Contributing

For repository setup, changesets, and pull requests, see the Contributing Guide. For the shared source workflow, see Canonical Source Workflow.

Package Boundary

Kimi Code is a declarative package with no runtime SDK or build step. The manifest (kimi.plugin.json), sync.config.ts, and tests are hand-authored; skills, commands, and SYSTEM.md are generated from packages/core/agent-directives/. The sync configuration adapts canonical prompts to Kimi’s skill, command, and system-prompt formats.

Manifest Contract

The manifest owns the Kimi-native entry points:

• skills and commands - generated skill and command directories
• systemPromptPath - the generated SYSTEM.md system-prompt contribution
• sessionStart.skill - the skill loaded at session start (orchestrator)
• skillInstructions - plugin-wide dispatch instruction text

Change Shared Methodology

Edit packages/core/agent-directives/, never packages/kimi-code/skills/, commands/, or SYSTEM.md. Regenerate and verify from the repository root:

scripts/sync-all
scripts/check-sync
pnpm --filter @maestria/kimi-code test

Kimi-specific changes belong in sync.config.ts or the manifest. Do not copy canonical prompt text into the package.

Add a Skill

Add the canonical prompt under packages/core/agent-directives/ (or a Kimi-only addition when the behavior does not belong in core). Add its transform and subagent profile in sync.config.ts, its entry in the orchestrator routing table when it is a specialist, and manifest and safety assertions in packages/kimi-code/tests/. Then run the commands above.

Read-only roles are advisory: Kimi permission rules apply to the session, not to an individual persona. The seven-specialist roster is part of the current routing contract; update the routing table and tests together when it changes.

See Also

• Installation
• Quick Start
• Kimi Code README