Overview

Meta_Kim is not another AI coding tool. It is a governance layer for durable AI coding work.

The hard part of AI coding is no longer getting a model to change files. The hard part is deciding what should happen first, which capability should own it, what evidence proves it worked, and how the lesson survives the next run.

Claude Code, Codex, OpenClaw, and Cursor are all hands: they can write code and change files. But who decides which file to change first? Who reviews the result? Who fixes the problems that show up? And how do we make sure the same mistake does not repeat next time?

Meta_Kim is built for that. It is the governance layer above the coding hands: a runnable set of agents, skills, contracts, hooks, scripts, and evidence gates that keeps complex work from turning into a mess.

One-line summary

First clarify what needs to happen -> then decide who should do it -> review after execution -> preserve what was learned -> feed that back into the next run.

This is not a new concept. Mature engineering teams already do this. Meta_Kim turns it into a runnable system instead of relying on human discipline alone.

Before / After

What 3.0 changes

Meta_Kim 3.0 makes the governance system easier to trust: advice, durable truth, user-facing views, installation state, and knowledge evolution no longer blur into one another.

The implementation follows an inward dependency rule:

This layering is intentionally incremental: existing script entrypoints remain compatibility facades where needed, while new stable logic lives behind the layered boundaries.

3-minute proof

Meta_Kim is easiest to understand by watching one governed run, not by reading every rule.

npm run meta:theory:demo
npm run meta:run-status:latest
npm run meta:theory:report -- --run-id latest
npm run meta:delivery:bundle

meta:run-status:latest is a minimal redacted status summary. Use the explicit meta:theory:report -- --run-id latest readback to inspect report content.

The proof path shows five things:

• a fuzzy request is turned into an explicit intent and success standard
• capability search happens before execution ownership is chosen
• work is split into bounded worker tasks instead of one giant chat response
• review and verification produce artifacts, not just reassuring prose
• compatibility evidence stays tiered, so smoke evidence is never promoted to native live proof

The executable core-loop contract is config/contracts/core-loop-contract.json; it binds the default path to npm run meta:theory:run -- "<task>" and keeps Critical -> Fetch -> Thinking -> Execution -> Review -> Meta-Review -> Verification -> Evolution testable. npm run meta:theory:demo is the zero-argument replay entry for the 3-minute proof.

Real stage execution is opt-in and read-only. Use npm run meta:theory:run -- --execute-stage-dag --stage-runner-runtime codex "<task>" or replace codex with claude; resume the exact interrupted run with --resume-stage-dag --run-id <id> --task "<same task>". Both runtimes consume the same coreLoop.stageDagPacket, record native session/tool/timing evidence, persist completed nodes through the durable kernel, and merge locally. The default remains planned-only. Ready sets use native concurrency by default; maintainers may explicitly install the currently tested @langchain/langgraph@1.4.8 and add --stage-runner-orchestrator langgraph to use its Functional API as an execution wrapper. LangGraph does not own topology or checkpoints, and this mode still does not execute writes or external side effects.

For a guided walk-through, start with examples/first-run/README.md.

Quick Start

If you just want to try it quickly, run:

npx --yes github:KimYx0207/Meta_Kim meta-kim

Or install it the traditional way:

git clone https://github.com/KimYx0207/Meta_Kim.git
cd Meta_Kim
npm install
node setup.mjs

💡 After install: setup.mjs prints where every artifact lives. A global install can use meta-kim status from any directory; an npx install can repeat npx --yes github:KimYx0207/Meta_Kim meta-kim status. An npx launch remains a supported entrypoint: before install or update writes persistent Claude Code or Codex projections, Meta_Kim fixes the exact package into an immutable store under the user home, keyed by package version and packed-package SHA-256, so Commands, Hooks, and merged settings never depend on the disposable npm cache. Repository maintainers may also use npm run meta:status.

At a fresh clone, Meta_Kim intentionally separates source files, generated projections, and local state:

Install scope

The default Enter path is global reusable capabilities. Agents, commands, MCP, and skills are installed into each selected runtime's official global/home locations when that runtime supports them. Global hook wiring is opt-in: pass --with-global-hooks when you intentionally want Meta_Kim to update Claude/Codex/Cursor hook settings. Projects reuse global capabilities directly; project-local agents, commands, MCP, hooks, or skills are created only when Fetch/Thinking proves project-specific customization, iteration, or a dedicated override is needed.

If you explicitly choose Project directory updates, setup asks which project directories to update and writes the target-selected project runtime projection there, including project hooks/config where that runtime supports them. This path does not install global reusable capabilities and does not run project cleanup.

Project files are still allowed, but they are not the default reusable capability store. Confirmed project bootstrap writes only project context/config/state plus proven project-specific overrides, preserving existing user config through managed blocks, add-only writes, protected JSON merge, backups, and manifests. Every applied project bootstrap records .meta-kim/ state and backup files.

Update also treats third-party installers as untrusted configuration producers. For Codex, Meta_Kim restores the user's pre-install configuration and applies only its narrowly owned native controls. This prevents an upstream installer from resurrecting a third-party MCP server the user deleted, silently adopting a new server, or persisting known Meta_Kim benchmark/test project registrations. User-owned MCP servers, projects, hooks, agents, and unrelated settings remain intact.

