← ClaudeAtlas

session-insightslisted

Analyze local agent session stores into fact-bound usage insight reports — session counts, active days, hour/project distribution, tool ranking, friction signals. Use when the user asks how they actually use their coding agent, wants a weekly report, or asks to be roasted on real numbers; read-only, aggregates only, raw transcripts never leave the machine.
TeFuirnever/azhou-ai-hub · ★ 1 · AI & Automation · score 70
Install: claude install-skill TeFuirnever/azhou-ai-hub
# Session Insights **🦊 阿舟 · Session Insights** > 📊 先有数字,再有故事。 Invocation class: user-invoked orchestrator, declared in frontmatter; the axis, its semantics and the composition rules are defined once in docs/skill-standard.md §2.2 (spec #126 ID-9) — invoked by name or by an explicit "我的 agent 使用报告" request; it never chains into another orchestrator and nothing chains into it. 它读取本机 harness 会话存储,由标准库脚本确定性聚合,产出事实绑定的使用洞察报告。品牌层协议见 [brand-layer](references/brand-layer.md),报告口径见 [report-contract](references/report-contract.md),prior art 与零 vendor 声明见 [provenance](references/provenance.md)。 ## 四条硬边界 1. **只读观察者**:脚本只读宿主会话目录,永不写入、永不外联;读取私有会话日志的授权就是本次显式调用本身。 2. **机器先于叙事**:aggregate JSON 是唯一事实源;报告中的每个数字来自它,LLM 只做综合叙述与 tone,不重算任何指标。 3. **原始记录不出机器**:聚合与脱敏摘录进产物;原始转写与用户路径不进 Git、不进上下文全文。产物默认落在 `.azhou/session-insights/`(仓库 `.gitignore` 已覆盖)。 4. **同源比较**:跨 harness 的 token 口径不同,报告只做同源比较,不做假等价。 ## 触发与用法 显式调用:"用 session-insights 给我出一份 agent 使用报告(或周报)"。CLI 单入口五个子命令(脚本在已安装包目录下): ```bash python <skill-dir>/scripts/session_insights.py detect python <skill-dir>/scripts/session_insights.py discover --harness claude-code python <skill-dir>/scripts/session_insights.py metadata --harness claude-code python <skill-dir>/scripts/session_insights.py aggregate --out .azhou/session-insights/aggregates.json python <skill-dir>/scripts/session_insights.py report --aggregate .azhou/session-insights/aggregates.json ``` - `detect`:报告各家 harness 存储是否在场;Codex 已接入同一管线,zcode 可检测但保持 `unsupported`。 - `discover`: