MCP Trust-Input Ingestion#

Decision. --input is not a required launch parameter, and the server no longer serves a silent baked-in example fallback. surface mcp runs input-agnostic by default: with no --input, it holds no default input, and each tool call supplies its own input (first-class multi-input for producers that emit many, evolving trust.bundle files). --input <file> remains an optional convenience default for the single-report case. A tool call with neither a per-call nor a startup input returns an honest error — "No trust input configured" — rather than quietly deriving demo data. The per-call input and adapter semantics are stated explicitly in the tool schemas (src/commands/mcp.ts).

This supersedes the earlier framing (keep --input as a defaulted single input): the original question was "why is input a launch parameter at all?" The honest answer is that it should not be a defaulted one — a hard default to the example bundle turned a forgotten --input into demo data masquerading as real trust state, which is worse than an explicit "unknown". Removing the default makes the multi-input case primary and keeps Surface honest about not having an input.

A current-aware resolver (a pointer/convention that resolves the active subject from a directory, glob, or moving "current" marker) is deferred, not rejected — the eventual answer for a producer whose "current" subject moves over a session without the consumer threading a path on every call. It is held until the Flow Agents workflow integration (flow-agents ADR 0011) concretely needs it, so the resolver is designed against a real access pattern rather than speculatively.

Why#

• A hard-defaulted --input (to the bundled example) is a footgun: a real producer that forgets it gets demo data presented as its trust state. Honest absence — an explicit error — is strictly better than a fake "known".
• The per-call input already exists and is exact: each call names its own input, so a producer with many bundles is fully served without a startup "primary" that may not exist.
• A directory/glob or resolver input is a larger surface (discovery semantics, "which is current", watch behavior) that should be shaped by the integration that needs it, not guessed. Shipping it early risks a mismatched convention.
• The related envelope-unwrap friction (issue #84) is solved independently by the veritas adapter preset, so ingestion ergonomics and adapter selection stay orthogonal.

Revisit when#

The Flow Agents workflow trust.bundle integration needs Surface's MCP to follow a changing "current" subject without per-call path threading. At that point, ratify the resolver shape (directory vs pointer vs convention) here and update status/rationale accordingly.