loongsuite/dsh-plugin

LoongSuite observability for DeepSeek Harness

English | 简体中文

@loongsuite/dsh-plugin is a standalone, open-source observability plugin for DeepSeek Harness (dsh). It observes DSH's native session, agent loop, LLM stream, and tool lifecycle, converts them into OpenTelemetry GenAI traces and metrics, and exports standard OTLP/HTTP protobuf to any compatible backend.

LoongSuite is an open-source observability collection ecosystem built on OpenTelemetry. This repository is its native DSH integration. The plugin does not depend on or require LoongSuite Pilot, a sidecar, a local JSONL tap, or any particular vendor's backend.

Status: stable 0.1.x release. Install @loongsuite/dsh-plugin from npm or the DSH plugin market.

One DSH turn exported over OTLP into self-hosted Langfuse: four react steps, per-call latency and token counts, a failed web_search followed by bash fallbacks, and the ENTRY span's GenAI attributes. Content capture was enabled for this capture; it is off by default.

Data model

DSH session/event + llm/stream
                │
                ▼
      lifecycle coordinator
                │
                ▼
   LoongSuite GenAI OTel utility
                │
                ▼
 private TracerProvider + MeterProvider
                │  OTLP/HTTP protobuf
                ▼
  any OpenTelemetry-compatible backend

One DSH turn produces a single trace with this shape:

ENTRY
└── AGENT
    └── STEP
        ├── LLM
        └── TOOL

Each real LLM attempt gets its own LLM span, so retries remain visible under the same step. Tool calls are correlated with their results by DSH call ID. Errors, aborts, incomplete streams, and plugin shutdown close live spans with an error status instead of leaving them open. Subagent sessions create their own trace and carry DSH parent-session and delegation attributes.

When content capture is enabled, ENTRY and AGENT input messages contain only the turn's direct source.kind=user input. Synthetic DSH context such as runtime snapshots, agent instructions, skill catalogs, goals, and coordinator relays remains visible on the LLM span, but prior-turn conversation history is excluded so every trace contains only its own turn context. Later LLM spans in a tool loop retain assistant tool calls and tool results produced earlier in the same turn. ENTRY and AGENT output messages contain only the final stop response; a turn that never reaches stop falls back to its last available assistant message.