fix: remove Think action and provide error feedback for invalid actions

[douglaz]

Summary

This PR improves agent robustness by removing the unused Think action and providing explicit error feedback when actions are malformed.

Changes

• Remove unused Think action variant from Action enum
• Replace silent Think fallback with explicit error messages
• Add error feedback when action format is invalid (e.g., missing use_tool: prefix)
• Store errors in agent memory for self-correction
• Add comprehensive tests for invalid action handling

Problem Solved

Previously, when the LLM generated malformed actions (like shell:run_command without the use_tool: prefix), the agent would silently fall back to the Think action which did nothing. This caused the agent to get stuck in confusion loops without understanding what went wrong.

Solution

Now the agent receives clear error messages like:

Invalid action format: 'shell:run_command'. Expected one of: use_tool:<tool>, remember:<key>, explore, wait

These errors are stored in memory as error_<timestamp> entries, allowing the agent to:

1. See what went wrong in its next observation
2. Learn from the error
3. Self-correct by using the proper format

Test Plan

• Unit tests pass
• New test added for invalid action error handling
• Tested with Fedimint challenge - agent receives and learns from errors
• All formatting and linting checks pass