Prevent agents from calling duplicate tools with same inputs

[Copilot]

This PR implements duplicate tool call prevention to address a common failure mode where frontier models (GPT-5, Gemini-Flash-2.5) get stuck calling the same tool repeatedly with identical inputs.

Problem

Agents would sometimes enter infinite loops calling tools like:

<tool name="lookup_memory">{"query": "project timeline"}</tool>
<tool name="lookup_memory">{"query": "project timeline"}</tool>
<tool name="lookup_memory">{"query": "project timeline"}</tool>
<!-- ... continues indefinitely -->

Solution

The implementation tracks tool calls within each AgentRun execution and blocks duplicates based on tool name + normalized arguments. However, agents can override this by providing unique reasoning:

<!-- First call - allowed -->
<tool name="lookup_memory">{"query": "project timeline"}</tool>

<!-- Exact duplicate - blocked -->
<tool name="lookup_memory">{"query": "project timeline"}</tool>

<!-- Duplicate with unique reasoning - allowed -->
<tool name="lookup_memory">{"query": "project timeline", "_duplicate_reasoning": "searching with broader context"}</tool>

Key Features

• Argument normalization: Excludes _nonce and _duplicate_reasoning from duplicate detection
• Reasoning validation: Ensures _duplicate_reasoning is unique within the current session
• Format support: Works with both JSON and XML argument formats
• Clear error messages: Guides agents when duplicates are blocked
• Whitespace handling: Normalizes reasoning to prevent trivial bypasses
• Stress level adjustment: Increases when duplicates are blocked to encourage different behavior

Components Affected

• AgentRun: Primary implementation with tool call history tracking
• AgentLearn: Extended to support duplicate prevention
• Both components now initialize self.tool_call_history = [] and reset it on each execution

Technical Implementation

Added helper functions:

• normalize_tool_args(): Removes special parameters for comparison
• extract_nonce_and_reasoning(): Extracts special parameters from parsed args
• is_duplicate_tool_call(): Checks if tool call matches previous calls
• is_reasoning_unique(): Validates reasoning hasn't been used before

The duplicate check is integrated into handle_tool_use() before tool execution, providing immediate feedback when duplicates are detected.

Fixes #8.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[Copilot]

@wmeddie 👋 This repository doesn't have Copilot instructions. With Copilot instructions, I can understand the repository better, work faster and produce higher quality PRs.

I can generate a .github/copilot-instructions.md file for you automatically. Click here to open a pre-filled issue and assign it to me. I'll write the instructions, and then tag you for review.

[wmeddie]

Thanks for the implementation. I found a couple of correctness issues that should be addressed before this is merged:

1. AgentLearn still subclasses Component, but its execute() now calls self.handle_tool_use(...). handle_tool_use is only defined on AgentRun, so any learning run that emits a tool call will raise AttributeError: 'AgentLearn' object has no attribute 'handle_tool_use' instead of executing/preventing the tool call. Either make AgentLearn inherit/reuse the AgentRun implementation or add the duplicate-prevention path to AgentLearn directly.

2. Plain-text tool arguments all normalize to None, so different raw arguments for the same tool are treated as duplicates. For example, <tool name="lookup_memory">project timeline</tool> followed by <tool name="lookup_memory">project milestones</tool> is blocked even though the arguments differ. parse_tool_args() returns (None, raw_string) for non-JSON/non-XML args, but duplicate detection only stores/compares the parsed dict. The normalized key should include the raw argument string when parsing does not produce a dict.

I also noticed the new _duplicate_reasoning / _nonce metadata is left in tool_args_str when the tool is executed. That may be intentional for _nonce per the doc, but _duplicate_reasoning will be passed to standard tools/MCP tools and could break strict schemas; consider stripping framework-only metadata before dispatch if tools are not expected to receive it.

Validation run locally:

• python -m py_compile agent_components.py passed.
• python -m pytest -q found no tests.