Decide what belongs in the AI context, as data

Situation

An AI client indexes what it finds: build outputs, the local state directory, a vendored binary, a data dump. The context fills with noise and, sometimes, with a secret. Every tool has its own ignore list, and none of them is the repository’s own statement of what matters.

What you run

• doctor: the scope policy exists and parses
• knowledge sources, knowledge nodes --scope shared: what the index discovers once the policy is applied
• over MCP: majordomus_scope reads the policy, majordomus_scope_classify answers for one path

Scenario

setup: installed-wired
given:
  - 'installed, and the two enforcements the policy declares are actually in place as hooks'
steps:
  - id: declared
    run: ['doctor']
    note: 'the scope file is present and well-formed; init seeded it'
    expect:
      exit: 0
      stdout_contains: ['^OK   layout      .ai/repo/scope.yaml', 'doctor: 0 failure']
  - id: sources
    run: ['knowledge', 'sources']
    note: 'the source classes and the files each discovers, after the scope policy is applied'
    expect:
      exit: 0
      stdout_contains: ['^policy +shared', '^scope +shared']
  - id: nodes
    run: ['knowledge', 'nodes', '--scope', 'shared']
    note: 'the shared objects only: nothing under .ai/local/ is a node'
    expect:
      exit: 0
      stdout_contains: ['shared']
      stdout_not_contains: ['operational']
then:
  - 'what is in and out of the context is one file in the layer, with a schema'
  - 'build outputs, local state, binaries and secrets are out by declaration, not by luck'
  - 'the MCP server classifies a path with the same data'

Outcome

What belongs in the AI context is declared once in the layer, checked, and applied by the shell tool and the server alike. A client that honours it indexes the repository, not the machine.