Kimi UI Agent

Kimi UI Agent

Use the bundled Bun CLI script as the deterministic controller for Kimi-powered frontend/UI work. The CLI owns repo scanning, adapter rendering, path validation, project config, run state, and artifact files. The active harness, Codex, Kimi Code, or Claude Code, owns judgment, edits, and final verification.

Start

1. Confirm the user explicitly invoked $kimi-ui-agent or explicitly asked for Kimi UI Agent.

2. Run:

   KIMI_UI_AGENT_SKILL_DIR="${KIMI_UI_AGENT_SKILL_DIR:-${CODEX_HOME:-$HOME/.codex}/skills/kimi-ui-agent}"
   KIMI_UI_AGENT_CLI="${KIMI_UI_AGENT_CLI:-$KIMI_UI_AGENT_SKILL_DIR/scripts/kimi-ui-agent.ts}"
   bun "$KIMI_UI_AGENT_CLI" --json doctor
   

3. If this repo has not been profiled, run setup in dry-run mode first:

   bun "$KIMI_UI_AGENT_CLI" --json setup --dry-run
   

4. Apply setup only when durable project intelligence should be written:

   bun "$KIMI_UI_AGENT_CLI" --json setup --apply
   

5. Read .agents/kimi-ui-agent/project-profile.md, .agents/kimi-ui-agent/frontend-map.md, .agents/kimi-ui-agent/design-system.md, .agents/kimi-ui-agent/verification.md, and .agents/kimi-ui-agent/protected-paths.md when present.

Workflow

• Use setup for first-run deterministic repo scanning and lean project context generation. Read setup-profile before changing profile behavior or interpreting profile files.
• Use install --dry-run to render project-local Codex, Kimi Code, and Claude Code adapters. Read install-targets before applying adapter writes.
• Use start --task "<task>" --dry-run to preview a plan-first isolated worktree run. Apply only after checking the intended worktree, branch, and artifact paths.
• Use status, reply, continue, finalize, and abort to manage a run. Read lifecycle before modifying run protocol or interpreting artifacts.
• Use mcp only as a stdio MCP server configured by Codex, Kimi Code, or Claude Code. Read adapter-contracts before changing MCP tool schemas.

Guardrails

• Keep this skill explicit-only. Do not enable implicit invocation for an editing/worktree orchestrator.
• Do not paste secrets or private tokens into Kimi, Claude, Codex, config files, or artifact files.
• Keep install and setup dry-run-first. Every write requires --apply.
• Treat generated project context as editable hints, not source of truth. Verify against current code before acting.
• Keep Kimi launch plan-first by default. Autonomous or yolo modes are opt-in and must be named by the user.
• Hooks are workflow guards, not a sandbox. Use worktree isolation, harness permissions, and path containment as the real safety boundary.

Resources

• scripts/kimi-ui-agent.ts: Bun TypeScript CLI entrypoint. Invoke it with bun "$KIMI_UI_AGENT_CLI" after setting KIMI_UI_AGENT_CLI to the installed or repo-local script path; do not assume a global kimi-ui-agent binary exists.
• references/security-and-privacy.md: sanitization, path, env, and artifact safety rules.
• references/output-contract.md: JSON and Markdown artifact contract.
• references/kimi-code.md, references/claude-code.md, and references/codex.md: harness-specific notes.
• assets/design-brief.md: optional prompt input for high-stakes UI work.