Skip to content

[BUG] Anthropic Haiku 4.5 fails to execute function calls #8702

@cmpdev23

Description

@cmpdev23

Problem (one or two sentences)

When using the new Anthropic Haiku 4.5 model with Roo Code, the model fails to execute codebase_search and read_file function calls properly. The exact same prompt works correctly with Claude Sonnet 4. The behavior seems to break during the multi-step analysis sequence.

Context (who is affected and when)

This issue affects users who use Anthropic Haiku 4.5 with Roo Code.
When trying to run a standard multi-step analysis (e.g. exploring a code module and integrating components), the model is unable to properly execute function calls. This happens systematically when using Haiku 4.5, while the exact same workflow works with Claude Sonnet 4.

Reproduction steps

1 Select Anthropic Haiku 4.5 as the model in Roo Code.

2 Ask the model to analyze a code module and create an integration plan (e.g. “Analyze the structure of the Analyses module and create a plan for integrating FormSummary with SyncContext”).

3 Observe the model’s generated function calls in Roo.

4 Roo displays an error after the first function calls, and the task does not proceed further.

Expected result

The model should generate valid JSON for codebase_search and read_file function calls, and Roo should execute these without errors — just like it does with Claude Sonnet 4.

Actual result

The model generates partial or malformed function call sequences. Roo fails to execute them and shows an error message

Variations tried (optional)

I tried explaining to the model manually how to call the functions properly within the conversation. This made the situation worse — the model started producing even more broken sequences, failing earlier in the process.

App Version

3.28.17

API Provider (optional)

None

Model Used (optional)

Haiku 4.5

Roo Code Task Links (optional)

https://app.roocode.com/share/8c20f097-f22c-4d66-93bb-e6ea6587cdfd

Image

Relevant logs or errors (optional)

Roo is having trouble… This may indicate a failure in the model's thought process or inability to use a tool properly, which can be mitigated with some user guidance (e.g., ‘Try breaking down the task into smaller steps’).

Metadata

Metadata

Assignees

No one assigned

    Labels

    Issue/PR - TriageNew issue. Needs quick review to confirm validity and assign labels.bugSomething isn't working

    Type

    No type

    Projects

    Status

    Triage

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions