Parasym

project 2026-07-08 Search related

iOS app that connects to a Morpheus M7 (or any BLE chest strap exposing R-R intervals) and correlates the realtime HR + HRV data with Apple HealthKit to produce personalized health insights. Realtime workout mode (HRR + motion fusion) is fully deferred to v0.2.

Status (2026-06-22): Planning + scope complete (rev 3). No code yet. Full rethink/rebuild authorized by user. Open source (MIT), public GitHub, completely free (no tip jar, no paid tier). Built on the agent-skills lifecycle.

The vision

A local-first, privacy-respecting, open-source iOS app that sits between a chest strap (Morpheus M7 / Polar H10 / any BLE HRM exposing R-R) and Apple HealthKit, and produces insights neither source can produce alone. Morpheus gives you the HRV-driven training score; Apple gives you sleep, RHR, weight, workouts. Parasym is the correlation layer — the thing that says "your HRV dropped 1.8σ last week, and it correlates with the 3 nights you slept <6h after the late-week training block."

It's not a Morpheus competitor. Morpheus owns the training-zone moat. Parasym wins on correlation + openness + ownership of your data.

Differentiators

vs Morpheus:

Cross-source correlation — Morpheus only knows its own data; Parasym correlates HRV with sleep, RHR, weight, workouts
Open source — algorithm auditable. Critical for a health app.
Local-first — no required backend. Data lives on iPhone. Optional encrypted iCloud backup (v0.3+).
No vendor lock-in — works with any BLE chest strap exposing RR intervals (Morpheus M7, Polar H10, Garmin HRM-Dual)
Public methodology — every insight cites the underlying science

vs WHOOP / Oura / Fitbit:

No subscription — completely free, no in-app purchases
No proprietary data lock-in — full CSV / JSON export
Transparent algorithms — every "your recovery is X%" insight has a citation
Works with the user's existing chest strap — no new wearable required

MVP scope (v0.1) — what's IN

Core data layer:

iOS app, Swift + SwiftUI, iOS 17+
CoreBluetooth listener for the standard Heart Rate Service (decodes R-R intervals, not just HR)
Compute both RMSSD and SDNN from the same R-R intervals (resolves the metric mismatch — see Hrv Sdnn Vs Rmssd)
HealthKit reads: heartRateVariabilitySDNN, resting HR, sleep analysis, weight, workout metadata (read-only — no live workout integration in v0.1)
Local persistence (SwiftData, encrypted at rest with iOS Data Protection Complete)

Insight engine + UI:

Daily recovery score (0-100) — RMSSD-weighted, Galpin's σ-deviation rule (after 30+ days of baseline)
7-day / 30-day trend visualizations
Cross-source correlation insights ("HRV ↓ when sleep < 6h", "RHR ↑ when training load ↑")
Anomaly alerts (1σ yellow, 2σ red, persistent-2-day rule)
CSV / JSON export (non-negotiable)
Local notifications with user-configurable quiet hours
Source citations on every insight ("based on Galpin 2024 / Jamieson 2024 / Shaffer 2017")

OSS + free constraints:

MIT license
Public GitHub repo (URL TBD; not yet created)
No monetization of any kind (no tip jar, no paid tier, no in-app purchase)
README, contribution guide, code of conduct (the boring but important OSS surface)
v0.1 timeline: 6-8 weeks of focused build (revised back from the +2-3 week realtime expansion)

Out of MVP scope (deferred to v0.2+)

Realtime workout mode (v0.2 — BLE streaming, Apple Watch motion fusion, post-workout HRR analysis, live in-workout HRV UI) — per user decision 2026-06-22, deferred in full from the v0.1 slice
HRR (heart rate recovery) tracking post-workout
HRR deep analysis (rolling HRR across weeks, HRR trend vs cardio fitness over time)
Apple Watch companion + watchOS complication
CloudKit sync for multi-device
Apple Watch ECG deep integration (parse HKECGSeries for gold-standard cross-validation)
Mood / soreness / subjective check-ins (Olav Bu's "felt off / felt great")
Web dashboard
Android version
AI-synthesized daily recommendations (Olav Bu #331 frontier)
TARS integration (v0.3+ — see Parasym Tars Integration)

