feat(kimi): support kimi-code (Moonshot's new Node CLI, successor of kimi-cli)

[whyihaveyou]

Context

Moonshot has released kimi-code, a ground-up Node.js rewrite of the Python kimi-cli, and it is the designated successor going forward:

• MoonshotAI/kimi-code repo created 2026-05-22, first release 0.2.0 on 2026-05-26, currently at 0.26.0 (2026-07-16)
• The Python kimi-cli README now states: "Kimi CLI is evolving into Kimi Code CLI … This project will be gradually wound down." (it is still being released in the meantime — 1.49.0 on 2026-07-16)
• cc-connect's kimi agent (added in feat(agent): add Kimi CLI support #616) targets the Python kimi-cli dialect. Verified: latest kimi-cli 1.49.0 still supports --print / --work-dir / --quiet / --resume / --output-format, so the current agent and current kimi-cli still match.

When users install kimi-code (which the official installer now does by default), the agent breaks in three ways. Related: #1476, #806, #1253, #1059, PR #1483.

Incompatibilities (all reproduced against kimi-code 0.26.0)

1. Unsupported flags (covers #1476, #806/#1253 for kimi-code)

cc-connect builds: kimi --print --output-format stream-json [--quiet] [--plan] [--resume SID] [--model M] --work-dir DIR --prompt ...

kimi-code 0.26.0 rejects / lacks:

cc-connect flag	kimi-code status	what works instead
--print	❌ error: unknown option '--print'	drop it — --prompt alone enters non-interactive print mode
--work-dir DIR	❌ unknown option	drop it — setting the process cwd is sufficient (cc-connect already sets cmd.Dir)
--quiet	❌ unknown option	drop it
--resume SID	❌ unknown option	-r SID
--output-format stream-json, --model, --plan, --prompt	✅ supported	—

2. stream-json schema change: assistant content is a plain string (the #1059 symptom, on kimi-code)

kimi-code emits:

{"role":"assistant","content":"17 × 23 = 391"}

but handleAssistant only handles content as []any of {type:"text",...} blocks, so every reply arrives empty. Rewriting the line to content:[{"type":"text","text":"..."}] fixes rendering.

3. Permission-mode flags conflict (worth knowing before "fixing" with --yolo)

--yolo / --auto are mutually exclusive with --prompt in kimi-code (error: Cannot combine --prompt with --yolo). Bare --prompt mode auto-approves tool calls (implicit-yolo semantics), which matches the agent's documented "default" mode expectation — so no extra flag is needed.

What already works unchanged: tool_calls / tool role events match the parser, and the session-resume hint ("To resume this session: kimi -r session_...") matches extractResumeSessionID as-is.

Working workaround (in production on my setup)

A small shim between cc-connect and kimi-code, selected via the existing cmd option — no cc-connect changes needed:

[projects.agent.options]
cmd = "/path/to/kimi-cc-shim"
model = "kimi-code/k3"

#!/usr/bin/env python3
"""kimi-cc-shim: translate cc-connect's kimi-cli dialect to kimi-code v0.26+.
Args: drop --print/--quiet/--work-dir; --resume SID -> -r SID; pass the rest.
Stdout: rewrite assistant string content -> [{"type":"text","text":...}].
stderr / exit code / signals pass through."""
import json, os, signal, subprocess, sys

REAL_KIMI = os.path.join(os.path.dirname(os.path.abspath(__file__)), "kimi")

def translate_args(argv):
    out, i = [], 0
    while i < len(argv):
        a = argv[i]
        if a in ("--print", "--quiet"):
            i += 1
        elif a == "--work-dir":
            i += 2
        elif a.startswith("--work-dir="):
            i += 1
        elif a == "--resume":
            out += ["-r", argv[i + 1]]; i += 2
        elif a.startswith("--resume="):
            out += ["-r", a.split("=", 1)[1]]; i += 1
        else:
            out.append(a); i += 1
    return out

def main():
    child = subprocess.Popen([REAL_KIMI] + translate_args(sys.argv[1:]),
                             stdout=subprocess.PIPE, stderr=None, text=True, bufsize=1)
    def forward(sig, _):
        try: child.send_signal(sig)
        except Exception: pass
    signal.signal(signal.SIGTERM, forward); signal.signal(signal.SIGINT, forward)
    for line in child.stdout:
        out_line = line
        try:
            msg = json.loads(line)
            if isinstance(msg, dict) and msg.get("role") == "assistant" \
                    and isinstance(msg.get("content"), str):
                msg["content"] = [{"type": "text", "text": msg["content"]}]
                out_line = json.dumps(msg, ensure_ascii=False) + "\n"
        except (json.JSONDecodeError, ValueError):
            pass
        sys.stdout.write(out_line); sys.stdout.flush()
    child.stdout.close()
    sys.exit(child.wait())

if __name__ == "__main__":
    main()

Verified end-to-end via cc-connect → Feishu: prompts, tool calls, session resume all work.

