Repository navigation
Kimi Code integration
- All 23 WebDev skills and nine bounded native Kimi roles.
- Project startup instructions through a managed .kimi-code/AGENTS.md pointer; clean staging for user-wide plugin installation preserves private project state.
- Opt-in project-filtered experience hooks: text prompts, tool results and failures, with missing/truncated content marked.
- Kimi proposal and text A/B evaluation with tools/subagents disabled, explicit model selection and configuration-drift checks.
- Safe hook configuration merge/removal and unique backups.
- Installation checklist for Context7, Playwright, GitHub, conditional Figma, Desktop browser, WebBridge and Computer Use, including Windows/macOS setup.
Installation
Download the matching client archive below and verify it against SHA256SUMS. Preserve existing project facts, learning data and custom instructions.
Kimi installation and MCP/browser setup
Verification and limits
All three PR workflows passed: source/package validation, quality checks and links, including Linux/Windows helper regressions. The local learning suite passed 104 tests with one skip.
Authenticated live Kimi skill/role discovery, capture, proposal/A-B execution and browser/computer workflows remain unverified. Native browser routing is documented, not certified; packaged subagent tool allowlists do not automatically grant browser access. Run browser QA in the main session with verified tools.
Kimi Stop events do not expose final assistant text. Tool output can be truncated upstream, and visual attachments remain metadata-only in the learning evaluator. Text benchmarks do not certify real coding behavior.
Release approved with these disclosed limits. No existing user installation or model credentials were changed.