unexpected `tool_use_id` found in `tool_result` blocks

[aheritier]

Describe the bug

Since few hours my Anthropic models are rejecting my requests with:

all models failed: error receiving from stream: POST "https://api.anthropic.com/v1/messages": 400 Bad Request (Request-ID: req_011CXwUGAoEeBVooyRv5365i)
{"type":"error","error":{"type":"invalid_request_error","message":"messages.298.content.1: unexpected `tool_use_id` found in `tool_result` blocks:
toolu_01VaSNuJZRTJH227epge8S24. Each `tool_result` block must have a corresponding `tool_use` block in the previous message."},"request_id":"req_011CXwUGAoEeBVooyRv5365i"}

Version affected

cagent version v1.20.6
Commit: Homebrew

How To Reproduce

Not sure. My agents definition didn't change when the problem started to occur. I can share a team definition if useful.

Expectation

No error, or a better management of the error, right now cagent is not working anymore.

Screenshots

N/A

OS and Terminal type

macOS + Ghostty

Additional context

N/A

[rumpl]

Thanks!

What would be useful for us is to get the session in which you got this error. Or at least screenshots of what the agent was doing before you got the error

[aheritier]

Hi @rumpl

My session ID is 3f8f21c7-07b5-4b72-bc73-242be25d13fb

I restored it

[aheritier]

Claude is considering that the context got corrupted:

  Issue 1: API tool_use_id Mismatch (Red Error)

  all models failed: error receiving from stream: POST "https://api.anthropic.com/v1/messages": 400 Bad Request
  unexpected 'tool_use_id' found in 'tool_result' blocks: toolu_01VaSNuJZRTJH227epge8S24.
  Each 'tool_result' block must have a corresponding 'tool_use' block in the previous message.

  This is a conversation state corruption error. It means cagent is sending a tool_result message back to the API that references a tool_use_id which no longer exists in the conversation history. This typically happens when:

  1. Context window compression truncated earlier messages that contained the original tool_use call, but cagent still tries to send its result
  2. A retry/resume after a failure caused the conversation to get out of sync — the agent loops "Let's resume" / "Let's fix the compilation issue" repeatedly, suggesting it kept hitting errors and retrying

  The fact that the same toolu_01VaSNuJZRTJH227epge8S24 appears in all three red errors confirms it's stuck in a loop trying to send the same orphaned tool result.

It probably makes sense as new sessions are ok.