Skip to content
Docs
DocumentationTransparent & five mode

Transparent & five mode

v0.0.4Revision 587c8ae

Start with transparent mode to keep the full tool collection available. Choose five mode when you want your agent to request a compact, task-specific lineup.

Transparent is the default

Section titled “Transparent is the default”

Roster lists each backend tool using a namespaced ID, such as filesystem__read_text_file. It preserves tool contract metadata and forwards calls and results. Derived outcome recording still happens locally.

Transparent mode does not expose draft and call, does not automatically engage five mode at a context threshold, and does not provide the skill-instruction bridge in its tool list.

Terminal window
npx --yes @npmmo/roster@0.0.4 serve --transparent

This command starts a stdio process for an MCP client. It is not an interactive chat command.

Enable five mode

Section titled “Enable five mode”

For a client already connected through sync, set the existing Roster config’s mode field to "five", preserving all other fields. Then restart the client’s MCP connection.

For an explicit process launch, the supported flag is:

Terminal window
npx --yes @npmmo/roster@0.0.4 serve --five

A command-line mode flag overrides the config for that process. The synced launcher normally uses serve without a flag and reads the configured mode.

Draft the next step

Section titled “Draft the next step”

The client-visible tool list is stable: draft and call. This example shows the parameters of an MCP tools/call request:

{
  "name": "draft",
  "arguments": {
    "need": "Read the checkout component",
    "k": 5
  }
}

The response’s text content contains a compact JSON object with need, draft_id, starters, and a usage hint. Each starter includes an id, a kind, and a short description. Tool cards can include a trimmed input schema; skill cards explain that calling them returns instructions.

The default maximum is five, not a guarantee of exactly five results. k is configurable from 1 to 10. Retrieval can use rated fallback candidates when direct matches are sparse, so a draft is a suggestion, not a relevance guarantee.

The agent chooses the call

Section titled “The agent chooses the call”

Using an illustrative draft ID and path:

{
  "name": "call",
  "arguments": {
    "tool": "filesystem__read_text_file",
    "args": {
      "path": "/example/shop/src/checkout/CheckoutForm.tsx"
    },
    "draft_id": "d1"
  }
}

Use the returned draft_id to associate the call with the right need. An absent or unknown draft ID does not borrow another draft’s context; the call can still execute without that attribution or its alternate suggestions. Recent draft context is held in memory for the connection.

A tool invokes its configured backend. A skill returns instructions and resources. A successful response means that call returned without a detected failure; it does not mean the checkout is fixed.

Drafting is not a permission boundary. Known capabilities remain callable by ID, and local draft quarantine is not a universal execution block. Your client and agent must retain their own authorization policies.

Next: Bring skills into the Playbook, or understand failure suggestions.

Draft/call implementationCapability card contract