Imported from tiangong-ai/agent-skills (gdelt-events-fetch/SKILL.md). Install upstream with npx skills add tiangong-ai/agent-skills --skill gdelt-events-fetch. Copyright stays with the author.

GDELT Events Fetch

Use the CLI-owned gdelt.events capability. This Skill supplies intent routing and result-use boundaries only; the CLI owns source discovery, input/output schemas, HTTP and archive handling, limits, validation, and receipts.

Before running

1. Read references/tiangong-data-requirement.json.
2. Use the caller- or workspace-resolved stable CLI. The requirement declares compatible capability and operation contract majors; it does not select a package build.
3. Run data describe with that same CLI. Continue only when the capability ID and required contract majors match, and copy the exact current capability/operation versions from that response into the run request.

tiangong-ai data describe gdelt.events --json

Use the returned Discovery Metadata to confirm current source coverage, freshness, restrictions, provides, and doesNotProvide. Do not substitute facts remembered from an older Skill revision.

Prepare the request

Build a tiangong.data.run-request.v1 envelope. Replace the two version placeholders with the exact versions from the same data describe response. This example selects a bounded range of source snapshots:

{
  "schemaVersion": "tiangong.data.run-request.v1",
  "capabilityId": "gdelt.events",
  "capabilityVersion": "<describe.manifest.capabilityVersion>",
  "operationId": "fetch",
  "operationVersion": "<describe.manifest.operations[0].operationVersion>",
  "input": {
    "mode": "range",
    "startDateTime": "2026-03-01T12:00:00Z",
    "endDateTime": "2026-03-01T12:45:00Z",
    "maxFiles": 4
  }
}

Use the operation input schema returned by the same data describe response when choosing latest or range. Range bounds need not align to a snapshot: the CLI selects the first published 15-minute timestamp at or after the lower bound and stops at maxFiles, even when the requested window is larger. Do not widen the user's range or increase a safety limit without approval.

Run

tiangong-ai data run gdelt.events fetch \
  --input /absolute/path/to/request.json --json

The command emits a tiangong.data.run-result.v1 envelope. Preserve its contract, warnings, errors, and receipt with data when handing the result to another workflow.

Result boundaries

• Treat rows as machine-coded event records, not verified real-world events, unique incidents, legal findings, or causal evidence.
• A source URL is lineage metadata, not an article body or proof that every coded field is correct. Do not claim full-text acquisition.
• Preserve GDELT identifiers and source timestamps when deduplicating or joining downstream; do not collapse rows merely because labels look similar.
• The capability returns normalized in-memory rows and execution metadata. It does not create a durable ZIP mirror, expose the master file list, or perform polling and incremental state management.
• Surface partial, truncation warnings, archive-validation failures, and empty results. Never reinterpret them as complete absence of events.
• Preserve file SHA-256 values and validation issues. Invalid UTF-8 or wrong-column rows are omitted locally while valid rows from the same verified ZIP remain usable; report that omission explicitly.
• Use the dedicated GKG or Mentions Skill for their row types; this Skill must not invoke or combine other feeds automatically.
• Cross-source comparison, persistence, scheduling, and research evidence admission belong to the caller or Auto Research.

Reference

• references/tiangong-data-requirement.json: stable capability requirement; it is not a package lock.