Overview

As the Gemini CLI grows in complexity, debugging agent behavior and reproducing issues has become increasingly difficult. This epic tracks the development of enhanced diagnostic tooling and a dedicated "devtools" suite to streamline development and troubleshooting.

Current Difficulties

• Non-deterministic Behavior: The stochastic nature of LLMs makes it hard to reproduce specific failure modes reliably.
• State Visibility: Limited insight into the internal state transitions of the agent, tool execution details, and the full conversation history during a session.
• Environment Context: Capturing the exact environment (environment variables, local file system state, project configuration) at the moment of failure is manual and error-prone.
• Reproduction Lag: User-reported issues are often hard to replicate without the exact sequence of events and model responses.

Proposed Solution: DevTools Suite

We propose implementing a diagnostic infrastructure that includes:

• Session Recording & Playback: Ability to export a full session log (including hidden tool outputs and metadata) that can be replayed to reproduce issues.
• Diagnostic Dashboard: A hidden or developer-facing UI (e.g., via a specific flag or command) to inspect the message chain, tool calls, and token usage in real-time.
• Enhanced Logging: Structured logging of tool interactions and decision-making processes.
• Environment Snapshotting: Automatic gathering of non-sensitive system info and configuration when an error occurs.