Feature Request: Claude-compatible local plugin support

What feature would you like to see?

Add a Claude-compatible local plugin compatibility layer to Kimi Code CLI.

This is intended to address a real user need already raised in the repo, especially around whether Kimi can work with Claude-style plugins and local plugin ecosystems. Related discussion: #1566. It is also adjacent to skill/plugin discovery requests such as #1705.

Proposed v1 scope:

• Load local Claude-compatible plugins via --plugin-dir /path/to/plugin
• Auto-discover plugins from ~/.kimi/claude-plugins/
• Recognize .claude-plugin/plugin.json
• Support these plugin components on a best-effort basis:
  • skills/
  • commands/
  • agents/
  • hooks/hooks.json
  • .mcp.json
  • settings.json (agent key only)
• Keep plugin hooks and MCP session-scoped only
• Keep slash-command invocation for plugin commands/skills
• Expose a concise plugin capability summary in model-visible context so goal-oriented user requests can benefit from installed plugin capabilities without relying on a broad natural-language hard router

Important non-goals for v1:

• No marketplace / install / reload-plugins workflow
• No .lsp.json
• No mutation of persistent user MCP config such as ~/.kimi/mcp.json
• No claim of full Claude runtime compatibility; unsupported fields should be warned and skipped or surfaced as advisory context

Why this seems valuable:

• It would let Kimi interoperate with an existing local plugin ecosystem instead of requiring every plugin to be rewritten into Kimi's current lightweight plugin format
• It helps users who already rely on Claude-style plugin layouts (commands, skills, hooks, .mcp.json) move workflows into Kimi more incrementally
• It provides a practical compatibility path without forcing Kimi to adopt Claude-specific product workflows like marketplace/install upfront

I already have a working implementation in my fork and can open a draft PR if this aligns with the roadmap. The implementation currently covers:

• local plugin loading and auto-discovery
• namespaced plugin skill/command registration
• plugin hooks and MCP loading as session-scoped runtime extras
• plugin default agent handling from settings.json
• model-visible capability summary for installed plugin skills/commands
• tests for discovery / commands / agents / hooks / MCP / router / capability summary / slash-command integration

Latest local verification from the forked branch:

• pytest: 99 passed
• ruff: clean
• pyright: clean

If this direction is acceptable, I can open a draft PR from my fork next.

Additional information

Fork branch with implementation:

• GTC2080/kimi-cli:GTC/claude-plugin-compat

Behavior notes from the current implementation:

• explicit slash commands remain the deterministic path
• broad natural-language hard routing was intentionally removed in favor of a narrower bare-invocation fallback plus model-visible capability summaries
• plugin agents are supported as selected default agents, but are not advertised as autonomous model-executable actions in the capability summary

Review the proposed behavior and the existing implementation in the GTC2080/kimi-cli:GTC/claude-plugin-compat fork, since the issue names no repository files or entry points. Run the reported pytest suite and verify the v1 boundaries, including session-scoped hooks and MCP, supported plugin components, and the listed non-goals.