Local, proof-driven autonomous engineering runtime powered by Kimi. Archived as of June 2026. This repository is read-only.
OMK was built around the Kimi Wire Protocol — a programmatic JSON-RPC interface that allowed external orchestrators to spawn, control, and intercept Kimi agents. This protocol enabled OMK's core capabilities:
- Approval Proxy — intercept and policy-gate every tool call
- Lifecycle Hooks — inject scripts at PreToolUse, Stop, and other checkpoints
- Replay — re-run a goal timeline from persisted events
- Steer — programmatically influence agent decisions
In mid-2026, MoonshotAI released kimi-code (v0.14.0), a TypeScript rewrite of the CLI that replaces the Wire Protocol with ACP (Agent Client Protocol). ACP is designed for IDE integration (editors like Zed and JetBrains driving a chat session), not for external orchestration.
ACP does not provide:
- Interception of tool calls (they auto-execute or use internal TUI flows)
- External hook injection
- Replay or steer methods
- Programmatic cancel of a running turn
Without these primitives, OMK cannot function as designed. A hotfix using kimi-legacy (the old Python CLI, v1.43.0) is possible but is a dead end — that CLI is no longer maintained.
The codebase remains a useful reference for several standalone components (see below).
OMK was an attempt to build a local, proof-driven autonomous engineering runtime — turning a high-level goal like "Add OAuth and rate-limiting to the API" into planned, verified, and delivered repository changes, entirely locally, with no cloud control plane.
Key differentiators that (at the time) no other tool offered:
| Capability | Description |
|---|---|
| Multi-agent team scheduler | Run N workers in parallel with role packs, task claims, leases, and inbox/outbox IPC |
| Proof-driven delivery | Every goal produced proof.json with gate results, readiness verdicts (ready / not_ready / blocked), and failure artifacts |
| Durable goal state | Goals survived process crashes. Pause, resume, cancel, replay from SQLite + JSONL state |
| Git worktree isolation | Parallel slice execution in isolated worktrees with conflict detection and auto-rebase |
| Six-review wall | Architect, code, test, security, performance, and anti-slop reviews before any PR |
| Approval policy engine | Never / Safe / Yolo / Pattern policies with timeouts and human-in-the-loop channels |
| Cost tracking & budgets | Token, USD, and time budgets with enforcement and reporting |
| Verification gates | Configurable gate runner (cargo test, clippy, fmt, security checks) with evidence capture |
The following components from the OMK codebase can be reused as standalone utilities or libraries. Each links to the relevant crate or module in this repository.
| Component | Source | Description | Standalone Value |
|---|---|---|---|
| Gate Runner | crates/omk-gates/, crates/omk-proof/ |
Configurable verification gate execution (test, lint, fmt, security) with structured proof.json output |
Drop-in CI quality gate. Like pre-commit but with machine-readable artifacts |
| Task Graph | crates/omk-scheduler/, crates/omk-db/ |
Task dependency graph with claims, leases, retries, ownership, and SQLite persistence | Lightweight embedded alternative to Temporal or Airflow |
| Event Log | crates/omk-events/ |
Append-only JSONL event envelopes with typed readers, writers, and sinks | Universal event-sourcing primitive for any Rust project |
| Budget Tracker | crates/omk-cost/, crates/omk-db/ |
Token/USD/time cost estimation and budget enforcement with SQLite storage | Attach to any LLM API client (OpenAI, Anthropic, etc.) |
| Worktree Manager | src/cli/team/worktree.rs (and related) |
Git worktree creation, isolation, auto-rebase, and conflict detection | Parallel feature development without merge conflicts |
| Approval Policy Engine | crates/omk-wire-worker/src/approval/ |
Pattern-matching approval policies with timeouts and human channels | Can be adapted as a pre-hook for any agentic tool |
The design and architecture documents remain valid references for building agent orchestration systems:
SPEC.md— Product spec, delivery contract, and goal lifecycleROADMAP.md— Staged roadmap (Stage 0–8) from foundation to long-horizon reliabilitydocs/ARCHITECTURE.md— System design, module boundaries, and data flowdocs/COMPETITIVE_POSITIONING.md— Market map vs Devin, OpenHands, Claude Code, AiderAGENTS.md— Multi-agent workflow rules, Rust safety rules, async architecture, error handling doctrinedocs/API.md— Machine-readable CLI output contractsdocs/TUTORIAL.md— Step-by-step first run (historical reference)
If you were using OMK or considering it, here are practical alternatives:
| Your Need | Alternative |
|---|---|
| Interactive AI coding in terminal | kimi-code — the official TypeScript CLI from MoonshotAI. Open source, MIT, ACP-based. |
| Multi-agent orchestration | OpenHands or Devin |
| Git worktree isolation | Use git worktree natively, or build on the reference logic in this repo |
| CI verification gates | Extract omk-gates logic, or use cargo-husky + custom JSON output |
| Cost tracking | Build on omk-cost crate, or use provider dashboards |
kimi-code is open source (github.com/MoonshotAI/kimi-code, MIT, TypeScript). Several ideas explored in OMK may be valuable there:
- Cost tracking & budgets — token/USD visibility and enforcement per session
- Approval policy engine — pattern-based approvals with timeouts (beyond the current
default/auto/yolomodes) - Structured proof artifacts — a machine-readable completion report after each task
- Verification gates — optional post-task test/lint gate runner
If you are interested in carrying any of these forward, the OMK codebase is a fully working reference implementation.
MIT © oh-my-kimi contributors