Ghealth Agent Skill

ghealth

CLI for the Google Health API v4. 40 verified data types.

Prerequisites: See ../ghealth-shared/SKILL.md for auth, setup, global flags.

Choosing the right operation

| Goal | Operation | Example |

|------|-----------|---------|

| Daily totals (steps, distance, calories) | daily-rollup | ghealth data steps daily-rollup --from 2026-03-22 --to 2026-03-29 |

| Individual readings (HR, weight, SpO2) | list | ghealth data heart-rate list --from today --limit 20 |

| Sessions (exercise, sleep) | list | ghealth data exercise list --from 2026-03-01 |

| Daily summaries (resting HR, HRV, resp rate) | list | ghealth data daily-resting-heart-rate list --from 2026-03-01 |

| Merged multi-source data | reconcile | ghealth data weight reconcile --from 2026-01-01 |

Why this matters: steps list returns minute-level intervals *without counts*. Use daily-rollup to get actual step totals (countSum). Same for distance (millimetersSum) and floors.

Types at a glance

Run ghealth schema types for the live version. Quick reference:

Use daily-rollup for totals:

• steps → countSum per day
• distance → millimetersSum per day
• total-calories → kcalSum per day (rollup-only)
• floors → countSum (rollup-only)
• active-minutes → (rollup-only)
• swim-lengths-data → strokeCountSum per day
• calories-in-heart-rate-zone → caloriesInHeartRateZones per day (rollup-only)

Use list for readings:

heart-rate, weight (writable), body-fat (writable), height (writable), oxygen-saturation, heart-rate-variability, altitude, vo2-max, active-zone-minutes, activity-level, basal-energy-burned, active-energy-burned, blood-glucose, core-body-temperature, respiratory-rate-sleep-summary, run-vo2-max, sedentary-period, swim-lengths-data, hydration-log

Use list for sessions:

• exercise (writable) — includes type, duration, calories, HR summary, notes
• sleep (writable) — includes summary by default. Add --detail for per-stage breakdown.

Cardiac (dedicated scopes, list-only):

• electrocardiogram — waveform samples + rhythm classification. Requires ecg.readonly.
• irregular-rhythm-notification — alert windows. Requires irn.readonly.

Nutrition:

• nutrition-log — logged food entries with nutrient/energy breakdown (list, get, rollup, daily-rollup, reconcile)
• food, food-measurement-unit — reference catalogs (list, get only). No time filter — --from/--to are ignored.

Daily summaries (one value per day, filter by date):

daily-resting-heart-rate, daily-heart-rate-variability, daily-oxygen-saturation, daily-respiratory-rate, daily-vo2-max, daily-sleep-temperature-derivations

Get a single point by ID: get --id <id> is supported on exercise, sleep, weight, body-fat, height, hydration-log, nutrition-log, blood-glucose, core-body-temperature, food, food-measurement-unit.

Patterns the CLI can't tell you

These require judgment that --help and schema don't provide.

Get the user's timezone before querying date-sensitive data:

ghealth user settings get   # → timeZone: "Europe/London", utcOffset: "3600s"
# Then use --from/--to with the correct local dates

This reports the account timezone for information only — to have the CLI resolve dates in that zone, set it explicitly with ghealth config set timezone <IANA zone>.

Sleep/exercise page size is capped at 25 per request (auto-paginated by CLI):

ghealth data sleep list --limit 5    # CLI handles pagination internally

Paging through large list results. list returns up to --limit rows (default 500). When more exist, the response carries a nextPageToken and a hint. Pass it back with --page-token to fetch the next page — it resumes exactly where the last page ended, no rows skipped or repeated:

ghealth data heart-rate list --from 2026-06-15 --limit 500
#   → {"dataPoints":[…500…], "nextPageToken":"ABC", "_hints":[…]}
ghealth data heart-rate list --from 2026-06-15 --limit 500 --page-token ABC
#   → next 500 rows

Correlate heart rate with exercise sessions:

# 1. Get exercise time window
ghealth data exercise list --from today --limit 1
#    → start: "2026-03-29T14:18:32+01:00", end: "2026-03-29T14:39:14+01:00"
# 2. Query HR for that window using --filter (raw API syntax, UTC required)
ghealth data heart-rate list --filter 'heart_rate.sample_time.physical_time >= "2026-03-29T13:18:32Z" AND heart_rate.sample_time.physical_time < "2026-03-29T13:40:00Z"'

