managed-provider-clilisted

# Managed Provider CLI Use this skill when a task touches how Longhouse starts, observes, steers, or repairs provider CLIs. The core rule: Longhouse manages session control, not provider binary distribution. See [`ARCHITECTURE.md`'s Session modes section](../../ARCHITECTURE.md#session-modes) for the canonical definitions of Shadow, Helm, Console, and managed/unmanaged sessions. ## Mental Model - **Provider CLI**: an upstream executable the user installs separately, such as `claude`, `codex`, or `agy`. - **Machine Agent**: `longhouse-engine`; ships local events, owns runtime hooks/state, and may run provider-specific bridge processes. - **Runtime Host**: FastAPI/web/database product runtime. It stores session state and exposes `/api/agents/*`. ## Provider Paths ### Claude - `longhouse claude` relies on Claude's native channel/MCP/stdin control path. - Claude channel send, interrupt, and active-turn steer are first-class local control operations. Steer uses `claude-channel send --meta intent=steer` and Runtime Host must gate explicit `intent=steer` on a fresh active runtime phase; idle channel injection is not steer. - Helm creation is terminal-originated through `longhouse claude`; the Machine Agent has no remote Helm-originating launch command. - Longhouse's channel is a private MCP server, not an Anthropic allowlisted channel plugin. Because it arrives via `--mcp-config`, Claude's startup banner always prints `server:longhouse-channel · no MCP server confi