fix: Exclude computer tool from tool_use modality

[dzorlu]

Summary

• Fixed bug where tool_use tasks were incorrectly exposed to the computer tool
• Added inverse filtering for tool_use modality to exclude computer tool
• Updated and added tests to verify correct behavior

Problem

The tool filtering logic in task_env.py only handled computer_use modality (keeping ONLY the computer tool). For tool_use modality, there was no filtering - all tools passed through, including the computer tool if the MCP endpoint exposed it.

This caused training runs to log call_tool(computer) for what should be pure tool-use tasks.

Changes

• src/envs/fleet_env/task_env.py: Add filtering to exclude computer tool for tool_use modality
• tests/envs/test_fleet_task_env.py: Update test name and assertions, add new test for function format

Test plan

• Run pytest tests/envs/test_fleet_task_env.py::TestFleetTaskEnvComputerUseFiltering -v - all 5 tests pass

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

Fixes tool filtering bug where tool_use modality tasks incorrectly received the computer tool alongside API tools. The fix adds inverse filtering logic to explicitly exclude the computer tool for tool_use tasks (lines 222-233 in task_env.py), complementing the existing computer_use filtering that keeps only the computer tool. The implementation correctly handles both tool formats (direct name field and nested function.name field). Tests were updated to verify the exclusion behavior for both formats.

Confidence Score: 5/5

• This PR is safe to merge with no issues identified
• The fix is focused, well-tested, and addresses a clear bug. The logic mirrors the existing computer_use filtering pattern, handles both tool formats correctly, and includes comprehensive test coverage for the new behavior. No mechanical issues, alignment concerns, or edge cases identified.
• No files require special attention

Important Files Changed

Filename	Overview
src/envs/fleet_env/task_env.py	Added inverse filtering logic to exclude computer tool for tool_use modality, preventing incorrect tool exposure during training
tests/envs/test_fleet_task_env.py	Updated test assertions and added new test case to verify computer tool exclusion for both name and function formats

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[reset_async called] --> B{Fetch tools if not fetched}
    B -->|Tools fetched| C{Check modality}
    C -->|tool_use| D[Filter: EXCLUDE 'computer' tool]
    C -->|computer_use| E[Filter: KEEP ONLY 'computer' tool]
    D --> F{Check tool format}
    E --> G{Check tool format}
    F -->|name format| H[Remove if t.name == 'computer']
    F -->|function format| I[Remove if t.function.name == 'computer']
    G -->|name format| J[Keep only if t.name == 'computer']
    G -->|function format| K[Keep only if t.function.name == 'computer']
    H --> L[Return filtered tools cache]
    I --> L
    J --> L
    K --> L
    E -->|No computer tool found| M[Warning: Config error - Clear tools cache]
    M --> L

Loading

_{Last reviewed commit: 26f3940}