Exporting data for analysis

Use -o <file> to write data to a file. When -o is set, stdout shows only a summary with the column schema — not the data itself. This means you can fetch data and immediately write analysis code using the column names from stdout, without reading the file.

ghealth data steps daily-rollup --from 2026-03-24 --to 2026-03-30 --format csv -o steps.csv

What stdout shows (this is all the agent sees):

Wrote 6 rows to steps.csv

Columns: countSum, date
Preview:
countSum,date
4062,2026-03-29
9122,2026-03-28
2469,2026-03-27

What the file contains (full CSV, not printed to stdout):

countSum,date
4062,2026-03-29
9122,2026-03-28
2469,2026-03-27
6541,2026-03-26
4025,2026-03-25
3995,2026-03-24

The agent now knows the columns are countSum and date, and can write pd.read_csv("steps.csv") without ever reading the file.

Do not pipe to file — use -o instead. Piping (> file.csv) sends the full data to the file but prints nothing to stdout, so the agent has no column schema and must read the file to learn the structure.

More examples:

# Sleep — nested stageMinutes auto-flattened to stageMinutes.AWAKE, stageMinutes.DEEP, etc.
ghealth data sleep list --from 2026-03-01 --format csv -o sleep.csv

# Exercise — metricsSummary.caloriesKcal, metricsSummary.averageHeartRateBeatsPerMinute, etc.
ghealth data exercise list --from 2026-03-01 --format csv -o exercise.csv

# Heart rate — 500 readings straight to file
ghealth data heart-rate list --from today --limit 500 --format csv -o hr.csv

Exercise time series (GPS/heart-rate track) → CSV. export-tcx --as csv flattens the TCX track to one row per trackpoint — pd.read_csv it directly instead of parsing TCX XML:

# Find the exercise id first, then export its track
ghealth data exercise list --from 2026-06-01 --limit 10
ghealth data exercise export-tcx --id <id> --output ride.csv --as csv   # or --output - for stdout

Columns (fixed, stable for dataframes): time, activity, lap, sport, latitude_deg, longitude_deg, altitude_m, distance_m, heart_rate_bpm, cadence_rpm, speed_mps, watts. Absent sensors are empty cells (NaN in pandas), never zeros. distance_m is cumulative. 0 rows = indoor/no-sensor activity (Google emits no track for those) — the session summary and workout notes come from data exercise list, not the track export.

Writing data

Writable types: exercise, sleep, weight, body-fat, height. Writes are async (API returns Operation object).

Discover the correct payload format by inspecting a real response with --raw:

ghealth data weight list --raw --limit 1
# Use the response structure as a template for your create payload

Write operations use create, update --id <id> [--update-mask fields], delete --ids <ids>.

Gotchas

• Missing days are NOT zeros (altitude, distance, floors, steps, total-calories): a date absent from rollup output means the device wasn't worn / didn't sync — NOT zero. countSum: "0" is a true zero (worn, no activity). Never coalesce missing buckets to 0 or average over absent days as zeros — that silently deflates weekly/monthly stats
• String vs number values follow protobuf JSON encoding: int64 fields (beatsPerMinute, countSum, minutesAsleep) are strings; int32/double fields (weightGrams, caloriesKcal, percentage) are numbers
• --filter raw syntax: only >= and < comparators. Civil time fields (no Z): interval types use {type}.interval.civil_start_time, sleep uses sleep.interval.civil_end_time (only end-time is filterable), daily types use {type}.date. Physical time fields (with Z): sample types use {type}.sample_time.physical_time
• Write operations are asynchronous — the API returns an Operation object, not the created/updated data. Use list to verify persistence
• Body fat delete returns HTTP 500 — this is an API bug
• Height update returns HTTP 400 ("updateMask not recognized") — API bug; use create + delete as a workaround
• daily-rollup aggregates by civil/local day (1-day windows; override with --window-days N). rollup aggregates by physical time (--window-size, default 86400s); bare --from/--to dates anchor at midnight in the configured timezone (ghealth config set timezone <IANA zone>), falling back to machine-local time when unset. For local-day totals use daily-rollup. Both send their window size explicitly — the API rejects requests that omit it