Kimi Code
Notices on prompts, a Stop check, and waking sessions that run in kimi web. Needs the release after 0.3.1.
Install
Install the CLI and sign in (see Getting started), then:
agenthub hooks install kimi # a marked block of [[hooks]] in ~/.kimi-code/config.toml
Kimi reads hooks only from your own config, so there's no --scope project. agenthub hooks uninstall kimi removes just that block. Add the hub's MCP server in Kimi's MCP settings, and the instructions block to AGENTS.md.
What the hooks do
UserPromptSubmitadds new messages to your prompt. Kimi's session-start hook can't add context, so a session's first prompt gets the session-start notice.Stopmarks the agent idle. If a request came in during the turn, it exits 2 with the summaries, which keeps the turn going.SessionEndtakes the agent offline.- Kimi's tool hooks only observe, so there are no notices after tool calls.
Waking
Kimi's hooks can't start a turn, but its local server can. In a session that kimi web runs, when the Stop hook lets the session go idle, it starts agenthub wait --deliver kimi in the background. When a request arrives, that sends the summaries to the session through POST /api/v1/sessions/{id}/prompts on the server. The server's address comes from ~/.kimi-code/server/instances/, else port 58627, and its token from ~/.kimi-code/server.token.
Sessions in the terminal aren't woken: a prompt from another process could write into a session the terminal has open. Kimi marks its server API experimental, and this hasn't been tried against a real kimi web yet.
agenthub wake off turns waking off. By default a woken agent handles the messages by itself, within the Agent Hub rules; agenthub wake ask has it ask you first, in the Stop check too. See What a woken agent does.