feat: add Kimi Code CLI support

[rlorenzo]

Adds Kimi Code (the kimi CLI) as a fifth supported harness.

Kimi discovers user-level Agent Skills as SKILL.md directories, the same mechanism Codex and Copilot use, so it slots into the existing generate → install pipeline without a new command format.

Changes

• tools/generate — derives .kimi-code/skills/<name>/SKILL.md alongside the Codex/Copilot/Antigravity skills, from the same .claude/commands/*.md sources. The 6 generated skill files are committed.
• setup — detects kimi, offers it in tool selection, and installs skills to $KIMI_CODE_HOME/skills (default ~/.kimi-code/skills), where Kimi invokes them as /skill:<name>. MCP servers are configured by writing ~/.kimi-code/mcp.json directly: Kimi has no mcp add subcommand, only the interactive /mcp-config TUI editor. A kimi_home helper honors the KIMI_CODE_HOME override throughout.
• lib/lib-review-loop — kimi is now a valid --editor / --reviewer agent, run as kimi -p <prompt>. Non-interactive mode already uses auto permission mode, and Kimi rejects --yolo / --auto when combined with --prompt, so no approval flags are passed.
• Tests — test/generate.bats covers drift in the new skill dir, test/lib-review-loop.bats validates the new agent name, test/smoke includes kimi in its agent list. .pre-commit-config.yaml watches .kimi-code/skills/ for generated-file drift.
• README — supported-tools table, per-command /skill: usage lines, config docs, MCP notes, and uninstall path.

Deliberately skipped

• gh agent skill and Impeccable — neither has an identifier for Kimi, so both steps skip it (the existing empty-id guards handle this; no new branching).
• MCP auto-approve permissions — Kimi's permission rules live in config.toml as a [[permission.rules]] array of tables rather than a JSON allow-list, so setup leaves them to the user, as it already does for Codex.

Known limitations (documented in the README)

• kimi -p takes the prompt as an argv with no stdin form, so on Windows/Git Bash a very large review-loop prompt can hit the ~32 KB OS argument limit — the same limitation Copilot has.
• Kimi has no per-run flag to disable MCP servers, so an autonomous loop run still loads whatever is in ~/.kimi-code/mcp.json (Claude, Codex, and Copilot runs disable theirs).

Verification

• tools/generate --check — in sync
• test/run — 47/47 passing
• shellcheck -x on setup, lib/lib-review-loop, tools/generate, both review loops, test/smoke — clean
• markdownlint-cli2 README.md — clean
• Exercised install_commands and the MCP helpers against a throwaway $HOME: 6 skills installed with the source marker, re-run reports all 6 up to date, mcp.json written with the expected shape, and KIMI_CODE_HOME respected.

Smoke tests against the real kimi CLI were not run — it is not installed in this environment.

🤖 Generated with Claude Code

https://claude.ai/code/session_019H18Vt8ToMYzD9gX64nKnk

---

Generated by Claude Code

[greptile-apps]

Greptile Summary

Adds Kimi Code CLI as a supported harness.

• Generates and installs six Kimi-compatible Agent Skills.
• Adds Kimi detection, custom home-directory support, and direct MCP configuration.
• Enables Kimi in both review loops and smoke-test discovery.
• Updates generated-file checks, tests, and user documentation.

Confidence Score: 5/5

The PR appears safe to merge with no concrete correctness or security failures identified.

The Kimi integration consistently extends the existing generation, installation, agent-dispatch, test-discovery, and documentation paths, with its known CLI and MCP limitations explicitly documented.

Important Files Changed

Filename	Overview
setup	Adds Kimi discovery, skill installation, KIMI_CODE_HOME handling, and MCP JSON configuration without an identified defect.
lib/lib-review-loop	Registers Kimi as an editor or reviewer and dispatches non-interactive prompts through kimi -p.
tools/generate	Extends the existing deterministic skill-generation pipeline to the Kimi output directory.
test/smoke	Includes installed Kimi clients in the existing editor and reviewer smoke-test flow.
README.md	Documents Kimi installation, skill invocation, MCP behavior, review-loop support, limitations, and removal.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[Claude command sources] --> B[tools/generate]
    B --> C[.kimi-code/skills]
    C --> D[setup]
    D --> E[$KIMI_CODE_HOME/skills]
    D --> F[$KIMI_CODE_HOME/mcp.json]
    G[Review loops] --> H[lib-review-loop]
    H --> I[kimi -p prompt]

Loading

_{Reviews (1): Last reviewed commit: "feat: add Kimi Code CLI support" | Re-trigger Greptile}