mycelium-hq/ai-brain-starter/longitudinal

When the user types /longitudinal, run the multi-year correlation pass and surface only the strongest signals across years of health-mcp + journal data.

Language

Generate the report in the language the user writes in. If Spanish, all sections including the panel commentary are in Spanish.

Scope resolution

Parse the argument for window:

• all -> earliest record in DB to today
• Ny -> last N years (e.g. 5y, 3y)
• Nm -> last N months (e.g. 18m)
• blank -> 365 days

If all, query the DB for the earliest record date first:

SELECT MIN(start_date) FROM records WHERE value IS NOT NULL

Step 1: top_signals first (the noise filter)

Always call health_top_signals(vault_root=..., lookback_days=N, min_strength="moderate") first. This is Lara Briden's dissent codified: most correlations are noise. The substrate has already filtered. Start with what's left.

If signal_count == 0, report "no signals above noise threshold for this window" and stop — do not invent. Surface what IS there: the deltas and r-values that didn't quite clear the threshold, in case the user wants to relax it.

Step 2: Floor x body fingerprints for the user's top 3 Floors

Load the journal index, count Floors in the window, take the top 3 by occurrence.

For each Floor, call:

health_floor_body_fingerprint(floor=<name>, vault_root=..., lookback_days=N)

Report the body fingerprint deltas (HRV, RHR, sleep efficiency, cycle phase distribution). If delta_pct exceeds ±10% AND n_on_floor >= 10, this is a real fingerprint. Below that, mention it but flag as "weak."

Step 3: Sleep architecture trend

Call health_sleep_architecture(start, end) for the window AND for the prior matching window (e.g. 1y now vs 1y prior). Compare REM%, Deep%, Core%, efficiency. Flag drift > 5 percentage points.

Step 4: Longitudinal markers

Call health_longitudinal_summary(start, end, granularity="quarter"). Pull HRV baseline, VO2max, lean body mass, walking steadiness, sleep efficiency by quarter. Compute trend slope per marker (rough linear regression: (last - first) / first * 100).

Surface only markers with > 5% drift across the window.

Step 5: Symptom correlate (if symptom data present)

Call health_symptom_correlate(symptom_type=None, vault_root=..., lookback_days=N). If any symptom has > 15% delta_pct on a body metric AND n >= 5 days with the symptom, surface it.

Step 6: Named loops (optional, if the user has /patterns output recently)

If [VAULT_PATH]/Meta/Patterns/loops-detected.json exists (created by /patterns), read it. For each named loop with date list, call health_loop_signature(loop_dates_iso=[...], vault_root=...). Surface the loop body fingerprint.

If the file doesn't exist, skip this step silently — no error.

Report format

## Longitudinal pattern review -- <window>
 
### What survived the noise filter
<bullet list of top_signals results, strongest first>
 
### Floor x body fingerprint
<for each top 3 Floor: name, occurrence count, body deltas with > 10% threshold>
 
### Sleep architecture
<current vs prior window comparison, flag drift>
 
### Longevity markers (per quarter)
<table or bullets of HRV baseline, VO2max, lean mass trend>
 
### Symptoms
<symptom correlates with > 15% delta>
 
### Named loops (if /patterns ran recently)
<loop body fingerprints>
 
### Health panel commentary
<3-5 voices from the Health & Body section of advisory-panel.md commenting on the strongest signals. Required: Peter Attia (longevity), Stacy Sims (cycle phase if any female-physiology pattern is present), Bessel van der Kolk (HRV-vagal-tone if Floor x body signal is strong), Lara Briden (dissent: did we pick signal or noise). Each voice: 1-2 sentences referencing a SPECIFIC number from the report.>
 
### What to do next
<3-5 concrete, action-shaped items based on the strongest signals. NOT "consider X" -- "do X this week.">

Critical: zero fabrication

Every number in the report comes from a health-mcp tool call. If a tool returns error, surface the error verbatim and skip that section. Never invent a metric. Never round so heavily that a 38.2 ms HRV becomes "around 40."

If the lookback window includes a period before the user started recording (e.g. they ran /longitudinal all but the DB only has 3 years), report the actual span you found: "earliest record: 2023-04-12, so the window is 2023-04-12 to 2026-05-10 (3.1 years)."

When to escalate

If a single marker shows > 25% drift in the wrong direction (HRV baseline dropping > 25%, VO2max dropping > 25%, lean mass dropping > 10%), surface it AT THE TOP of the report with the marker name in bold and a one-line note. These are health-significant changes worth flagging directly. Do not bury them in section 4.