google-health-readonly

Google Health read-only workflow

Call the connector before stating facts about the connected user's Google Health data. Never replace a failed or empty tool result with an inferred measurement.

Translate the request into its metric, result granularity, closed-open range, query time zone, and data-source family. Select the operation with this routing table:

Use references/query-catalog.md for the exact type-to-method matrix, range limits, source families, and pagination details. Common mappings include:

• steps, distance, activity minutes, activity energy, heart-rate aggregates, and other day totals → the matching daily rollup type;
• resting heart rate → daily-resting-heart-rate;
• daily HRV → daily-heart-rate-variability; HRV samples → heart-rate-variability;
• daily blood oxygen → daily-oxygen-saturation; blood-oxygen samples → oxygen-saturation;
• sleep temperature → daily-sleep-temperature-derivations; sleep sessions → sleep;
• daily respiratory rate → daily-respiratory-rate; sleep respiratory summary → respiratory-rate-sleep-summary;
• daily, running, and sample VO2 max → daily-vo2-max, run-vo2-max, and vo2-max respectively.

Keep every range closed-open and as narrow as the request permits. Resolve relative dates in the user's intended time zone. Respect the tool's 14-day limits for heart-rate, active-minutes, calories-in-heart-rate-zone, and total-calories rollups; the 90-day general limit; pageSize<=25 for sleep/exercise reconcile; other page-size limits; and fixed partial-response masks. For daily rollup, retain pageSize at least the civil-day count (up to 50); split longer history into non-overlapping ranges for upstream compatibility. Daily rollup, physical rollup, reconcile, and IRN support continuation through pageToken. Follow a returned token only with every other input unchanged and only when the request needs another page. Any returned nextPageToken means the query is incomplete even when that page has no records; do not classify the query as empty until the required page walk reaches a response without a token.

Disambiguate completeness. “Cover every supported category” means the 34 non-IRN data types plus IRN, using the catalog's preferred projection for each type. “Return every raw record” additionally requires following every token-bearing query to completion and can be large; physical rollup is still used only when the user explicitly requests sub-day buckets. For ordinary analysis, select the daily, session, interval, or sample types that directly answer the question and use the requested granularity.

For scheduled summaries, save a self-contained prompt that names the MCP connection, the exact tool/data types, the intended query time zone and closed-open range, and the page limit. Set and verify the scheduler's own time zone separately; a time zone written only inside the data-query prompt does not control when the task runs. This repository's local Skill is documentation and evaluation material: a Developer Mode MCP connection receives the server instructions, tool descriptions, and schemas, but does not automatically import this Skill into web ChatGPT. Verify the real app association and keep the necessary query contract in the Scheduled prompt itself. Treat connector unavailable, OAuth invalid, permission denied, rate limited, upstream failure, and an actual empty result as different outcomes.