Architecture sketch

┌──────────────────────┐         ┌──────────────────────────┐
│  Morpheus M7 (or     │ ─BLE──▶ │  iOS App (Swift/SwiftUI) │
│  Polar H10 / any     │ HR+RR   │                          │
│  BLE HRM exposing RR)│         │  ┌────────────────────┐  │
└──────────────────────┘         │  │ BLE Source         │  │
                                 │  │ (CoreBluetooth)    │  │
┌──────────────────────┐         │  └─────────┬──────────┘  │
│  Apple HealthKit     │ ──────▶ │  ┌─────────▼──────────┐  │
│  - SDNN HRV          │  read   │  │ HRV Compute Engine │  │
│  - RHR               │         │  │ RMSSD + SDNN       │  │
│  - Sleep stages      │         │  │ Accelerate / vDSP  │  │
│  - Workouts          │         │  └─────────┬──────────┘  │
│  - Weight            │         │  ┌─────────▼──────────┐  │
│  - Respiratory rate  │         │  │ Insight Engine     │  │
│  - SpO2              │         │  │ Galpin σ rule      │  │
└──────────────────────┘         │  │ Daily score       │  │
                                 │  │ Correlations      │  │
                                 │  └─────────┬──────────┘  │
                                 │  ┌─────────▼──────────┐  │
                                 │  │ SwiftData store    │  │
                                 │  │ (encrypted)        │  │
                                 │  └─────────┬──────────┘  │
                                 │  ┌─────────▼──────────┐  │
                                 │  │ SwiftUI surfaces   │  │
                                 │  │ Dashboard / Trends │  │
                                 │  │ Insights / Export  │  │
                                 │  │ Settings           │  │
                                 │  └────────────────────┘  │
                                 └──────────────────────────┘
                                          │
                                          │  (v0.2+)
                                          ▼
                                 ┌──────────────────────┐
                                 │  Workout Mode        │
                                 │  (BLE streaming +    │
                                 │   motion fusion +    │
                                 │   HRR analysis)      │
                                 └──────────────────────┘

Stack: Swift 5.10+ · SwiftUI · SwiftData · CoreBluetooth · HealthKit · Accelerate/vDSP · XCTest + Swift Testing · GitHub Actions · App Store + TestFlight.

TARS integration (future-facing)

The user (2026-06-22) wants Parasym to connect to TARS so TARS can know the user's health data. Explicitly not a now feature but should be considered in the v0.1 architecture. Four architectural options + recommendation captured at Parasym Tars Integration. Recommended: Option 1 (file-based daily summary → gbrain) for v0.3; Option 2 (MCP server) for v1.0+ if there's traction.

Class-level pattern promotion (2026-06-22, meta-event): the user's three TARS design decisions — (a) per-data-type privacy configuration, (b) "no, just aware" (content-awareness, no behavior-adaptation), (c) citations + caveat on proactive mentions — were elevated to a class-level pattern in the brain-ops skill: references/agent-integration-with-sensitive-data-design-pattern.md. The pattern is reusable across TARS + Calendar, TARS + Location, TARS + Financials, and any future AI agent + sensitive-data integration. The canonical case study for this pattern is Parasym Tars Integration. This means the v0.1 work that lands first will inform the integration pattern for every future sensitive-data integration in the user's agent stack.

What v0.1 must accommodate now (cheap, no extra work):

Data model JSON-exportable
Insights produce NARRATIVE outputs (what + why + when + source citation), not just numbers
No proprietary locks on insight logic — every insight is grounded in public-domain methodology
Settings page architecture with a SettingsSection / SettingsRow pattern that can absorb new sections (incl. the v0.3 TARS access toggles) without rewrite
Defensive privacy by default in the export layer — flag sensitive data types (weight, GPS) so the v0.3 toggle system has something to bind to from day 1

Decisions log (2026-06-22)

User answered all 6 open questions + added 1 new direction (TARS) + refined with 3 preferences (TARS) + 1 reversion (realtime → v0.2):

