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. 4. **Resolve every `UNKNOWN` with the user** before writing code. If a needed capability is not supported, surface it and propose the fallback — do not silently work around it. 5. **Plan the executable proof**: a real sandbox/API call that exercises the path. The integration is not "done" until that call succeeds. 6. Before completion, have the **adversarial-verifier** agent challenge each "it works" claim against the cited ev