Skip to content

Automation with Kimi Code CLI

Running little-loops orchestration on the Kimi Code CLI via KimiRunner (scripts/little_loops/host_runner.py, FEAT-2914). Flag behavior on this page was machine-verified on kimi 0.30.0 by the FEAT-2911 spike (thoughts/research/kimi-cli-surface.md).

Running orchestration CLIs

All orchestration tools (ll-auto, ll-parallel, ll-action, ll-harness, ll-loop, ll-sprint, plus the FSM evaluators and handoff) support kimi as the backing host CLI. Set LL_HOST_CLI=kimi-code to select it explicitly:

LL_HOST_CLI=kimi-code ll-auto
LL_HOST_CLI=kimi-code ll-parallel --workers 3
LL_HOST_CLI=kimi-code ll-action manage-issue --json
LL_HOST_CLI=kimi-code ll-loop run my-loop
LL_HOST_CLI=kimi-code ll-sprint run v2-launch

Auto-detection

You do not need to set LL_HOST_CLI explicitly if kimi is on PATH. resolve_host() in host_runner.py probes available binaries and selects KimiRunner when kimi is found and no override is set.

Detection order (first match wins):

  1. LL_HOST_CLI environment variable
  2. LL_HOOK_HOST environment variable
  3. Binary probe: claude → codex → pi → gemini → omp → kimi → error

You can also set the host permanently in .ll/ll-config.json (or .kimi-code/ll-config.json):

{
  "orchestration": {
    "host_cli": "kimi-code"
  }
}

What the runner emits

KimiRunner.build_streaming assembles argv as:

kimi --output-format stream-json [--continue] [--agent <name>] [--add-dir <dir>] [--model <alias>] -p "<prompt>"

with LL_NON_INTERACTIVE=1 and DANGEROUSLY_SKIP_PERMISSIONS=1 in the child environment. Kimi-specific facts:

  • Flags first, prompt last. Kimi's argument parser treats a bare positional after options as a subcommand, so -p <prompt> always terminates argv.
  • No permission flags, ever. kimi -p runs under the auto permission policy implicitly, and --yolo / --auto / --plan are rejected in combination with -p (verified: error: Cannot combine --prompt with --yolo.). The runner never emits one.
  • Resume maps to --continue (most recent session in the cwd). The terminal meta event in the stream carries the session_id, which the runner layer can use for resume bookkeeping.
  • Model maps to --model <alias>; the alias must exist in the [models] table of kimi's config.toml.
  • --add-dir (from workspace_root) is an additive workspace dir, not a sandbox — it widens rather than confines access.

Blocking JSON consumers

Kimi has no single-blob --output-format json mode — text and stream-json are the only formats. build_blocking_json therefore also streams, and blocking consumers take the final role:"assistant" content event before the terminal meta line (the same consume-the-final-event contract as Codex). There is no type:"result" summary event like Claude's.

Detached runs and print_background_mode=steer

In print mode, kimi defaults to print_background_mode = "steer": the process stays alive after the main turn while background tasks (including background subagents) are pending, feeding completions back as synthetic user messages. It is bounded by print_wait_ceiling_s / print_max_turns, which are effectively unbounded by default. Callers of build_detached must not expect the process to exit at the first final answer when the model backgrounds work.

--agent + resume conflict

Agent selection is native (--agent <name>) and works with -p — but kimi rejects --agent in combination with --continue / --session (the agent is bound at session creation). When ll's orchestration layer requests a persona on a resume, KimiRunner drops the agent and emits a CapabilityNotSupported warning rather than failing the run.

Conformance

The generic host-parametrized conformance harness covers kimi-code; the golden paths pass:

pytest -m conformance --conformance-host kimi-code scripts/tests/conformance/
# 4 passed

(Point pytest at the conformance directory itself — the --conformance-host option is registered by scripts/tests/conformance/conftest.py, which is not loaded for broader path args.)

See CONFORMANCE for the harness contract.

Current limitations

No token reporting

Kimi 0.30.0 emits no token-usage events anywhere in the stream-json output (usage is visible in the TUI /usage only). Loops run under kimi produce no usage.jsonl file and no per-state cost table in ll-loop run output.

No structured-output flag

Kimi has no --json-schema flag and no single-blob JSON mode. A json_schema passed to build_blocking_json is dropped with a CapabilityNotSupported warning, and the FSM evaluators gate on HostCapabilities.structured_output ("unsupported" for kimi) and fall back to prompt-and-parse — with the BUG-2626 <StructuredOutput> tag recovery — instead of appending an inline schema flag.

No tool allowlist

Kimi has no --tools CLI flag. Tool policy lives in agent files (tools / disallowedTools frontmatter) or the global [tools] table in kimi's config.toml. A tools list passed to the runner is dropped with a CapabilityNotSupported warning.

See also