feat: Interactive breakpoint and step-through debugging (AgentStepper)

[acailic]

Paper Reference

• Title: AgentStepper: Interactive Debugging of Software Development Agents
• Authors: Robert Hutter, Michael Pradel
• Year: 2026
• URL: https://arxiv.org/abs/2602.06593
• Venue: arXiv preprint

Paper Summary

First interactive debugger for LLM-based software engineering agents. Represents agent trajectories as structured conversations among LLM, agent program, and tools. Supports breakpoints, stepwise execution, and live editing of prompts/tool invocations. User study shows 60% bug identification vs 17% with conventional tools, and reduced frustration from 5.4/7.0 to 2.4/7.0.

Proposed Feature

Implement breakpoint and step-through debugging for agent sessions in Peaky Peek:

Core Capabilities

• Breakpoints: Allow users to set breakpoints on specific agent decisions, tool calls, or event types. When replaying a session, execution pauses at each breakpoint.
• Step Controls: Step into (next decision), step over (skip tool internals), step out (return to parent context).
• Live Prompt Editing: At any breakpoint, allow inline editing of the agent's prompt/context to test alternative reasoning paths.
• Variable Inspection: Show current agent state (context window, memory, tool results) at each breakpoint.

Technical Approach

• Extend the existing checkpoint system to support breakpoint markers
• Add step-through controls to the session replay UI
• Implement state inspection panel showing agent context at each step
• Add "branch and replay" capability from any breakpoint

Impact

This would make Peaky Peek the first open-source agent debugger with full interactive debugging capabilities, significantly differentiating from LangSmith (read-only tracing).

Labels

enhancement, paper-inspired, high-priority