kimi-plugin-cc

Use Kimi as Claude Code's second reviewer, independent thinker, and delegated worker — without building your own multi-agent stack.

This is a Claude Code plugin that drives the kimi-code CLI (the Node.js successor to Kimi CLI) as a subprocess. Claude can ask Kimi for a structured code review, delegate a bug hunt, or have Kimi double-check its own work before stopping — all through slash commands, with persistent job state, session resume, and per-command safety enforced by a PreToolUse hook installed in ~/.kimi-code/config.toml.

• Independent model, independent perspective. Kimi reasons differently from Claude. A second opinion from a different model catches things self-review misses.
• No orchestration layer required. No ACP, no cloud broker, no shared API keys. The plugin talks directly to a locally-installed kimi binary using kimi -p --output-format stream-json.
• Full agent capabilities, safely bounded. Kimi can read files, write code, run shell commands, and resume where it left off — all bounded by a workspace allowlist invoked via the PreToolUse hook. Read-only commands enforce read-only at the hook layer, not the prompt.

Migrating from v0.4? v0.4.x targeted the Python Kimi CLI and stays available at the v0.4.0 tag (v0.4-maintenance branch is cut from that tag for ongoing fixes — see the linked tag if the branch is not yet pushed). v1.0 is a hard cut to kimi-code — install kimi-code first, then /plugin update kimi will upgrade you in place. See docs/migration.md for the step-by-step upgrade.

Try it in 60 seconds

# Prerequisite: install kimi-code from https://kimi.com/code/docs
/plugin marketplace add linxule/kimi-plugin-cc
/plugin install kimi@kimi-marketplace
/kimi:setup
/kimi:review "review my current diff"

Kimi reads your working-tree diff and returns a review as markdown:

## Verdict: concern

One high-confidence issue in the auth middleware; the rest looks fine.

### src/middleware/auth.ts:42-47 — JWT expiry not checked before token refresh

The refresh handler calls `getNewToken()` without first checking whether the
current token has actually expired. On a slow-clock client this triggers a
refresh on every request.

Suggested fix: add an expiry check before the refresh call.

Claude reads the review and can act on it directly. For programmatic access to the same content plus job metadata, companion.sh result <jobId> --json returns a structured envelope {job_id, kind, status, summary, error, artifact_path, body, ...} where body is the full markdown. To go further:

/kimi:rescue "fix the top review finding"

Kimi opens the file, writes the fix, runs the relevant tests, and reports back. The session persists — if you restart Claude Code, /kimi:rescue --resume picks up where it left off.

When to use what

The plugin ships seven Claude Code subagents — one per delegated capability — that the main thread can dispatch via the Agent tool: the read-only kimi-review, kimi-challenge, kimi-ask, and kimi-swarm (parallel review fan-out), plus the write-capable kimi-rescue (a bounded delegated task), kimi-pursue (experimental autonomous goal mode), and kimi-swarm-write (a parallel edit fan-out that returns a reviewable patch — the plugin never applies it). The heavy modes — kimi-swarm, kimi-pursue, and kimi-swarm-write — carry a mandatory --budget ceiling. Safety does not widen on the agent path: the index-0 PreToolUse hook fires on every tool call of every turn — read-only labels deny writes; rescue/pursue/swarm-write apply the workspace allowlist and cannot mutate git state (write-swarm confines edits to a throwaway worktree and hands back a patch the human merges). Each agent's description is Kimi's own statement of what it's good for — Claude matches the moment and dispatches; no prescriptive skill manual in between.

How it works

The architecture is modeled after OpenAI's codex-plugin-cc. Both follow the same pattern: thin plugin shell, rich local runtime, one process per job.

