2b]

Problem (one or two sentences)

When using a Claude model (Anthropic Messages API) in Roo Code, requests intermittently fail with HTTP 400 if the assistant triggers multiple tool calls concurrently. This appears to be caused by tool result messages being appended in completion order rather than the original tool_use order, which violates the API’s expected message/blocks ordering.

Context (who is affected and when)

Who is affected: Anyone using Roo Code with Claude models (Anthropic Messages API) and enabling parallel/concurrent tool execution.

Reproduction steps

Configure Roo Code to use a Claude model via the Anthropic Messages API (any Claude variant is fine).

Ensure tool calling is enabled and the tool runner executes tools concurrently/in parallel (default in many setups).

Use a prompt that reliably triggers multiple tool calls in a single assistant turn, e.g.:

“Search the web for X and Y, then summarize both sources.”

“Run tool A and tool B at the same time, then compare results.”

Run the agent multiple times (or increase likelihood by making one tool slower—e.g., add an artificial delay to one tool implementation).

Observe that when tools finish in a different order than they were requested, Roo Code sends tool_result blocks back out of the original tool_use order.

The Anthropic API then rejects the request with HTTP 400 (conversation invalid / tool results ordering mismatch), and the run fails.

Expected result

Tool results should be sent back to Claude in the same order as the model’s tool_use blocks (matched by tool_use_id), so the request succeeds without HTTP 400.

Actual result

Tool results are sometimes returned out of the original tool_use order during concurrent execution, causing Claude to respond with HTTP 400 and the run to fail.

Variations tried (optional)

No response

App Version

3.50.5 (961d340)

API Provider (optional)

Anthropic

Model Used (optional)

No response

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I've analyzed this issue and here's my proposed implementation plan:

Root cause: When the Anthropic API returns multiple tool_use blocks in a single assistant turn, Roo Code executes them sequentially via presentAssistantMessage, pushing tool_result blocks into userMessageContent as each tool completes. The existing validateAndFixToolResultIds() function (in src/core/task/validateToolResultIds.ts) already handles missing and duplicate tool_results, but it does not reorder them to match the original tool_use order from the assistant message. The Anthropic Messages API requires tool_result blocks to appear in the same order as their corresponding tool_use blocks, so out-of-order results trigger HTTP 400.

Proposed fix: Add a reordering step at the end of validateAndFixToolResultIds() in src/core/task/validateToolResultIds.ts. After all existing validation (deduplication, ID correction, missing result injection), sort the tool_result blocks in finalContent so they appear in the same order as the tool_use blocks in the preceding assistant message, matched by tool_use_id. Non-tool-result blocks (text, image) remain in their current positions relative to the tool results.

Specifically:

1. Build an order map from toolUseBlocks: { tool_use_id -> index }
2. After constructing finalContent, extract tool_result blocks and non-tool-result blocks separately
3. Sort tool_result blocks by their corresponding tool_use index
4. Reconstruct the content array with tool_results first (in correct order), followed by non-tool-result blocks
5. Add tests covering the out-of-order scenario

This is a minimal, targeted change in the existing validation function that already serves as the centralized safety net for tool_use/tool_result issues. No changes to the tool execution flow or streaming logic are needed.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[little2b]

I've analyzed this issue and here's my proposed implementation plan:

Root cause: When the Anthropic API returns multiple tool_use blocks in a single assistant turn, Roo Code executes them sequentially via presentAssistantMessage, pushing tool_result blocks into userMessageContent as each tool completes. The existing validateAndFixToolResultIds() function (in src/core/task/validateToolResultIds.ts) already handles missing and duplicate tool_results, but it does not reorder them to match the original tool_use order from the assistant message. The Anthropic Messages API requires tool_result blocks to appear in the same order as their corresponding tool_use blocks, so out-of-order results trigger HTTP 400.

Proposed fix: Add a reordering step at the end of validateAndFixToolResultIds() in src/core/task/validateToolResultIds.ts. After all existing validation (deduplication, ID correction, missing result injection), sort the tool_result blocks in finalContent so they appear in the same order as the tool_use blocks in the preceding assistant message, matched by tool_use_id. Non-tool-result blocks (text, image) remain in their current positions relative to the tool results.

Specifically:

1. Build an order map from toolUseBlocks: { tool_use_id -> index }
2. After constructing finalContent, extract tool_result blocks and non-tool-result blocks separately
3. Sort tool_result blocks by their corresponding tool_use index
4. Reconstruct the content array with tool_results first (in correct order), followed by non-tool-result blocks
5. Add tests covering the out-of-order scenario

This is a minimal, targeted change in the existing validation function that already serves as the centralized safety net for tool_use/tool_result issues. No changes to the tool execution flow or streaming logic are needed.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

@roomote I approve this plan — reordering tool_result blocks to match the preceding tool_use order (by tool_use_id) should fix the Anthropic 400.

One small note: please keep non-tool blocks (text/image) in their original positions and only reorder the tool_result blocks relative to each other (stable), rather than moving all tool_results to the front. Otherwise, the content structure might change unexpectedly.

Thanks — happy to test once a PR is ready.

[roomote]

I've seen the discussion and the approved plan. I'll implement the tool_result reordering fix in validateAndFixToolResultIds() now -- keeping non-tool blocks in their original positions and only reordering tool_result blocks relative to each other. Will open a PR shortly.

[roomote]

Opened a draft PR with my attempt at addressing this: #11806

The fix adds a reorderToolResults() step at the end of validateAndFixToolResultIds() that sorts tool_result blocks to match the tool_use order, while keeping non-tool-result blocks in their original positions. Feedback welcome!