feat!: rebuild as a generic Kimi CLI wrapper

[userFRM]

Complete breaking rebuild. The old surface mirrored Kimi's internal file/shell/web tools (14 of them), duplicating what the MCP client already has. This replaces it with a thin, generic wrapper over the Kimi CLI's headless print mode.

Tools

• kimi_run — send a prompt to the Kimi agent, optionally scoped to a working directory; returns the final answer.
• kimi_resume — continue a prior session by id with a follow-up prompt.
• kimi_status — installed binary (native vs legacy), version, and auth health.

Robustness

• Classifies Kimi's documented exit codes (0 ok, 1 permanent, 75 transient), retries transient failures with backoff, and enforces a wall-clock timeout via a process-group kill.
• Detects quota/balance failures even when the CLI exits 0 (observed in the wild).
• Binary resolution prefers the native installer path ~/.kimi-code/bin/kimi, then legacy ~/.local/bin/kimi, then $PATH.
• Auth detection covers OAuth credentials, kimi.json, and an api_key in config.toml.

Safety

Tool auto-approval (-y) is off by default and opt-in per-call (yolo) or via KIMI_YOLO, so an unsupervised agent cannot mutate a repo or run shell without consent.

Verification

• typecheck, build, and 37 unit tests pass.
• End-to-end smoke against the real local binary: tools list correctly and kimi_status reports installed + authenticated.
• Not yet exercised against a live model call here because the test account is out of API balance (HTTP 429).

🤖 Generated with Claude Code