deep-groundinglisted

# Deep grounding (Workflow-driven) The heavyweight form of `ground-integration` for **L3/L4 external integrations**. It orchestrates a multi-agent Workflow so a large or critical API doc is read in parallel and **every capability claim is adversarially verified** before you trust it. This is the strongest defence against the worst failure mode — confidently assuming a capability the provider does not have. ## When to use - **Explicit invocation only**, **L3/L4 only.** For L0–L2 or a quick check, use `ground-integration` (single-agent) — say so and stop. - This skill **drives the `Workflow` tool** (multi-agent, ~15× tokens). Invoking it is the opt-in. - **If `Workflow` is unavailable** (plan/version), fall back to `ground-integration`'s single-agent flow and tell the user — never hard-fail. ## How to run 1. Establish the provider / product / API version, the **list of capabilities** the task needs (one row per capability), and the doc **sources** (URLs). 2. Invoke the `Workflow` tool with `name: "groundwork:deep-grounding-run"` and `args: { capabilities: [...], sources: "<official doc URLs>" }`. The shipped script (a) fans out one `grounded-researcher` per capability — each returns a cited finding; (b) runs an adversarial panel (2 `adversarial-verifier` skeptics, distinct lenses, refute-by-default) per row; (c) synthesizes the matrix. Do **not** rewrite it inline. It refuses to run without `capabilities`, and logs the fan-out size so the cost is visi