perrylink/dsh-observe

Observability export: session/event flows to OTLP traces+metrics and Langfuse observations (turn/step/tool/LLM span, token/cost count, desensitized prompt/completion), asynchronous batch + bounded persistent offline buffer + backoff retry, turned off by default and turned on explicitly

dsh-observe turns DeepSeek Harness's session/event stream into standard observability protocols: per-turn traces with deterministic ids linking turn, step, tool-call (duration, status, retry derivation) and LLM generation spans; per-provider/model token counters, USD cost counters from a configurable pricing table, and an optional context-pressure gauge from ctx.tokenMeter. Capture is sanitized before anything is queued or sent — prompt and completion bodies are redacted (structural key names + built-in secret patterns + your patterns) and truncated. Reliability features include async batching (size- and timer-triggered), a bounded durable offline buffer (storage-domain) with oldest-first eviction, deterministic exponential-backoff retries, and a runtime kill switch via an optional Typert remote (observe/status, observe/setEnabled). Off by default: nothing is captured or exported until enabled: true plus a backend are set.

Other ★ 0 updated 2026-08-21 — untested
View on GitHub ↗

Install

dsh plugin --profile web add dsh-observe

npm dsh-observe 0.2.4 verified 2026-09-04 (repository field → github.com/PerryLink/dsh-observe; README EN primary with zh/es/pt/hi editions). Install: dsh plugin --profile web add dsh-observe (npm) or dsh plugin --profile web add "github:PerryLink/dsh-observe#main" (git). Export is OFF by default — set enabled: true plus at least one backend in the profile patch row and restart.

Compatibility

DeepSeek Harness 0.1.2-alpha.5 (adapted 2026-09-02 per README); Node ^22.19.0 || >=24.0.0; backends: OpenTelemetry OTLP/HTTP (traces + metrics, JSON encoding) and Langfuse — either or both; model-agnostic, no model calls.

Details

Recent updates

0.2.4 is the current npm latest (verified 2026-09-04).

FAQ

What does it export?
OTLP traces + metrics (JSON encoding) to any OpenTelemetry-compatible backend, and/or Langfuse observations — turn/step/tool/LLM spans, token and USD-cost counters, optional context pressure.
Does it capture my prompts?
Only when enabled, and sanitized: prompt/completion bodies are redacted (structural keys + built-in secret patterns + your patterns) and truncated before anything is queued or sent.
Is exporting on by default?
No — off by default. You must set enabled: true plus at least one backend in the profile patch row; otherwise nothing is captured or exported.

Alternatives

linyp/dsh-plugin-langfuse · loongsuite/dsh-plugin

More plugins in Other

Browse more in Other

Guides for Other plugins