agent-insightslisted

# agent-insights Cross-tool version of Claude Code's builtin `/insights`. Everything runs locally; no session data leaves the machine. `SKILL_DIR` below means this skill's directory; `SCRIPT` means `python3 <SKILL_DIR>/scripts/agent_insights.py`. All intermediate files live in `~/.agent-insights/` (override with `--data-dir` on every subcommand, e.g. for tests). ## Arguments One optional argument: number of days to analyze. Default 30, clamp to 1..365. Non-numeric input: tell the user, then proceed with 30. ## Pipeline ### Step 0 — Detect runtime ```bash SCRIPT detect ``` Detects whether this skill is running on the host machine (where the user's agent logs live) or inside a **sandboxed runtime** such as Claude Cowork, whose Bubblewrap container does not mount the host filesystem. Reads the JSON result: - If `sandboxed` is `false`, continue to Step 1. - If `sandboxed` is `true`, the host machine's agent log directories are **not reachable** from this runtime. Do **not** run a full report from here: it would be empty or cover only the sandbox's own session. Instead, tell the user: - the detected `runtime` and `confidence`, and the `signals` that fired; - the `recommendation` verbatim (run `/agent-insights` from a non-sandboxed agent with direct host access, e.g. Claude Code in the terminal); - `reachable_sources` (what little, if anything, is visible from inside the sandbox). Then stop, unless the user explicitly asks to analyze only what is reachable