mycelium-hq/ai-brain-starter/health-context

health-context, biometric context for substrate skills

Wraps the vault-aware tools in health-mcp so other skills can pull biometric context without each one re-implementing the connection logic.

This skill is invoked WITH another skill, never standalone. It's the "look up what the body was doing during this moment" layer.

When to use

• User invokes /journal or /daily-journal: pull health_journal_context(today) and prompt with HRV, sleep duration, sleep efficiency, steps, workouts
• User invokes /coaching: pull health_coaching_context(start, end, vault_root) for the coaching window
• User invokes /panel or strategy/decision question fires the advisory panel: pull health_panel_context(today, vault_root) for delta-vs-7d
• User invokes /patterns: pull health_floor_correlation(metric, days, vault_root) for HRV-vs-Floor and other biometric-vs-Floor patterns
• User invokes /weekly or /monthly (insights): pull health_weekly_rollup(week_start) and fold into the review

Do NOT use for:

• Querying live data without a prior import (suggest /ingest-health first)
• Writing to the vault (this skill never writes)
• Replacing medical advice (the scores are directional, not diagnostic)

Graceful degradation

If health-mcp is not registered, has no data, or the tool call errors:

• Fail silently. Continue the host skill (journal, coaching, etc.) without health context.
• Surface a one-line note in the host skill's output: "(health context not available; run /ingest-health to set up)"
• Never block the host skill on missing health data.

Composition by host skill

Inside daily-journal

When the journal skill assembles its prompt, call health_journal_context(date) for today. Add to the prompt:

Body context (today, from health-mcp):
  HRV: 28ms (vs 42ms 30-day baseline -- 33% below)
  RHR: 65 bpm
  Sleep: 5h 12m asleep (efficiency 87%, REM 38m, deep 22m)
  Steps: 4,820
  Workouts: 0
  Mindful: 0 min

This becomes a journal-prompt input: "Your body had a tough night. How does that map to what you noticed?" The journal skill owns the prompt; health-context owns the data fetch.

Inside coaching

When the coaching skill spans a multi-day window (e.g. last 14 days of accumulated tension), call health_coaching_context(start, end, vault_root). Surface:

• Average recovery score over the window
• Days with low recovery (< 50)
• Days with low restorative sleep (REM + deep < 60 min)
• Floor distribution from journals in the window

This gives the coaching skill a "body track" alongside the emotional track.

Inside advisory-panel

When a panel fires on a strategy/decision moment, call health_panel_context(today, vault_root). Surface:

• Today's recovery score
• 7-day average
• Delta (today minus 7-day avg)
• Today's Floor tag (if a journal exists for today)

If delta < -10 AND today's floor_level is in the lower third, the panel can fold "your body and floor are both low; consider deferring high-stakes calls" into its synthesis. Never override the user's decision; only inform.

Inside patterns

When the patterns skill runs, call health_floor_correlation(metric, days, vault_root) for each of: HRV, RHR, sleep duration, steps, recovery score. Surface any correlation with n >= 10 and |r| >= 0.25. Examples the patterns skill can render:

• "Low-Floor days correlate with low HRV (r=-0.42, n=38)"
• "Steps below 5k correlate with floor level <= 6 (r=-0.31, n=42)"

Inside insights (weekly / monthly)

When the insights skill runs /weekly or /monthly, call health_weekly_rollup(week_start) for each week in the review. Add to the rendered review:

• HRV avg / min / max
• RHR avg / min / max
• Steps total / daily-avg
• Workout count + total minutes
• Recovery trend (positive / flat / negative)

vault_root

Every vault-aware tool needs vault_root. Resolve it from:

1. The host skill's environment, if it sets VAULT_ROOT
2. The CLAUDE.md path's parent directory (search up from cwd)
3. Ask the user only if neither is available

Never hardcode a path. Pass as a tool argument every call.

Invocation pattern

This skill does not have its own / slash command. It piggybacks on the host skill via the host skill's invocation. The host skill MUST call health-context at the start of its prompt-assembly phase, before LLM synthesis, so the body data is available throughout.

Example (inside daily-journal's orchestrator):

# Inside daily-journal
try:
    body = await call_tool("health_journal_context", {"date_str": today.isoformat()})
    prompt += f"\nBody today: HRV {body['hrv_ms']}ms, RHR {body['rhr_bpm']} bpm, sleep {body['sleep_asleep_min']}min..."
except Exception:
    # health-mcp not available; continue without
    prompt += "\n(health context not available; run /ingest-health to set up)"

Privacy

Reads only. Never writes the vault, never writes the health DB, never sends data over the network. Health Auto Export TCP-live is the one exception, and that's local-Wi-Fi only.