https://docs.all-hands.dev/ logo
Join Slack
Powered by
Has anyone using local inference come across the “...
# general
b
Has anyone using local inference come across the “Failed to parse tool call arguments as JSON” issue? The issue is not that the LLM is generating invalid tool calls, but that OH is sending them in the prompt to llama-server: https://github.com/OpenHands/OpenHands/issues/13949 Very nasty, because conversations cannot be continued.
👀 1
r
I have but for different reasons. Having set a limit on token generation to low, the tool calls get truncated and its not parsed correctly.
👀 1
b
I updated the issue, because here the problem is not that the LLM is generating invalid tool calls, but that OH is sending a prompt with invalid tool calls from previous messages and llama-server is rejecting it without doing inference.
c
Can you include the version of the OpenHands you are using 👀
If it is post V1, we can move this issue to the sdk repo
b
I'm using commit c78b92346830e1c83d090de4042f2d85b996bf02: Author: HeyItsChloe <54480367+HeyItsChloe@users.noreply.github.com> Date: Wed Apr 15 222827 2026 -0700 feat(frontend): Remove onboarding feature flags (#13947)
Basically what's happening here is that agent-server is passing invalid tool calls from previous failed events in the prompt and llama-server is trying to parse them and rejecting the request. I believe llama-server shouldn't parse previous messages, or OH should pass those previous invalid tool calls in a way that llama-server doesn't try to parse them.
We are trying now with: https://github.com/OpenHands/software-agent-sdk.git Commit: b3ad71a72685f85c888dbb14f6d58fe0211550d5
Issue updated
👀 1
c
Yeah, looks like it is worth moving the issue to the SDK
👍 1
Open in Slack
PreviousNext
Powered by