Fusengine Kimi Code Plugins

Port of the Fusengine Claude Code plugins to Kimi Code CLI, re-architected for K3 (1M-token context, native reasoning, native vision) — not a literal translation.

• 24 plugins · 196 skills · 37 agents · 34 commands · 68 hook rules
• Same agentic core as the Claude version: APEX methodology, sniper/challenger verification, domain experts, SOLID/DRY guards — wired through the same @fusengine/harness binary
• Measured parity with Claude Code on the benchmark suite: 3/3 tasks, 3.6–6.3× lower latency — see docs/BENCHMARK.md

Layout

kimi-code/
├── KIMI.template.md          # Legacy duplicate of plugins/kimi-rules/templates/KIMI.md.template
├── AGENTS.md                 # Thin project-only instructions (global rules live in the template + corpus)
├── marketplace.json          # Plugin catalog for /plugins marketplace
├── plugins/                  # 24 plugins (kimi.plugin.json manifests)
│   ├── kimi-rules/           # Core rules corpus 00-08 + template
│   ├── ai-pilot/             # sniper, challenger, explore-codebase, research-expert, …
│   └── …                     # domain experts (nextjs, laravel, react, rust, go, …)
├── scripts/
│   ├── migrate.ts            # Claude→Kimi migration pipeline (6 steps, JSON report)
│   ├── install-kimi.ts       # Installer → $KIMI_CODE_HOME
│   ├── validate.ts           # Conformance checks (24/24 green)
│   ├── hooks/kimi-hook-shim.mjs   # Env bridge → @fusengine/harness
│   └── benchmark/            # 3-task comparative benchmark (kimi vs claude)
└── docs/
    ├── MIGRATION.md          # Architectural analysis Claude→K3
    └── BENCHMARK.md          # Protocol + measured results

Install

# Install the plugin
/plugins install https://github.com/fusengine/kimi-code

# Setup (hooks + API keys + MCP servers)
~/.kimi-code/plugins/managed/fusengine/setup.sh        # macOS / Linux
~\.kimi-code\plugins\managed\fusengine\setup.ps1       # Windows

Then /plugins reload (or start a new session). One plugin, one namespace: 196 skills, 34 commands (/fusengine:<cmd>), 63 harness-backed hook rules. API keys for MCP servers: cp ~/.kimi-code/plugins/managed/fusengine/.env.example ~/.kimi-code/.env, fill in, re-run setup.sh.

How it works

The installer (run by setup.sh): backs up existing config, stages @fusengine/harness from npm, copies plugins/kimi-rules/templates/KIMI.md.template as ~/.kimi-code/AGENTS.md, merges the rules corpus between idempotent fences, merges MCP servers into ~/.kimi-code/mcp.json, and materializes the 37 agents into ~/.kimi-code/agents/. Kimi Code does NOT auto-load .env — the installer reads it when resolving ${VAR} placeholders into mcp.json, and the harness reads it at runtime.

User prompt → Hook detects project type → Expert agent activated
            → Hook loads SOLID references → Agent reads docs via MCP
            → Hook blocks if DRY violation → Agent writes code
            → Hook checks file size → Agent runs Sniper validation
            → Commit gate blocks on lint/type errors → Commit with version bump

Every step is intercepted — same harness, same enforcement as the Claude Code version.

Requirements

• Bun ≥ 1.1 · Kimi Code CLI ≥ 0.29
• Le harness @fusengine/harness est installé depuis npm (registre officiel) par l'installateur — aucune dépendance envers une autre CLI

Statusline (terminal daemon)

Kimi exposes no user statusline surface, so the port renders outside the TUI — terminal title + ~/.kimi-code/statusline.txt (tmux-ready):

bun run plugins/core-guards/statusline/src/daemon.ts start   # stop | status

Shows ⎇ branch*dirty · model · effort · ctx % · edits · age · dir, read from the live session's state.json/wire.jsonl. Details: plugins/core-guards/statusline/README.md.

Regenerating

bun run scripts/migrate.ts              # claude-plugins/plugins → plugins/ (24/24, JSON report)
bun run scripts/scrub-claude-refs.ts
bun run scripts/apply-core-overrides.ts # re-applies the hand-written K3 core (kimi-rules, ai-pilot agents, shims)
bun run scripts/validate.ts
bun run scripts/benchmark/run.ts --cli=both

The hand-written core lives in core-overrides/ (kimi-rules/ plugin + the 8 ai-pilot agents) so a full mechanical regeneration never loses it.

Known accepted regressions

• One type:"prompt" hook (core-guards Stop) dropped — Kimi hooks are command-only; noted in plugins/migration-report.json.
• Agent model: routing (sonnet/opus) removed — all subagents run on the main model (K3).
• See docs/MIGRATION.md §5 for the full list and rationales.