fv-capabilities

fv-capabilities

Answer one question: what can this user actually do today, and what is the one thing they should connect next?

This assumes FullVision is already connected. Getting there is someone else's job: fv-login connects this machine, and fv-onboard covers the whole path from no account to first data.

Read shared/reading-fullvision-data.md before calling anything.

Steps

1. Probe FullVision. Call fullvision:list_views. If it fails on auth, stop and route them by what they are missing: an existing user who just needs this machine connected wants /fullvision:fv-login — one click in the browser; someone with no account, tracker or revenue data yet wants /fullvision:fv-onboard, which covers the whole path. Either way, do not ask them to go find a key.
2. Probe each other server by making its cheapest read call. Record connected / not connected / errored. Do not treat "not connected" as a failure; most users start with one.
3. Run the data-health precondition — call fullvision:query_view on view:health-identity-recon, view:health-checkout-coverage, view:health-event-coverage. Summarise in one line each. Full detail is fv-data-health's job.
4. Build the capability matrix below from the skill catalog, marking each skill available / read-only / unavailable per the degradation rule.
5. Name exactly one next step. Not a list. Rank by revenue unlocked per hour of setup effort, using the auth-burden column in docs/mcp-servers.md: Meta and Webflow are 1-click OAuth; Brevo is an API key; Google Ads needs a developer token (days); LinkedIn needs a dev app plus app review (weeks). Do not recommend LinkedIn to someone who has connected nothing.

Degradation rule

• Missing a read dependency ⇒ the skill is unavailable. Say so, name the server.
• Missing a write dependency ⇒ the skill is read-only: it runs the full analysis and emits a change-list the user applies by hand. This is a normal mode, not an error.

Output

Follow shared/report-format.md, with the verdict replaced by the matrix:

# fv-capabilities — <workspace> — <date>

**Usable today:** <n> of 10 analysis skills (<m> read-only)
**Data health:** <✅ | ⚠️ + what it biases>

| Skill | Status | Blocked on |
|---|---|---|
| fv-data-health | ✅ available | — |
| fv-cut-wasted-spend | ⚠️ read-only | google-ads is read-only by design |
| … | | |

## Connect next: <server>
<why this one, what it unlocks, how long it takes>

Things to state plainly, every run

• Google Ads is read-only. The official Google server exposes GAQL reads only. Every Google change is emitted as a reviewed change-list the user applies in the Ads UI. This is permanent for v1, not a setup problem the user can fix.
• fv-build-audience is read-only in v1 — audience activation is not on the FullVision MCP surface yet. It sizes, floor-checks and consent-gates the segment, then hands off to the FullVision app.
• Six MCP servers is a lot of tool schema. Recommend enabling deferred tool loading (ENABLE_TOOL_SEARCH) — tool-selection accuracy degrades measurably under heavy MCP load.

Refuse when

• FullVision itself is unreachable or unauthorised. Nothing downstream is meaningful; do not produce a partial matrix.