dsh-visual-trace

In DeepSeek Harness, Web consoles, headless profiles, ACP hosts, SDK adapters, and Hooks plugins each interpret execution traces in their own way, forcing developers to hop between interfaces and manually compare raw event fields during debugging. dsh-visual-trace treats the official persistent session/event log as the single source of truth and applies one shared set of node semantics, natural-language explanations, and review signals across every runtime mode, so trace inspection no longer depends on a particular canvas.

Capabilities span three layers. The Web tab offers a filterable timeline, expanded tool sub-calls, inline review confirmation, and right-click quoting into the dialogue draft, with summaries in Chinese or English that require no extra model calls. In headless mode, traces are written to stderr while the final model answer stays on stdout, leaving existing pipelines intact. ACP, SDK, custom UI, and Hooks plugins call ctx.visualTrace.build() and render() to obtain the same standard node structure as the Web timeline, usable in protocol responses, audit files, or external dashboards. Quoted content automatically redacts common API keys, tokens, secrets, and passwords and flags the material as untrusted execution evidence.

Ideal for developers working across Web, headless, ACP protocol, and custom terminal environments on DeepSeek Harness 0.1.0-rc.6 or later, and for platform teams that need a unified trace report inside Hooks audit plugins. When official event types evolve, only the session-adapter and client adapter files need updating; no output interface requires a separate rewrite.