Suggestion for native support

Probe the CLI flavor once at agent init (e.g. parse kimi --help for --print, or kimi --version output format) and branch both the flag set and the content parser. PR #1483's probe-gating approach for --work-dir generalizes naturally to this. Happy to test any PR against kimi-code 0.26.x on Feishu.

[github-actions]

👋 Hi @whyihaveyou! Thanks for opening your first issue here!

We'll review it as soon as possible. While waiting, please make sure:

• You've provided enough details about the issue
• Any error messages are included in full
• For feature requests, describe your use case

Thanks for your feedback!

[chenhg5]

感谢 @whyihaveyou 的高质量首 feature + 完整 root-cause analysis (4 incompatibilities + 1 working shim + 1 native support 建议) + willingness to test PR.

PM 立场: feature / has-pr-ready / waiting-on-reporter-self-implement (跟 #1543 n-wen feishu topic workspace isolation + #1556 Codex workspace dependency dynamic tool bridge 同 reporter-self-implement 路径).

PM 评估

• 你的 root-cause 分析很扎实: 4 incompatibilities (--print/--work-dir/--quiet/--resume + stream-json content schema change + permission-mode flags conflict)
• 你的 working workaround shim (Python) 已经在 production 跑 cc-connect → Feishu 验证通过
• native support 建议 (probe CLI flavor at init + branch flag set + content parser) 跟 PR fix(kimi): gate --work-dir flag on probe, like --print (#1476) #1483 probe-gating approach 自然延伸
• 改动面: agent/kimi (新增 kimi-code flavor detection + flag translation + content rewriter), 中等改动

PM 推荐 path

走 has-pr 路线: 你发 PR 后, PM 派 qa-claudecode review (5+ 必验点 + acceptance criteria).

Pre-dispatch PR check 已 run (per feedback-pm-pre-dispatch-pr-check): gh pr list "1561" search = 0 result (你还没发 PR).

5+ PM acceptance criteria (pre-define, 等你 PR 提交后 dispatch QA review)

1. CLI flavor probe: agent init 时正确检测 kimi-cli (Python) vs kimi-code (Node) — 例如 parse kimi --help for --print 或 kimi --version 输出
2. flag translation: kimi-code 路径正确 drop --print/--quiet/--work-dir, --resume SID → -r SID
3. stream-json content rewrite: assistant content 字符串重写为 []any blocks ([{"type":"text","text":...}]), 跟现有 handleAssistant 对齐
4. permission-mode 默认: 不传 --yolo/--auto (跟 kimi-code Cannot combine --prompt with --yolo 兼容), bare --prompt mode 走 implicit-yolo
5. session resume: -r SID 走通 + extractResumeSessionID 仍然 match hint "To resume this session: kimi -r session_..."
6. fallback: kimi-cli 路径保持原 dialect 不动 (避免 regression, PR fix(kimi): gate --work-dir flag on probe, like --print (#1476) #1483 行为保持)

Willingness to contribute

(注: 你 issue 没勾选 willingness-to-contribute, 但提供了完整 shim — 我解读为 implicit willingness). 如果你发 PR, PR 描述建议包含 Fixes #1561.

Risk

• CLI flavor probe 可能误判 (新版本 kimi-cli 可能新增 flags), 需要 robust version detection
• 内容 rewriter 复杂度 (嵌套 tool_calls 的 content 数组处理)
• 跨平台 (kimi-code 是 Node.js, 跨 macOS/Linux/Windows 安装路径不同)
• 跟现有 fix(kimi): gate --work-dir flag on probe, like --print (#1476) #1483 probe-gating 整合要避免重复检测
• 维护负担: kimi-code 后续版本 flag 变化需要跟进

Sibling / 相关 issue

• feat(agent): add Kimi CLI support #616 (kimi agent 原 PR) — 当前 cc-connect kimi agent targets Python kimi-cli dialect
• fix(kimi): gate --work-dir flag on probe, like --print (#1476) #1483 (probe-gating approach for --work-dir) — 你的 patch 可以基于这个扩展
• [Bug] ❌ 错误: error: unknown option '--work-dir' #1476 (kimi-cli flag 问题) — 不同 flavor 同类问题
• Kimi CLI 不支持 --quiet 参数，导致 /mode quiet 切换后报错 ｜ Kimi cli 错误地自动重置session #806/fix(kimi): remove unsupported --quiet flag, fix session reset on quiet mode (#806) #1253 (kimi-cli compat) — 相关
• [Bug] stream-json 解析逻辑与 Kimi CLI 实际输出格式不匹配，导致 assistant 内容丢失 #1059 (stream-json content schema) — 你的 content rewriter 修这个
• PR fix(kimi): gate --work-dir flag on probe, like --print (#1476) #1483 (probe-gating) — 整合基础

不派 dev

reporter self-implement 路径 (你 shim 已经 production 验证), PM 不重复派. 等你 PR 后 PM dispatch QA review.

— cc-connect/pm · t-20260717-7dajak