  /kimi:review "check the auth flow"
       │
       └─ companion.sh → node dist/companion.js
               │
               ├─ spawns: kimi --output-format stream-json -p "<prompt>"
               │            with KIMI_PLUGIN_CC_CMD=review in the env block
               ├─ parses OpenAI-shaped NDJSON records (assistant content + tool_calls)
               ├─ PreToolUse hook (installed by /kimi:setup) enforces the
               │   per-command policy: review/challenge/review_gate/ask are
               │   read-only; rescue uses the workspace allowlist
               ├─ persists job + stream log to SQLite (node:sqlite, zero native deps)
               └─ returns structured result to Claude

Subprocess-first transport. v1.0 drives kimi -p --output-format stream-json as a one-process-per-job subprocess. kimi-code mints the session id and announces it via a role:"meta", type:"session.resume_hint" record on stdout (kimi-code 0.2.0+) — earlier 0.1.x announced on stderr instead. The runtime consumes both channels (first-announce-wins) and round-trips the captured token verbatim via --resume. The v0.4 Wire JSON-RPC client is gone (kimi-code dropped it); the v0.4-maintenance branch keeps the Wire path alive for Kimi CLI users.

Safety via PreToolUse hook. kimi-code's kimi -p mode hard-codes permission: auto and auto-approves every tool call. The plugin's safety contract therefore lives in a PreToolUse hook that /kimi:setup or $kimi-setup installs as a host-scoped managed block in ~/.kimi-code/config.toml. The hook reads KIMI_PLUGIN_CC_CMD from the env block we set per spawn and applies the right policy — read-only for review/challenge/review_gate/ask, workspace allowlist for rescue. Without a verified hook, every model-spawning command refuses before Kimi starts; the review gate skips rather than running un-enforced.

Workspace allowlist. Rescue's allowlist (runtime/rescue-approval.ts) is the security boundary — not the prompt. Symlink-aware path containment, .git/ exclusion, a curated set of read-only check tools, mutating-flag detection on git/find/sed, and explicit rejection of package-manager run <script> (opaque scripts are a supply-chain risk). The hook calls evaluateRescueHookRequest for every rescue tool call.

Job lifecycle. Every job gets a SQLite record, a stream-json diagnostic log, and a kimi_session_id captured from kimi's stream-json meta record (0.2.0+) or stderr announce (0.1.x fallback). Jobs go through running → completed/failed/cancelled. Use /kimi:status, /kimi:result, /kimi:cancel, /kimi:replay to manage them.

Kimi Code session titles. For plugin-created user-command sessions, the runtime performs best-effort post-run title sync after Kimi returns a session id. Titles are deterministic (Kimi <Command>: <summary>), derived from plugin command metadata rather than model output, and manual/custom Kimi titles are preserved. The review gate is excluded so Stop-hook checks do not clutter the Kimi Code session list.

Zero native dependencies. The runtime uses Node 22.5's built-in node:sqlite — no better-sqlite3, no node-gyp, no compilation step. dist/ is precompiled and committed, so installed plugins work immediately with just node on PATH.

490+ tests, drift gate. The test suite covers the stream-json parser (including the kimi-code 0.2.0 session-meta record and the 0.8.0 goal.summary record), cli-client lifecycle, approval policy, approval hook entry script, rescue allowlist, command handlers (incl. /kimi:pursue goal-mode parsing/exit-mapping), job lifecycle, setup managed-block installer, kimi version probe, Codex surface generation/lock (subfolder layout, runtime-mirror + orphan checks, Claude/Codex surface separation), and more. bun run check rebuilds dist/, typechecks, runs the suite, and fails if the rebuild produces uncommitted changes — forgotten rebuilds can't ship. Opt-in real-binary smokes (bun run smoke:real) spawn the actual kimi -p and prove the read-only commands' write denial end-to-end, plus that autonomous goal mode is hook-gated on every continuation turn; they're skipped by default (they need a real kimi binary + authenticated home).

CI (.github/workflows/, see docs/ci.md). bun run check runs on every push to main and every PR (ci.yml) — free, no secrets, smokes auto-skip. The real-binary smokes (including the /kimi:pursue goal-mode safety gate) run on manual dispatch (smoke.yml), authenticated via an API-key secret so they don't depend on the expiry-prone OAuth subscription; the workflow is inert until that secret is added.

