RURALOracle and MCP
The local MCP query and replacement-view polling subset, with unsupported Oracle operations.
12 October 2026 reconciliation (T-R106): the executable adapter configures exactly query and changes; the service harness checks that discovery pair and rejection of catalog, context, expand and transact. This is a source/evidence review, not a fresh runtime qualification; the T-R004 baseline records that native tests never launched because the build-guard lock write was denied.
rural mcp <profile-dir> <ordinary-collection> exposes query and process-local changes through newline JSON-RPC stdin/stdout. Start the sole service first. Launch configuration fixes branch 1 and one ordinary collection; the adapter never opens database files. It uses the trusted local owner credential rather than issuing a separately scoped service credential. Sources: local adapter and service harness.
Startup loads credentials once and verifies a one-record query through that session before accepting input. Store and session remain pinned; a service restart or disconnect requires restarting the adapter. No credential refresh or write dispatch occurs. Source: session preflight and dispatch.
Available tools versus the planned surface#
| Surface | Evidence-backed boundary |
|---|---|
query, changes | Configured and dispatched by local.zig; check_mcp_service.py checks discovery, CLI query/view parity and polling refusals. |
catalog, context, expand | W13 freeze specifies future schemas and service prerequisites; local.zig supplies no dispatch for them. |
transact | Disabled in local.zig; the freeze leaves it outside read-only Oracle scope. |
| Reusable six-name catalog | tools.zig defines names for generic dispatchers; server.zig uses available_tools for discovery and dispatch. Names alone do not establish executable availability. |
| Bounded Oracle context | context.zig owns facts and evidence strings with default limits of 32 facts, 16 handles and 64 KiB; the freeze explicitly says it is not an MCP route. |
The gateway guide still says the executable configures only query; that conflicts with the current adapter and harness. The query-only 28 September Unit A statement in Oracle evidence is explicitly historical there; Sprint 41 Units C/D/E record the later two-tool surface and W13 freeze, while retaining O01/V09 as partial.
query#
Arguments may contain only predicates, projection, sort and limit. Limit is a canonical decimal string from "1" to "200"; default is 200. At most 32 predicates and 64 projection fields are accepted. Predicates use operator (exists, eq, ne, lt, lte, gt, gte) and field; comparisons require a scalar value, while exists omits it. Sort accepts field and optional boolean descending. Collection, branch, operation, credentials and every unknown field are rejected even if they match launch configuration. Sources: argument builder, input schema and local argument tests.
Each successful query returns two MCP text content parts: (1) the validated service response including view evidence; (2) a plan_id meta object {"plan_id":"<32-hex>","contract":"mcp-changes-v1"} for process-local changes polling. Plans are capped (32) and die with the adapter process. Sources: query result construction and saved-plan table.
changes#
Arguments: required plan_id; optional cursor; optional limit ("1"…"200", default "50"). Re-executes the saved plan against the live service. The outcomes below follow the polling implementation and the overflow/deletion/cursor checks in the service harness.
Unchanged stable result/view hash and coverage ->
{"status":"unchanged"}only.Changed and current query-page row count ≤ poll limit ->
status: "changed"withentriesequal to the complete current matching page (replacement view, not a CDC delta),next_cursor, and coverage stamps. An empty matching set after deletions is a valid changed replacement (entries: []); it is still not a CDC tombstone log.Current query-page row count > poll limit -> error result; the poll cursor is not advanced (omitted rows must not become permanently unreachable through a truncated then-unchanged sequence). Caller may raise
limiton the sameplan_idor issue a newquery. Error text may say "resnapshot with query" as the safe recovery path.Unknown/stale plan or cursor, foreign
plan_idembedded in an otherwise well-formed cursor, and authority mismatch -> fail closed with non-sensitive text.next_cursoralways binds the saved plan identity (1:<plan_id>:…).
This is C11 process-local polling, not a durable feed and not Application /v1/changes. Sole-owner MCP does not implement a separate dynamic revoke API. Source: C11/W13 freeze.
The saved query's own limit remains in force when polling; raising the poll limit does not expand that query or prove complete collection coverage (buildQuery and changes).
Transport/session failures and invalid arguments return fixed, non-sensitive text content with isError: true. Framing is bounded to 64 KiB per line and 256 KiB buffering. Each request uses at most 16 MiB temporary allocator storage. Service exchanges use a ten-second timeout. Sources: adapter allocation, timeout and error paths and framing defaults.
Disabled tools and gateway admission/lifecycle errors instead return JSON-RPC errors; not every refusal is an MCP isError result (server.zig, harness).
Five-answer acceptance remains open#
The five-question test dispatches literal synthetic answers through Demo.call; P22 evidence limits that proof to protocol interoperability and transfer accounting, with no model, network, token counter or external source connected. The following acceptance mapping comes from the W13 freeze:
| Golden plan | Missing sole-owner evidence or definition |
|---|---|
island_signups_purchases | Authorized catalog/schema/metric discovery and exact scalar evaluation; the service surface is absent. |
compatible_conversion_per_million | Owner-defined formula, time window and compatible coverage; document query does not supply this metric. |
changes_around_deployment | The fixture deployment marker is not the saved-query replacement view supplied by local changes. |
projects_on_older_governance | Implemented context and an owner-defined governance version/age relation. |
revenue_after_refunds_by_asset | A real evidence-handle issuance/expansion path and authorized asset/scale sum; fixture totals alone are insufficient. |
The bundled owner question remains unresolved in the freeze. Oracle evidence retains O01/V09 as partial and leaves cost/latency/token claims outside its acceptance; the capability map likewise disclaims an accepted cold/warm/token/latency envelope. A matching current five-answer execution receipt and actual installed-client acceptance remain gaps in this reading set (backlog B04/ER01); no real Oracle walkthrough is established by the fixture test.
Related evidence and unresolved links#
B19 remains separate: collection-feed polling describes bounded canonical transaction/tombstone polling, while NEXT-GEN R02 requires authorized snapshot-to-durable-feed handover; neither changes local MCP saved-plan semantics (W13 freeze).
B02 and B15 remain separate: embedded reconciliation retains PF10/PF11 qualification/status uncertainty, and formal refinement retains unknown current receipt-repair closure; local MCP source presence closes neither gate.
B17/B35 remain separate: UBI consumer preparation leaves actual corpus, live import/reopen/restore and native-host acceptance open; synthetic Oracle answers establish none of those prerequisites.
B42 link/evidence gaps remain explicit: Sprint 41 evidence names raw W12 logs under
.local/change-checks/w12-mcp/, absent in this checkout; the collection-feed guide links to a missing pre-archivePROJECT-INTERNALcontract path. Matching raw logs and a public contract replacement remain gaps, without substituting private archive content or links (backlog reading boundary).