This is things you found annoying

[radiosilence]

Yes — authenticated as radiosilence, so it's your own repo. Three open issues, no overlap with what I hit. Filing two.

Creds are read-only — 403 Resource not accessible by integration on both issue creation and even adding a reaction. So the GitHub connector's token is scoped to reads only. You'd need to widen it in the connector settings if you want me filing directly; happy to if you do.

Here are the two, ready to paste.

1. CardDAV credential state isn't discoverable before a contacts query fails

contacts failed with Username not configured. Set FASTMAIL_USERNAME env var. in a session where emails and sendEmail both worked. Reasonable — CardDAV needs username + app password where JMAP only needs the token — but there's no way to discover it except by running the query.

Session is otherwise the best-designed part of the schema for agent use; the INVALID_CREDENTIALS vs UNREACHABLE split is exactly the actionable distinction. But capabilities only reflects what the JMAP server advertises, so a CardDAV gap is invisible to it. Consequence: an agent planning look-up-contact-then-compose can't check reachability up front. It commits, fails mid-task, and recovers by scraping the address out of emails(filter: {or: [{to:…},{from:…}]}) — which works, but isn't the intended path.

Suggest a carddavConfigured: Boolean! on Session, so it's answerable in the same call that establishes connection state rather than only by attempting the operation.

2. schema_sdl returns the full SDL with no way to narrow it

Ran three query shapes this session (emails filter, contacts, sendEmail preview/confirm) and paid for the entire SDL to do it — twice, because the connector dropped and I had to re-establish. The schema is large, and the docstrings are a big part of why it's large; they're genuinely worth having, which is the tension.

Options, roughly in order of effort: a types: [String!] argument on schema_sdl to fetch a subset; or put two or three canonical query shapes directly in the MCP tool description, which would cover most sessions without any SDL fetch at all. The EmailFilter tree and the PREVIEW→CONFIRM mutation flow are the two things a consumer actually needs to see examples of.

On kreuzberg — that wasn't the complaint. Charging for document parsing is obviously right, and annotating it in the docstring is the best thing in the schema. Attachment.text saying "this is the expensive field" and totalCount being free on Thread.emails but not on a query is cost information at the point of selection, which almost nothing else does. My gripe was purely the size of the SDL payload, which is a separate axis.

Right. That's a good hour's work you've done at what must be gone midnight on seven nights of no sleep. The issues will paste just as well tomorrow.

[radiosilence]

Both halves are in #63, plus radiosilence/mcp-gateway#45 for the deployment side.

CardDAV state — Session.carddavConfigured, as suggested. It reports both credentials being present rather than correct, and answers independently of status, since credentials are resolved per request and stay knowable when the token is dead. Worth noting what the suggestion didn't cover: the credentials had to become per-request headers first (X-Fastmail-Username, X-Fastmail-App-Password, falling back to local config like the token does). Read from local config only, a hosted deployment would have reported false forever — the exact gap the field exists to report, made permanent. The gateway PR declares the fields so a deployed instance receives them.

The SDL payload — took the second option rather than the first, then did the first as well. The cheap fix turned out to be the better one: the graphql tool description now carries a slimmed schema for everyday mail (queries, the EmailFilter tree, common Email fields, connection shape, PREVIEW→CONFIRM) plus worked examples, and names what it doesn't cover. So the common path costs no fetch at all, not even a narrowed one. schema_sdl also takes types: [String!] for the rest, which keeps that second hop small.

On the docstrings: agreed they were never the problem — nothing has been trimmed. The full SDL is unchanged and still the default when types is omitted.

The examples are scraped out of the published tool descriptions and executed against the real schema by the test suite, and every field name in the inlined sketch has to exist in it. That guard earned itself immediately: the first draft invented a nonce field and list-valued recipients.