Scheduled Report Refresh

The dominant automated job on this MCP server: custom agents and scheduled runs re-pulling a known set of charts and summarizing changes.

The arc

1. Anchor on saved content. use_amp_dashboards action: 'get' (batch ≤3) for the report's chart list, or get_from_url if the run input is a link. Saved charts are the contract — do not rebuild definitions from scratch.
2. Re-pull data. get_amplitude_charts include: 'data', ≤3 chart IDs per call, at the report's granularity (daily for daily/weekly briefs). For each chart compare the current period against the trailing baseline the report uses.
3. Fill gaps ad hoc. query_amplitude_data (typed chart) only for numbers no saved chart covers (a new funnel step, a one-off slice).
4. Context for movement. use_amp_flags action: 'list_deployments' once — recent deploys are the first hypothesis for any metric movement.
5. Report. Delta table first (metric, current, baseline, % change), then notable movers with hypotheses. State explicitly when the current period is partial ("today is tracking at X vs Y full-day yesterday").

Expectations to set

• get_amplitude_charts include: 'data' on a saved chart returns exactly what the UI shows — trust it; don't re-validate every number with query_amplitude_data.
• Keep definition reads (include: 'definition' / 'typed') minimal on refresh runs — the dashboard chart list plus data is enough unless a chart looks misconfigured.
• If a chart errors or returns empty, report it as a broken report item rather than silently dropping it — a missing tile in a recurring report is itself a finding.