Fix hierarchical delegation with language-agnostic tool matching

[devin-ai-integration]

Fix hierarchical delegation with language-agnostic tool matching (Issue #3925)

Summary

Fixes issue #3925 where the hierarchical process behaves differently depending on the language used. When using non-English prompts, the manager agent would repeatedly delegate to itself instead of using specialized coworkers because tool names were hardcoded in English.

Root Cause: Tool name matching relied on exact English strings ("Delegate work to coworker", "Ask question to coworker"), which failed when LLMs generated tool calls in other languages or used variations of the names.

Solution: Added language-agnostic tool matching using stable short identifiers:

• delegate_work / delegate_work_to_coworker → "Delegate work to coworker"
• ask_question / ask_question_to_coworker → "Ask question to coworker"

Changes:

1. Tool Selection (tool_usage.py): Added _normalize_tool_name(), _get_tool_aliases(), and enhanced _select_tool() to support:

   • Exact name matching (case-insensitive) - backward compatible
   • Normalized/slugified name matching (spaces/hyphens → underscores)
   • Stable short identifier matching
   • Fuzzy matching fallback (0.85 threshold) - existing behavior

2. Delegation Detection (tool_usage.py): Updated delegation counting logic to recognize short identifiers in addition to English names

3. Memory Filter (base_agent_executor_mixin.py): Updated short-term memory filter to exclude delegation actions using both English names and short identifiers

4. Tests (test_tool_usage.py): Added 7 unit tests covering normalization, alias generation, tool selection with various inputs, and memory filtering

Review & Testing Checklist for Human

• CRITICAL: Test with actual non-English prompts - The issue was reported with Japanese, but I haven't tested with real non-English prompts. Please verify that hierarchical crews now work correctly when the LLM generates tool calls in Japanese, Spanish, or other languages.
• Verify backward compatibility - Test existing English-based hierarchical crews to ensure they work exactly as before. The changes should be transparent to existing users.
• Check delegation detection patterns - Review the patterns in _create_short_term_memory() to ensure they're not too broad (could exclude legitimate actions from memory) or too narrow (could include delegation actions in memory).
• Edge case testing - Test with unusual tool name variations (mixed case, extra spaces, special characters) to verify the normalization logic handles them correctly.

Recommended Test Plan

1. Create a simple hierarchical crew with a manager and 2-3 specialized agents
2. Test with English prompts - should work as before
3. Test with Japanese prompts (or use an LLM configured to respond in Japanese) - manager should now properly delegate to coworkers instead of delegating to itself
4. Check the short-term memory to verify delegation actions are not being saved

Notes

• All CI checks passing (lint, type-checker, tests across Python 3.10-3.13)
• Maintained backward compatibility by keeping exact name matching and fuzzy matching as fallback
• The fix uses stable identifiers rather than full i18n support, which would require translating all tool names and invalidating VCR cassettes

Link to Devin run: https://app.devin.ai/sessions/8edde7b0966342a09a28a2c9bb976301
Requested by: João (joao@crewai.com)

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring