Two installable Claude Code Agent Skills that parse local JSONL session logs to produce a forensic regression report with no network calls or telemetry.
Key Takeaways
Reads ~/.claude/projects/**/*.jsonl already written by Claude Code; dedupes on (message.id, requestId) using the same scheme as ccusage to avoid double-counting resumed sessions.
Tracks 9+ headline metrics including read:edit ratio, write share of mutations, reasoning loops per 1K tool calls, thinking redaction rate, and tokens per user turn, each with published healthy/transition/concerning bands.
Auto-detects a drift inflection date via composite health score and argmax of |before - after| over candidate dates with a 0.75-sigma floor; falls back to median-timestamp split if no break clears.
Report generation splits work: a stdlib-only Python script (no pip, no Node) runs in ~2.5s to build a filled skeleton with ~20 narrative slots, then Claude fills prose in 10-20s inside your session.
Install via npx skills add delta-hq/cc-canary; outputs paste-ready markdown or a dark-theme HTML dashboard, both written to the invocation directory and never uploaded.