feat: add Kimi Code (Moonshot AI) preset and hook installer

[Krishnachaitanyakc]

Summary

Adds first-class support for kimi-cli, Moonshot AI's Python coding assistant, against the new agent-preset framework that landed on main in May 2026. Replaces the closed PR #751, which was scoped to the old per-preset arms and is structurally incompatible with the rewritten orchestrator.

Closes #639. Continues the work from #751 per @svarlamov's request to rebase on the simpler agent preset structure.

Preset (presets/kimi_code.rs)

• Pure parser implementing AgentPreset::parse. Handles PreToolUse / PostToolUse for WriteFile, StrReplaceFile, and Shell.
• Strict event handling: lifecycle events (SessionStart, Stop, SubagentStart, PostToolUseFailure, etc.) return PresetError rather than fabricating a phantom file-edit checkpoint.
• transcript_source: None for now — kimi-cli stores sessions on disk under $KIMI_SHARE_DIR/sessions/... but the hook payload doesn't surface that path. A dedicated reader can land in a follow-up.
• agent_id.model defaults to "unknown" because kimi-cli does not surface the active model in hook payloads.

Installer (mdm/agents/kimi_code.rs)

• Writes [[hooks]] entries into ~/.kimi/config.toml per the upstream hook documentation. No matcher field — the preset itself filters tool events.
• Idempotent: repeat install with the desired command is a no-op (compared structurally, not textually, so TOML round-trip whitespace doesn't cause spurious diffs).
• Token-based ownership matching (is_git_ai_kimi_command): a hypothetical sibling preset like kimi-code-pro is never misidentified.
• Uninstall removes only kimi-code entries; user-defined hooks survive. Empty hooks key is removed entirely so the file isn't left with a dangling hooks = [].

Tool classification (bash_tool.rs)

• New Agent::KimiCode variant: WriteFile/StrReplaceFile → FileEdit, Shell → Bash, all else → Skip.

Wiring

• presets/mod.rs::resolve_preset accepts kimi-code.
• mdm/agents/mod.rs::get_all_installers includes KimiCodeInstaller.
• git_ai_handlers help text lists kimi-code.
• tests/integration/repos/test_file.rs::AI_AUTHOR_NAMES and test_repo.rs::is_known_checkpoint_preset learn kimi-code.

Test plan

• cargo fmt --check clean
• cargo clippy --all-targets -- -D warnings clean (no #[allow(...)] annotations added)
• RUSTDOCFLAGS="-D warnings" cargo doc --no-deps clean
• 14 preset unit tests (presets::kimi_code::tests::*) — pre/post for WriteFile/StrReplaceFile/Shell, lifecycle event rejection, unsupported tool rejection, missing field errors, invalid JSON, alias key acceptance, tolerance for extra tool_output field
• 16 installer unit tests (mdm::agents::kimi_code::tests::*) — fresh install (s1), idempotency (s2), preserves unrelated hooks (s3), updates outdated path (s4), dedup with same and differing paths (s5/s5b), no matcher (s6), dry-run (s7), creates parent dir (s8); uninstall variants (u1-u5); invalid TOML/wrong type errors (e1-e3); ownership-token tests (substring lookalikes never match)
• 11 integration tests (tests/integration/kimi_code.rs) — preset routing for all 4 event types, full lifecycle/tool rejection sweep (11 events × 6 tools), 3 E2E flows via real git-ai checkpoint kimi-code subprocess asserting assert_lines_and_blame and session metadata
• Full lib test suite (1467 tests) passes locally on macOS in daemon mode

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 4 additional findings.