dsh-llm-call-inspector

In DeepSeek Harness Web, the Trajectory view narrates the full session story, yet it leaves a gap: there is no window where you can inspect exactly what the model received and what it returned for each normalized llm/stream call. dsh-llm-call-inspector fills that gap with an adjacent LLM Calls view that lists every session-scoped call in newest-first order, preserving the captured request snapshot and ordered response chunks without altering what the model receives or the caller gets back.

Key capabilities include a three-tab inspector (Request, Response, API Comparison), search with status and purpose filters, expandable JSON with copy actions, and session-scoped state restoration so your selected call, filters, and scroll positions survive view switches. The API Comparison tab is a semantic shape projection rather than a packet capture: the left side shows the actual DSH-normalized data observed, while the right side maps to mainstream shape families such as OpenAI, Anthropic, Gemini, Bedrock, and Mistral based on call-level protocol evidence, with a manual reference fallback available. Opening the tab or copying content never triggers an additional LLM call or incurs provider charges.

This plugin suits developers working with DeepSeek Harness 0.1.2-alpha.2 who need to audit normalized call boundaries during local debugging, multi-agent orchestration, or prompt engineering, compare protocol shape families, and track down token truncation or tool-argument anomalies without changing runtime behavior.