[safeblood]

Summary

Adds a CLI backend that drives the authenticated Kimi Code CLI (kimi --prompt <task> --output-format text), so --backend kimicode works alongside claude/codex/cursor/pi/opencode. Kimi Code (Moonshot's agent CLI/desktop app) already has its sessions harvested via #342; this PR completes the loop by letting the sleep cycle replay, judge, and reflect through the same agent.

Changes

File	What changed	Why
skillopt_sleep/backend.py	New KimiCliBackend (name="kimicode", aliases kimi/kimi_code/kimi-cli) + get_backend/build_backend routing	Prompts go inline via --prompt (kimi's -p takes a value, unlike claude -p which reads stdin). Output is captured as bytes and decoded as UTF-8 (the CLI emits UTF-8 regardless of the Windows console codepage). Binary resolution: PATH first, desktop install (~/.kimi-code/bin/kimi) as fallback for the bare default name only; an explicit --kimi-path is kept verbatim. Auth failures set last_call_error and log a warning.
skillopt_sleep/config.py	kimi_path default	Same pattern as the other CLI paths
skillopt_sleep/cycle.py	Pass kimi_path through both build_backend call sites
skillopt_sleep/__main__.py	--backend kimicode, --kimi-path	CLI surface
tests/test_backend_kimi.py	8 offline tests	Aliases, path resolution, command construction, UTF-8 decoding, auth-error detection, routing — no network

Known limits (documented)

• Each call takes ~60s (Kimi CLI cold start), so a full cycle is markedly slower than other backends; the 180s default timeout fits.
• The CLI has no --bare equivalent: the user's global ~/.kimi-code/AGENTS.md is still injected into optimizer calls (the Claude backend disables the analogous leak via --disable-slash-commands). Flagged here so the trade-off is explicit.

Verification

• pytest tests/test_backend_kimi.py — 8 passed.
• Live single call: _call('Reply with exactly: OK') returns a real model response.
• End-to-end on Windows: dry-run --backend kimicode over a 2-task reviewed file completes attempt/judge/reflect/gate with live Kimi calls (baseline 1.0, candidate 1.0, gate reject — no headroom on trivial demo tasks, as expected).