Refine retry logic for parallel tool calling

[DouweM]

Fixes #2311

Refines the retry logic to move retry count management from _call_tool to for_run_step method. This ensures retry counts are scoped per tool name per run step rather than per individual failure.

Changes

• Added failed_calls: set[str] field to track tools that failed in current run step
• Modified _call_tool to add failed tool names to failed_calls instead of incrementing retry count
• Updated for_run_step to build new retry dict with only failed tool names and their updated counts
• Removed success clearing logic from _call_tool since it's now handled in for_run_step

Generated with Claude Code

[DouweM]

@claude Rename failed_calls to failed_tools, updated_retries to retries, and add tests

[claude]

Claude finished @DouweM's task —— View job

---

Todo List

• Read the current code to understand variable names and retry logic
• Rename failed_calls to failed_tools
• Rename updated_retries to retries
• Add tests for the retry logic
• Commit and push changes

✅ Changes completed successfully!

Variable renames in pydantic_ai_slim/pydantic_ai/_tool_manager.py:

• failed_calls → failed_tools (field, initialization, and all references)
• updated_retries → retries (variable and usage)

Added comprehensive tests in tests/test_toolsets.py:

• test_tool_manager_retry_logic() - tests the new retry logic with failed_tools and for_run_step method
• test_tool_manager_multiple_failed_tools() - tests multiple tools failing in same run step
• Tests verify that retry counts are properly scoped per tool name per run step

Note: I wasn't able to run make lint and make typecheck due to permission restrictions. You may want to run these commands locally to ensure code quality standards are met.

The changes have been committed and pushed to the PR branch.

[DouweM]

@claude You're now able to run make install, make test, and make lint

[claude]

Claude finished @DouweM's task —— View job

---

Todo List

• Install dependencies with make install
• Run tests with make test
• Run linting with make lint
• Report results

❌ Unable to run validation commands

I attempted to run make install but encountered missing system dependencies:

/bin/sh: 1: uv: not found
/bin/sh: 1: pre-commit: not found
✖ Error: deno is not installed, but is needed for mcp-run-python

The environment is missing required tools:

• uv (Python package manager)
• pre-commit (Git hook framework)
• deno (JavaScript/TypeScript runtime)

✅ Previous changes completed successfully

The core task was already completed in the previous interaction:

Variable renames in pydantic_ai_slim/pydantic_ai/_tool_manager.py:

• failed_calls → failed_tools (field, initialization, and all references)
• updated_retries → retries (variable and usage)

Added comprehensive tests in tests/test_toolsets.py:

• test_tool_manager_retry_logic() - tests the new retry logic with failed_tools and for_run_step method
• test_tool_manager_multiple_failed_tools() - tests multiple tools failing in same run step

The changes have been committed and pushed to the PR branch. The code validation will need to be run in an environment with the required system dependencies installed.

[github-actions]

Docs Preview

commit:	7325617
Preview URL:	https://1296ac04-pydantic-ai-previews.pydantic.workers.dev