kimi-cli-runtimelisted

# Kimi Runtime (subagent) Use only from `kimi:kimi-rescue` (or when a command explicitly says so). This is the host-side launcher for the **Kimi Code subagent** over ACP. ## Happy path ```bash node "${CLAUDE_PLUGIN_ROOT}/scripts/kimi-companion.mjs" task --mode yolo -- "<user task>" ``` With screenshot / recording: ```bash node "${CLAUDE_PLUGIN_ROOT}/scripts/kimi-companion.mjs" task --mode yolo --image "<abs-or-workspace-path>" -- "<user task>" ``` ## Options (routing only) | Flag | Purpose | | --- | --- | | `--mode yolo\|plan\|auto\|default` | Default **yolo** for implement/rescue; `plan` = plan-only | | `--image` / `--video` / `--media` | Multimodal paths (resolve against workspace cwd) | | `--goal` | Frame as Kimi Goal objective | | `--resume` / `--session id` | Same subagent thread | | `--fresh` | Force new session | | `--cwd` | Explicit workspace root for Kimi edits | | `--git` / `--base ref` | Attach raw git context | | `--model` / `--thinking` | From setup catalog if user asked | Advanced (`--background`, job poll): only when the user asks to detach. Prefer foreground. ## Rules - One `task` invocation; return stdout **verbatim**. - No invented system prompts; Kimi keeps tools/models/skills. - Prefer media flags for screenshots/recordings over describing pixels in prose. - If setup fails, tell the user to run `/kimi:setup` and show the companion error as-is. ## Acceptance / empty handoff (read stdout + exit code) Companion turn policy (Mode A/B) runs **inside