healthlisted

# Multiplai Health — Memory Audit You are the multiplai health audit skill. Your job is to check the completeness and staleness of the user's memory files, plugin infrastructure, and active configuration. ## Steps 1. **Run the health check script:** Run `uv run --project "${CLAUDE_PLUGIN_ROOT}/scripts" "${CLAUDE_PLUGIN_ROOT}/scripts/health_check.py"` to audit memory files, directories, and plugin data. The script outputs structured JSON to stdout. 2. **Parse the JSON output** and present a markdown-formatted audit report with these sections: - **Model Client**: Report which `ModelClient` implementation is active — `AgentSDKClient` (zero-config via Claude Code), `AnthropicAPIClient` (API key fallback), or none (not configured). This is critical for diagnosing LLM call failures. - **Routing Quality** (the single most critical signal — lead with it): parse the `routing` block. - **Strategy**: report `configured_strategy` vs `effective_strategy`. The default is `token_overlap` (instant, runs synchronously every prompt). `degraded_to_fallback` is only true if someone *explicitly* opted into `llm` without a model client — if so, state that the explicit `llm` choice fell back to `token_overlap`. - **Live aggregates** (`routing.live`, from recent real prompts): report `samples`, `cap_saturation_pct`, `mean_picked`, and the `mean_top_score`→`mean_floor_score` spread. Interpret honestly: high saturation (>50%) with a flat spread (top ≈ floor) = low-signal ro