Codebase Learning Extractor

Overview

This agent is designed to act as a specialized knowledge curator for software development projects. Its primary function is to review raw session transcripts (e.g., debugging sessions, complex feature implementations) and distill them into concise, non-obvious learnings.

The goal is not to summarize what was done, but to capture how it was discovered—the hidden relationships, tricky workarounds, or undocumented constraints that are crucial for future developers working on the codebase.

Capabilities

• Insight Extraction: Identifies non-obvious technical discoveries, such as misleading error messages leading to breakthroughs or unexpected file dependencies.
• Contextual Placement: Automatically scopes learnings to the most relevant directory level (e.g., root AGENTS.md for project-wide issues, or a specific module's AGENTS.md).
• Redundancy Filtering: Skips obvious facts, standard framework behavior, and information already present in existing learning files.
• Structured Output: Updates or creates dedicated AGENTS.md files with highly condensed entries (1-3 lines per insight).

Example Use Cases

• Debugging Breakthroughs: If a session reveals that an API endpoint only works correctly when called from a specific parent module, this agent captures that dependency.
• Build System Quirks: Capturing non-standard build flags or required environment variables not documented in the main README.
• Architectural Constraints: Documenting instances where two seemingly unrelated modules must be modified together to prevent runtime errors.

By systematically updating these decentralized AGENTS.md files, it builds a living, actionable map of the codebase's true operational knowledge.