feat(kimi-code): emit commands onto the skills surface

[dyoshikawa]

Summary

Kimi Code has no custom-command directory outside plugins, but it registers every external Skill as a slash command (/skill:<name>, plus a /<name> shorthand when the name is free) and expands $ARGUMENTS, positional $0/$1 and named $<name> placeholders in the Skill body. That is the same contract a rulesync command has, yet the plain kimi-code target still reported commands as unsupported.

This PR emits rulesync commands onto Kimi Code's skills surface, following the existing Devin / Warp / Hermes Agent commands-as-skills pattern.

• New KimiCodeCommand writes one <slug>/SKILL.md per command to .kimi-code/skills/ (project) and ~/.kimi-code/skills/ (global; $KIMI_CODE_HOME/skills/ when KIMI_CODE_HOME is set).
  • name and description are always written, since Kimi fails to parse a directory-form SKILL.md that lacks either.
  • disableModelInvocation: true keeps the command-derived skill user-invoked. A kimi-code: section in the command frontmatter can set type, whenToUse, disableModelInvocation and arguments. Kimi's alias spellings (when-to-use/when_to_use, disable-model-invocation/disable_model_invocation) are normalized to those keys. A blank description falls back to <name> command.
  • Import and --delete are no-ops for commands on this target because the skills feature owns the tree.
• KimiCodeSkill.isDirOwned makes the skills feature skip directories named after a current command that targets kimi-code in the Kimi-specific root (rulesyncCommandSlugExists gained an optional toolTarget filter; Devin and Warp are unchanged), so it neither imports them as skills nor deletes them as orphans. The import-only .agents/skills/ root is not affected. KimiCodePluginSkill opts out, because kimi-code-plugin writes commands to commands/.
• Registered kimi-code in the commands processor and tuple, regenerated the support tables, .gitignore and the embedded docs, and added a Kimi Code note to docs/reference/file-formats.md.
• Tests: unit tests for the new class and the ownership hook, project/global e2e happy paths plus the KIMI_CODE_HOME e2e, and the Devin orphan-protection e2e case now covers kimi-code as well.

Upstream references:

• https://moonshotai.github.io/kimi-code/en/customization/skills.html (frontmatter fields, required name/description, body placeholders, skill locations)
• https://moonshotai.github.io/kimi-code/en/reference/slash-commands.html ("Skill Dynamic Commands": /skill:<name> and the /<name> shorthand)

Remaining on #2502

• Plugin MCP and hooks for kimi-code-plugin. Kimi only reads these from the kimi.plugin.json manifest, and the target does not write that manifest yet. Whether to write it is a maintainer design decision.
• Reading the [permission] allow / deny / ask array form on import. This needs an ownership decision: reading the arrays without owning them would duplicate rules, and owning them would delete hand-written arrays.

Refs #2502

🤖 Generated with Claude Code

https://claude.ai/code/session_012y9fNo2AAJdGocxvsHtDhv