By name: mentat-debrief description: | Generate the modernized end-of-session insight report. Replaces and exceeds the stock session-report plugin: full state-machine timeline, Q-table delta, drift event log, contradiction map, source-grounding audit, sub-agent provenance, and a forward-motion next-actions list. Renders as a webpage artifact in the user's modernized aesthetic. when_to_use: | Trigger phrases: "debrief", "session report", "wrap up", "end-of-session", "what did we do", "summarize the session", "session review". Fire on explicit request only — debrief is expensive and shouldn't auto-run. disable-model-invocation: false allowed-tools: "Read Write Bash(python3 ${CLAUDE_PLUGIN_ROOT}/bin/mentat*)"
You are generating the session debrief. The renderer is canonical — you do not author HTML. A deterministic Python script reads the insight bus + Q-table
- session record off disk and produces the full multi-section artifact in Daeron's modernized aesthetic.
Procedure:
Run the renderer:
python3 ${CLAUDE_PLUGIN_ROOT}/skills/mentat-debrief/scripts/render.py \ --session ${CLAUDE_SESSION_ID} \ --out /tmp/mentat-debrief-${CLAUDE_SESSION_ID}.htmlThe script prints the artifact path and a one-line size summary to stdout.
Surface the path to the user. The HTML is self-contained (Google Fonts + inline CSS/SVG only) — no follow-up rendering or styling required.
If a forward-action recommendation is missing or under-specified, you may add a one-paragraph qualitative narrative in your reply, but DO NOT regenerate the HTML — the renderer is canonical. Twelve sections are guaranteed to populate: banner, state timeline (Gantt), state machine graph (SVG with weighted edges + Gaussian glow on most-visited node), tool routing learned (Q-table top-3 per state), drift audit, sub-agent ledger, reward distribution (top/bottom sparkline bars), source-grounding score, contradiction map, entropy spikes, tool ledger, and three forward moves tied to state + insight evidence.