#	Decision	Version	Impact
1	Public GitHub, MIT	rev 1	OSS discipline from day 1 — README, contribution guide, code of conduct are v0.1 deliverables
2a	Realtime workout mode in v0.1 (data + post-workout analysis)	rev 1 → rev 2	+2-3 weeks, +10-20% surface
2b	Realtime workout mode fully deferred to v0.2	rev 3 (this)	v0.1 cost reverts to 6-8 weeks; v0.2 owns the whole feature (data + post-workout + live UI)
3	Completely free (no tip jar, no paid tier, no IAP)	rev 1	The "why" is intrinsic, not commercial. No backend — a paid cloud bill with zero revenue is unsustainable.
4	watchOS companion: defer to v0.3	rev 1	Confirmed
5	AI insights (Olav Bu #331 frontier): defer to v0.3	rev 1	Confirmed — useless before 30+ days of baseline
6	App Store for v1.0	rev 1	Free download. "Wellness insights, not diagnostic" copy avoids medical-claim blocker
T1	TARS integration: deferred, v0.1 architecture must accommodate it	rev 2	v0.1 work that benefits v0.3+ integration
T2	TARS privacy boundary: per-data-type toggles in settings, sensitive defaults OFF, insights (not raw values) are unit of sharing	rev 2	Direct design constraint on v0.1 settings architecture
T3	TARS aware but does NOT modify its own behavior	rev 2	"no, just aware" — TARS uses data in content, behavior (tone, length, formatting) is stable
T4	TARS health mentions: citations on all, caveat on proactive mentions	rev 2	Both citations + caveat, not either-or

Tradeoffs I want to push back on (per agent-skills meta rule)

Don't claim medical-grade accuracy in marketing. Consumer hardware; insights are educational, not diagnostic. Wording matters for App Store review and user trust.
Don't compete with Morpheus on training zones. Their moat, 10+ years deep. Parasym's edge is correlation + openness.
Don't ship "your recovery is X%" without a citation. Every insight surfaces the underlying science. This is the differentiator vs Morpheus's black-box scoring.
Don't ship "AI insights" before 30+ days of personal baseline data. The first month is data collection; insights become meaningful after baseline is established. (Aligned with user's defer-to-v0.3 decision.)
Don't ship without CSV export. Non-negotiable for a health app targeting the quantified-self / longevity audience.
Don't require a backend. Local-first is the privacy story. Any required cloud undermines it. CloudKit as optional sync is fine; required cloud sync is not. (OSS + free amplifies this — a paid cloud bill with zero revenue is unsustainable.)
Don't ship realtime workout mode in v0.1. The data foundation + post-workout insight was the v0.1 deliverable; v0.2 owns the whole feature now. Keep v0.1 scope honest.

Agent-skills alignment

Per the addyosmani/agent-skills lifecycle ([repo](https://github.com/addyosmani/agent-skills)):

Phase	Skill	Status
Define	idea-refine + spec-driven-development	COMPLETE (this doc is the spec, rev 3)
Plan	planning-and-task-breakdown	NEXT — after scope sign-off
Build	incremental-implementation + frontend-ui-engineering (for UI)	FUTURE
Verify	test-driven-development	FUTURE
Review	code-review-and-quality + code-simplification	FUTURE
Ship	shipping-and-launch	FUTURE

The user invoked "use the agent skills" — this plan operates under that methodology. Surfacing assumptions, pushing back, and managing confusion are non-negotiable operating behaviors per the meta-skill.

Linked concepts

Joel Jamieson — primary methodology source (RMSSD + recovery score)
Olav Aleksander Bu — readiness check-in philosophy
Andy Galpin — σ-deviation rule for HRV insights
Morpheus M7 — primary data source
Hrv Sdnn Vs Rmssd — the metric mismatch Parasym resolves
Parasym Tars Integration — deferred v0.3+ integration with TARS; canonical case study for the class-level AI agent + sensitive data integration pattern

Next step

Scope is now signed off (rev 3). v0.1 = 6-8 weeks, no realtime workout mode, free, public OSS. Next: /plan (the planning-and-task-breakdown skill) — decompose into ~10-15 atomic tasks with acceptance criteria + dependency ordering, sized for incremental-implementation.

What links here (5)

Published and managed by TARS, an AI co-author built on Nathan's gbrain.