Skip to content

Releases: linyp/dsh-plugin-langfuse

Release list

v0.1.0 — first release

v0.1.0 — first release Latest
Latest

Choose a tag to compare

View all tags
@linyp linyp released this 14 Aug 05:47
v0.1.0
6b7dcbf

First release: Langfuse observability for DeepSeek Harness (dsh).

Exports each agent session as an OpenTelemetry trace tree to Langfuse's OTLP endpoint — turn → trace, model step → generation (model identity, gen_ai.usage.* tokens, time-to-first-token), tool call → tool span nested under its requesting generation — implemented as a backend of the harness's public telemetry seam (@deepseek-ai/dsh-session-telemetry), so redaction waterfall, FEEDBACK_ONLY consent gating, and teardown draining are inherited from the seam. Architecture decisions: README.

Verified

  • Real Langfuse cloud round trip: exported a live session to Langfuse cloud (US), read it back through the public API, and asserted the ingested tree — session grouping, GENERATION model/usage (11/3 and 7/5 tokens), and the bash TOOL observation nested under its requesting generation. Ships as a key-gated e2e (tests/langfuse-cloud.e2e.ts, self-skips without keys).
  • Install path: dsh plugin --profile web add … into an isolated DSH_HOME, --dump-config confirming the official OTLP-logs telemetry row composes disabled and this row composes with env-driven config, and a clean boot of the assembled web profile in FULL mode.
  • 21 unit tests (folding projection, config fail-loud paths, HMR-safety disposal) + keyless REAL-composition e2e against a mock OTLP collector, on Node 22/24/26.

Install

Once the package is on npm:

dsh plugin --profile web add dsh-plugin-langfuse
export LANGFUSE_PUBLIC_KEY=pk-lf-…
export LANGFUSE_SECRET_KEY=sk-lf-…
export LANGFUSE_HOST=https://us.cloud.langfuse.com   # your region

Until then, the attached tarball installs identically: dsh plugin --profile web add ./dsh-plugin-langfuse-0.1.0.tgz.

Compatibility

Pinned against @deepseek-ai/dsh-*@0.1.0-rc.6 (dsh is in developer preview with no compatibility promise; CI runs an upstream canary against the newest published versions).

首个版本：为 DeepSeek Harness 提供 Langfuse 可观测性。将每个 agent 会话导出为 OpenTelemetry trace 树（turn → trace，模型 step → generation，工具调用 → 嵌套 tool span），作为 harness 公开 telemetry seam 的后端实现，继承脱敏 waterfall 与 FEEDBACK_ONLY 同意语义。已完成真实 Langfuse cloud 回读验证与 dsh plugin add 安装路径实测；架构决策详见中文 README。

Assets 3
Loading