PDPP

Mcp Adapter

Created openspec/changes/unify-read-evidence-surface/specs/mcp-adapter/spec.mdView on GitHub →

MODIFIED Requirements

Requirement: Stdio MCP Adapter Uses Scoped PDPP Tokens

The MCP adapter SHALL expose model-controlled read tools over grant-scoped PDPP client tokens. The adapter SHALL NOT issue grants, request new authorization, run connectors, access reference owner-control endpoints, or broaden the grant scope.

Scenario: Agent follows omitted content through a model-controlled read

WHEN a tool result's visible content[] omits, truncates, summarizes, or references underlying record content THEN the result SHALL expose a working model-controlled continuation path, such as a fetch handle, read_record_field arguments, or a cursor AND the continuation SHALL remain inside the same grant and source/stream/field scope.

Scenario: Host hides structured content

WHEN an MCP host exposes only visible content[] to the model THEN search/query/fetch results SHALL still include enough bounded evidence and continuation instructions for the model to decide what to inspect next AND SHALL NOT require structuredContent as the only path to record identity or continuation.

Requirement: MCP Adapter Surfaces Resource Links Without Depending On Them

The MCP adapter MAY return MCP resource_link blocks and resource templates for host-controlled reads. The adapter SHALL NOT make resources/read the only path to additional model-needed content.

Scenario: Resource read is unavailable

WHEN an MCP host cannot read a returned pdpp://... resource URI THEN the same omitted or truncated content SHALL remain reachable through a model-controlled tool path when the active grant authorizes it.

Scenario: Small text inspection avoids materialization

WHEN an agent inspects a bounded text field or small record excerpt that fits within the configured inline read window THEN the MCP adapter SHALL provide a model-visible inline read path AND SHALL NOT require full file/export materialization as the ordinary path for that inspection.

Requirement: MCP Adapter Uses Shared Evidence Semantics

The MCP adapter SHALL render evidence cards, truncation descriptors, binary metadata, declared-role presentation, and continuation descriptors using the shared read/evidence semantics rather than MCP-only presentation rules.

Scenario: Manifest roles determine presentation

WHEN a record field has a manifest-authored display role THEN MCP evidence presentation MAY use that role AND SHALL NOT use connector-specific or field-name guessing to promote undeclared fields.

Scenario: Binary field appears in a record

WHEN a record contains a large binary, blob, image, attachment, or base64-like field THEN MCP visible output SHALL keep that field metadata-only by default AND SHALL expose an authorized resource/export continuation when available, using the grant-scoped blob read route such as GET /v1/blobs/:blob_id for stored blob content AND SHALL NOT inline the large binary body into model-visible text.