ground-integrationlisted

# Ground an external integration External integrations are where guessing does the most damage. Do this **before** writing integration code. Boost does not cover third-party APIs — go to their official documentation. For an **L3/L4** integration with a large or critical API, escalate to **`deep-grounding`** — a multi-agent Workflow that builds an adversarially-verified capability matrix (each row refuted-or- survived). This single-agent flow stays the default for L0–L2. ## Steps 1. **Identify the exact provider, product, and API version** in use (or to be used). 2. **Fetch the official documentation** with WebFetch/WebSearch (and the provider's own MCP if one exists). For a deep sweep, spawn the **grounded-researcher** agent. 3. **Build a capability matrix** — one row per feature the task needs: | Feature needed | Supported? | Evidence (URL + quote / endpoint) | Fallback if not | |---|---|---|---| Mark anything you cannot confirm as `UNKNOWN — must verify`. Never assume a capability exists because similar providers have it. Three rows are **mandatory** for every provider, because the answers decide whether the integration can be made safe at all — and none of them is guessable: - does the provider accept an **idempotency key** on an outgoing call (header name, scope, TTL)? - is webhook/callback delivery **at-least-once or at-most-once**, and does it retry? - does an event carry a **stable id** that is identical across redeliveries? A