devmom/dsh-trajectory-debug

README 摘要

DSH Plugin: Trajectory Debug Workbench (trajectory-debug) English 中文 Target host: DeepSeek Harness ( @deepseek-ai/dsh v0.1.0-rc.x, developer preview) Turn DSH's event-sourced sessions into debuggable assets: a waterfall trajectory view, deterministic single-step replay (zero token cost), breakpoints, sandboxed edit-and-rerun, fork comparison, and performance analytics — plus OTel GenAI trace export and trajectory model tools for agents to introspect their own runs. Packages Commands Feature Status Capability Notes Waterfall engine buildTrajectoryPage : turn/step/tool rows, status, timing, tokens, error codes, filters, paging Deterministic replay stepContextAt : model view + action per step; ReplayCursor step/seek — zero token, zero tool execution Performance analytics analyzePerf : success rates/percentiles, failure taxonomy, token distribution, TTFT/decode, turn stats (optional price table → cost) Fork compare compareTrajectories : step alignment, tool changes, result diffs, summary Breakpoints BreakpointManager : agent/pre-step waterfall short-circuit + timeout auto-resume Edit-and-rerun rerunTool through the full ctx.tools pipeline; record\ sandbox\ ask (ask fails closed without…

分类