Module agents 

Agent roster, MCP tool loop

Functions

agent_backend_kind

Backend kind of a roster agent: "local" | "remote" (unknown/empty slug → "local").

agent_remote_connection_test

Verify that a configured remote-MCP agent can answer the non-generative tools/list handshake. This never sends a chat prompt.

agent_roster_add_remote

Convenience: create/update a REMOTE-MCP agent from primitives so the UI never hand-builds the backend enum. transport_kind ∈ "tcp" | "http" | "stdio"; endpoint is host:port / a URL / a command line respectively.

agent_roster_get

agent_roster_list

agent_roster_remove

agent_roster_upsert

agent_runtime_status

Truthful runtime projection for one roster agent. Definitions are durable; this response describes only the currently effective local-model residency and recent decode measurement, so callers do not mistake “configured” for “loaded on GPU”.

agent_set_allowed_mcp_tools

Convenience: set allowed_mcp_tools on a roster agent (persist via upsert).

ingest_chat_cml

Store a chat turn’s inline CML context (#project: / #topic: / #task: / [[concept]]) into the person’s inforg (their private library). No-op if the message has no tags. Returns concepts stored.

mcp_call_tool_gated

Principal-gated MCP tool call. principal_permitted = false → Err without MCP. Tool must be on the agent’s allowed_mcp_tools (or *); empty allowlist denies all.

mcp_ensure_safe_tool_allowlist

If allowlist is empty, seed list_capabilities + computer_vision (dogfood-safe). Does not Permit any call — only widens the roster allowlist.

mcp_list_local_tools

List local in-process MCP tools (tools/list) for Talk / allowlist UI.

provider_credential_remove

Remove a user-owned connection credential from the operating-system keychain.

provider_credential_store

Store a bearer credential for a user-owned connection in the operating system keychain. The secret is intentionally write-only at this API boundary.

run_remote_agent_turn

Run one turn against a REMOTE-MCP agent from the roster (native-only). Privacy-gated via the job router (Classified/sanctuary never leaves the device), then issues an MCP tools/call to the provider and appends the reply as an agent message. Returns a ChatInferenceResult-shaped JSON.