Also surfaces in Codex desktop. The Codex plugin is a self-contained subfolder, plugins/kimi-codex/ (its own .codex-plugin/plugin.json, skills/, and a byte-mirror of the runtime), with the Codex repo marketplace at .agents/plugins/marketplace.json pointing into it. It lives in a subfolder — not at the repo root beside the Claude files — so Claude Code does not auto-discover the Codex skills (the Claude Code surface stays commands + agents). Codex skills call the same shell companion runtime as Claude commands; there is no MCP server or second execution path. The plugins/kimi-codex/ tree is generated by bun run generate:surfaces.

Install

Via the Claude Code marketplace (recommended)

/plugin marketplace add linxule/kimi-plugin-cc
/plugin install kimi@kimi-marketplace
/kimi:setup

/kimi:setup writes the PreToolUse hook to ~/.kimi-code/config.toml and probes it both directly and through /bin/sh -c so launch-from-GUI/LaunchAgent setups (where node may not be on the shell's PATH) get caught up front rather than silently auto-approving every tool call. As of v1.8, setup also parses the complete TOML, validates every configured hook, and serializes the read-modify-write transaction with a private adjacent lock; malformed foreign config or simultaneous Claude/Codex setup fails safely instead of silently dropping or overwriting enforcement.

From a local clone

git clone https://github.com/linxule/kimi-plugin-cc ~/kimi-plugin-cc
claude --plugin-dir ~/kimi-plugin-cc

Then run /kimi:setup to install the safety hook.

Via the Codex repo marketplace

The Codex marketplace sidecar lives at .agents/plugins/marketplace.json and points at the self-contained plugin root plugins/kimi-codex/ (source.path: "./plugins/kimi-codex"). Add the marketplace and install the kimi plugin:

codex plugin marketplace add linxule/kimi-plugin-cc
codex plugin add kimi@kimi-marketplace

Then run the $kimi-setup skill to install/check the kimi-code safety hook. The Codex skills mirror the Claude surfaces: $kimi-review, $kimi-challenge, $kimi-ask, $kimi-rescue, $kimi-pursue, $kimi-swarm, $kimi-swarm-write, $kimi-status, $kimi-result, $kimi-cancel, and $kimi-replay.

Using both Claude Code and Codex

Both hosts share one ~/.kimi-code/config.toml. As of v1.7.0 each host manages its own host-scoped PreToolUse block, so run /kimi:setup in Claude Code and $kimi-setup in Codex once each — they coexist and no longer overwrite each other. (Before v1.7.0 a single shared block was clobbered whenever you switched hosts, forcing a re-setup.) See docs/safety.md § Host scoping.

Removing the integration

/kimi:setup --uninstall        # removes THIS host's block only
/kimi:setup --uninstall --all  # removes every host's block from the shared config

Removes the managed block(s) from ~/.kimi-code/config.toml. The plugin itself can be uninstalled via /plugin uninstall kimi.

Prerequisites

• kimi-code installed and authenticated. The plugin spawns kimi -p; set KIMI_PLUGIN_CC_KIMI_BIN to override the binary location.
• Node >= 22.5 — for built-in node:sqlite. Set KIMI_PLUGIN_CC_NODE_BIN to override.
• bun — only for contributor tooling. Not required at runtime.

Still on Python Kimi CLI? Stay on the v0.4.0 tag (or its v0.4-maintenance branch, once published) — see docs/migration.md.

Development

bun run check    # rebuild dist/, typecheck, run the full test suite, drift gate
bun test <path>  # run a single test file

See CONTRIBUTING.md for the full contributor workflow and AGENTS.md for the architecture invariants coding agents should preserve.

How this was built

This plugin was built through the same multi-model collaboration it enables — Claude, Kimi, and Codex working together at every stage from design through pre-public audit. The v1.0 cutover (PRs 1-5) used the same pattern: each PR landed with paired Claude code-reviewer + Codex codex-rescue adversarial reviews, contradictions resolved against the kimi-code source, and convergent findings applied before commit.

Acknowledgments

• Kimi (Moonshot AI) — the reasoning model this plugin delegates to, and a design/review collaborator throughout development
• Codex (OpenAI) — architecture consults, parallel implementation, and independent pre-ship reviews via the codex-plugin-cc companion plugin
• Claude Code (Anthropic) — the host environment, primary development agent, and the platform this plugin extends

License

Apache-2.0