dsh-system1-observer
在代理循环的配置节点调用 System One 模型，并将每次调用、请求和响应写入 JSONL trace。不做任何决策。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-system1-observer@0.2.1System1 Observer 以 dsh-system1-observer 发布，当前版本为 0.2.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Human's words: This is vibe-coded. Agents could take initiative to, and are mostly free to gather rubric and evaluate using system1 call on parts or whole of sessions. Almost all is volatile. Agent can change configurations which will be hot-loaded. Measurement being important, the actions to make use of the evaluation are interesting and useful. The "seams" are there to feedback to the human or agent, and the current state of this project provides a flexible basis.
Following is written by agents, self-evaluated and compared notes, with README writing standards they researched and formulated. These words are results of two LLM in DeepSeek Harness and a TypeSafe jev model in dsh-system1-observer plugin.
A seam observer for the DeepSeek Harness. It asks a System One model what the agent loop is doing — at the points you choose — and writes every call, every skip and every answer to a JSONL trace.
It decides nothing. Each listener returns the loop's own decision by reference; every trace line says
enforcement: "declarative", verified: false. Its product is evidence about the decision model's replies at each point
of the loop, not judgements about the session it is observing.
This page is short on purpose: it is capped at 100 lines by test/readme.test.js, and the reference material —
every setting, the seams, the trace, and the measured claims with the ones that are not verified — is in the
manual.
dsh plugin --profile web add dsh-system1-observer
web is the profile to install into — substitute the name of yours. Or from a clone, which needs no npm account:
git clone https://github.com/johnlam1968/dsh-system1-observer.git
cd dsh-system1-observer && ./install.sh web
It needs the profile's harness for @deepseek-ai/cordis and @deepseek-ai/dsh-tools (they are peerDependencies,
not bundled), and a model to call: the profile's system1 service, or the HTTP wire at wireUrl (default
http://127.0.0.1:8766).
system1_explain what it can do, what it is set to now, what a reading covers
system1_sessions { action: 'list', search: 'XYZ' } find the session
system1_evaluate_session { sessionId: '…', segmentChars: 57600, package: true } judge it, and package the run
system1_measurements { action: 'interpret', package: '…', text: '…', by: 'you' } attach your reading
system1_evaluate_session does the mechanical work in one call: it reads the session, segments it if it is large, asks
every question of every segment, aggregates in code and writes a report package. The last step is a separate call, and
deliberately yours — prose cannot be derived from numbers.
Nine tools in all: system1_explain, system1_settings, system1_decide, system1_evaluate_session,
system1_measurements, system1_sessions, system1_question_sets, system1_battery, system1_trace.
~/.dsh/logs/system1-observer.jsonl — one line per call, per skip and per start-up, read by the
system1_trace tool, the trace card, or scripts/trace.mjs.segmentChars — a large session needs segments.system1_explain prints these in full, derived from the running row.
npm run ci # the six gates in order: test, coverage, citations, compat, composition, deploy
The rules this repository holds itself to, each with the measurement that produced it, are in the conventions.
Ask in GitHub issues; pull requests are welcome. How a release happens, and what it does not ship, is the maintainer's runbook, kept off-repository with the other notes.
| the chart | the whole row in one picture: the data path from the loop to the trace, the nine tools, the repository around it |
| manual | the reference: hooks, every setting, the trace, mounting, the measured claims and the unverified ones |
| docs/settings.md | every setting: what it does, what it costs, why it exists |
| docs/measurement-depth.md | what evidence a judgement gets (G0–G4) |
| docs/question-suitability.md | which question type fits which evidence |
| case study | two agents, one rubric and this page's own redesign, measured — the method, the numbers and the mistakes |
MIT — see LICENSE.