Cavan-Ou/dsh-observation-journal

README 摘要

dsh-observation-journal Zero-touch runtime telemetry for DeepSeek Harness: every session writes its own report card. 简体中文版见 README.zh.md A pure observer plugin (zero tools registered, zero LLM calls, zero agent involvement). When a session ends, it writes the run's facts — task, model tier, tools, failures, duration, status — into a human-readable journal with an auto-updated stats section. Why it exists Failures have a recorder (dsh-fail-logger). Successes and run facts didn't. This is the sibling: what happened — not how to solve it, not what to remember. No tools, no injection, no retrieval. The harness writes passively; humans and projects read. What it is not (boundaries, stated plainly): Plugin Records Injects back into agent? Consumption this run facts (telemetry) never human/project file dsh-task-planner solutions ("how to solve") yes (recall) agent planning dsh-mneme / dsh-memento / dsh-memory agent memories yes (retrieval) agent context dsh-fail-logger failures via skill agent skill loading 60-second verification What the output looks like The journal is the UI. A marker section that survives manual edits, plus an auto-stats block: raw sidecar ( obsFile + '.jsonl' , appen…

分类