GRANT ... --mcp is unreachable: the module-security read is routed to MCP and aborts before the PED write path (#704)

[ynnckw]

Summary

#704 shipped AddEntityAccessRule over PED, and mxcli mcp capabilities reports it as available. But every GRANT … --mcp fails before reaching that code, because the role-validation read is routed to the MCP backend instead of to -p.

Version: v0.18.0 (2026-08-14), Windows, Studio Pro 11.12.2.

Reproduction

Studio Pro open, MCP server enabled at Preferences → AI → MCP Server.

mxcli --mcp http://localhost:7782/mcp -p app.mpr \
  -c "GRANT MyModule.MyRole ON MyModule.MyEntity (READ (SomeAttr));"

Actual:

Error: failed to read module security for MyModule: GetModuleSecurity: not supported
by the MCP backend; run without --mcp to author against a local .mpr

Expected: the rule is added, as #704 describes and as mcp capabilities reports.

The role exists and is in the same module as the entity. Unchanged with --engine legacy. --mcp-verbose prints no PED tool call at all, so the abort happens before the write path is entered.

Why this may be worth separating from #704

mxcli mcp capabilities --mcp http://localhost:7782/mcp -p app.mpr prints, on the same binary and in the same session:

✓ Entity access rules (GRANT) — GRANT a NEW entity access rule for an existing role — the rule lives on DomainModels$Entity.accessRules (the domain-model document PED authors), NOT the security document. ADD-only: PED refuses to remove DomainModels$AccessRule / DomainModels$MemberAccess, so REVOKE and replacing an existing rule in place are rejected (edit in Studio Pro). The referenced module role must already exist

A capability report that contradicts the very next command is more costly than a missing feature, because it is the surface an agent consults before deciding which write path to take. Either the module-security read should come from -p (the role is a by-name reference that Studio Pro resolves anyway), or the capability should report itself as unavailable so the report stays trustworthy.

Workaround

Calling PED directly over HTTP JSON-RPC — ped_read_document / ped_update_document with documentType: "DomainModels$DomainModel" — works, because it avoids the module-security read entirely.