Repository navigation
Releases: beibeihk/dsh-agent-observability
Release list
v0.1.0 — Event-grounded agent observability
v0.1.0
First community release of an event-grounded, privacy-first observability and reliability plugin for DeepSeek Harness.
- Installs through the official plugin manager from npm or the identical prebuilt
.tgzasset, usingdsh.bundle.patch; no core changes. - Reconstructs session/turn/step/tool trajectories through durable events and async session-query leases; provides
/observewhen human commands are available. - Computes reliability metrics and nine deterministic F-rules plus informational tool-error evidence; exports JSON/JSONL/CSV/evaluation cases, plaintext and offline HTML.
- Defaults to metadata only, redacts optional captured content, excludes reasoning/stream data, and sends no telemetry.
- Includes 36 offline tests, five golden fixtures, semantic with/without comparison, real persistence resume/fork/disposal tests, the official CLI installation demo, 130 synthetic traces and transparent performance results.
- English/Chinese README, technical report, blog draft, study guide with 35 answered interview questions, and factual CV/profile material.
Tested: published DeepSeek Harness 0.2.0-rc.2, Cordis 4.0.4. No compatibility claim for the alpha channel or future releases.
Install:
npm install --global @deepseek-ai/dsh@0.2.0-rc.2 pnpm
dsh plugin --profile web add @beibeihk/dsh-agent-observability@0.1.0
dsh webRun a task and submit /observe; open its local report.html. npm v0.1.0 was published on 2026-10-15 (Asia/Hong_Kong). Its integrity matches the original attached release tarball, which remains an alternative installation source. Publication and registry installation verification.
Known limits: pattern-based privacy protection; usage only where reported; no task-correctness or progress judge; individually atomic files rather than a multi-file transaction; no automatic disk pruning; no source-watcher HMR stress claim. Metadata observation added 34.20 μs/event (+79.67%) in the recorded 10,000-message official Session-dispatch microbenchmark. This excludes model/disk latency. Synthetic precision/recall 1.0 is not real-world accuracy.