Summary

This PR refactors the reflection agent system to use structured tool calls instead of text pattern matching. The reflection agent now uses a dedicated reflect tool to report its analysis decision, making the system more robust and type-safe.

Key Changes

Reflection Configuration & Agent Definition

• Added reflection_agent field to ReflectionConfig to allow custom reflection agents (defaults to built-in)
• Replaced enable_reflection: bool with reflection: Option<ReflectionConfig> in StandardDefinition for more flexible configuration
• Added validate_reflection_agent() method to ensure reflection agents have the "reflect" tool configured
• Updated helper methods: is_reflection_enabled() and new reflection_config() getter

Reflect Tool Implementation

• Implemented new ReflectTool as a built-in tool with structured parameters:
  • quality: enum (excellent/good/fair/poor)
  • completeness: enum (complete/partial/incomplete)
  • should_continue: boolean (retry decision)
  • reason: optional explanation
• Tool stores its structured result as the final result in ExecutorContext for downstream processing
• Added to builtin tools list in get_builtin_tools()

Reflection Agent Loop

• Updated run_reflection_agent() to return ReflectionResult containing both text content and structured tool call result
• Modified reflection decision logic to extract should_continue from the tool call output instead of text pattern matching ("Should Continue: YES/NO")
• Removed reliance on string matching in reflection agent output

Reflection Agent Markdown

• Updated reflection_agent.md to:
  • Configure the "reflect" tool in its tools.builtin
  • Instruct agent to use the reflect tool instead of text output
  • Update decision rules to match tool parameter values

Execution Result Formatting

• Added truncation logic to ExecutionResult::as_observation() to prevent token bloat:
  • Text parts: max 1000 chars
  • Data/ToolResult parts: max 500 chars
  • Includes truncation indicator with total character count

Cleanup & Removals

• Removed unused BrowserHooksConfig enum and related field from StandardDefinition
• Removed browser_hooks configuration option

Test Improvements

• Added test_store_config() helper to create in-memory SQLite databases for tests
• Updated tests to use in-memory stores instead of filesystem dependencies
• Added API key environment variable checks to skip tests when credentials unavailable
• Fixed test isolation issues in prompt store and auth provider tests

Implementation Details

• The reflection agent now signals its decision through a tool call rather than text parsing, improving reliability
• Custom reflection agents can be specified but must explicitly include the "reflect" tool
• The built-in reflection agent automatically gets the reflect tool
• Execution results are now truncated in observations to prevent context window bloat while preserving full data in storage

https://claude.ai/code/session_019JhhSoRA6gQZNvpWsGWxNN