Codex agent fan-out has a bounded default of two threads and one nested level. Explicit user limits are preserved; only an absent value or Meta_Kim's former default of six threads is migrated.

If you plan to maintain the repository, edit the canonical sources first: canonical/agents/, canonical/skills/meta-theory/, config/contracts/, and config/capability-index/. Then run (requires Node.js >= 22.13.0):

npm run meta:sync
npm run meta:validate

Recommended reading order:

1. This file, README.md
2. AGENTS.md
3. CLAUDE.md when working on Claude Code behavior
4. canonical/runtime-assets/cursor/rules/meta-enforcement.mdc when working on Cursor rules

Usage Paths

After the default install (node setup.mjs or npx) or a confirmed project bootstrap, humans should be able to use plain task language. Slash commands remain maintainer shortcuts, not the normal user path.

Platform Support Tiers

Meta_Kim now tracks platform support in tiers instead of treating every compatible surface as a full runtime projection.

Source of truth: config/runtime-compatibility-catalog.json.

Dependency-project install targets are handled by those upstream projects and are not part of Meta_Kim's public runtime support list.

Surface compatibility is intentionally weaker than runtime support. A tool can share Meta_Kim-compatible primitives and still need adapter design, profile/layout generation, sync tests, and live validation before it becomes a formal projection. Dependency-project install matrices are not repeated here as Meta_Kim support claims.

Contact

GitHub KimYx0207 | X @KimYx0207 | Website aiking.dev | WeChat Official Account: 老金带你玩AI

Feishu knowledge base: long-term updates

Buy me a coffee

If Meta_Kim has been useful, support the project with a coffee.

WeChat Pay	Alipay

Method basis

Meta_Kim’s methodological foundation comes from research on meta-based intent amplification, authored by this project’s maintainer (KimYx0207):

• Paper: https://zenodo.org/records/18957649
• DOI: 10.5281/zenodo.18957649

Architecture: Hidden Skeleton + Dynamic Dealing

This is the core design idea of Meta_Kim. If you only read one section, read this one.

First, split the core terms so they do not get mixed up later

If you only want one sentence to remember:

The 8-stage workflow moves execution forward, gates decide whether a stage can pass, contracts define the required outputs, and dealing adds dynamic intervention.

8 stages = the hidden skeleton

Meta_Kim has 8 fixed execution stages. This is the hidden skeleton:

flowchart LR
    C[Critical<br/>Clarify the request] --> F[Fetch<br/>Search capabilities]
    F --> T[Thinking<br/>Plan the approach]
    T --> E[Execution<br/>Dispatch the work]
    E --> R[Review<br/>Inspect the result]
    R --> MR[Meta-Review<br/>Review the review]
    MR --> V[Verification<br/>Verify reality]
    V --> EV[Evolution<br/>Write back lessons]

    style C fill:#fbbf24,color:#000
    style F fill:#34d399,color:#000
    style T fill:#60a5fa,color:#000
    style E fill:#f87171,color:#fff
    style R fill:#a78bfa,color:#fff
    style MR fill:#a78bfa,color:#fff
    style V fill:#34d399,color:#000
    style EV fill:#fbbf24,color:#000

Critical - pin down the real problem first

When the request is vague, ask clarifying questions instead of guessing. This stage produces intentPacket, which locks down the real user intent, success criteria, and exclusions. If the request is already clear, the system records an explicit skip reason instead of quietly skipping.

Fetch - search existing capabilities before inventing new ones

Search whether existing agents, skills, tools, or MCP integrations already cover the need. The core idea here is capability-first: define the capability first, then search for the owner that declares it, then dispatch to the best match. Capability-index lookup goes config/capability-index/ -> runtime mirror -> local inventory -> fallback. Do not start by hardcoding a specific agent name.

Governance Decision Engine

Meta_Kim is not only the 8-stage spine. It first identifies the governance trigger, checks runtime and OS capability, checks dependency capability, separates owner from weapon, filters by Win/Mac/runtime support, asks the user only for branch-changing choices, executes deterministic parts, verifies whether the user goal actually landed, and writes reusable learning back. Reference-only projects are absorbed into Meta_Kim data, not silently promoted into dependencies; see config/governance/decision-pattern-catalog.json.

Automation is assistive, not authoritative. It may gather evidence, draft options, run deterministic checks, surface blockers, and prepare readable status, but branch-changing judgment in Critical, Fetch, Thinking, and Review remains a human decision point. A selected capability, hook match, report, or validator pass must not be relabeled as human acceptance or native runtime evidence.

Thinking - define boundaries, owners, sequence, deliverables, risks, and stop conditions

Break the task into subtasks, assign owners, and make dependencies and parallel groups explicit. This stage produces a dispatchBoard: who does what, what can run in parallel, and who is responsible for merging the result. At least two solution paths should be explored; do not lock into a single route too early.

Execution - produce the actual work while still under governance

Dispatch the subtasks to specialist agents. Each subtask is wrapped in a workerTaskPacket, including file context, constraints, review owner, and verification owner. Independent subtasks should run in parallel when possible. Execution is not completion - the output still has to pass review and verification.

Review - check quality and boundary compliance

Inspect code quality, security, architecture compliance, and boundary violations. Produce a structured reviewPacket with findings. Each finding has a severity from CRITICAL to LOW. This is not a formality - unresolved findings cannot move forward.

Meta-Review - inspect whether the review standard itself is biased or too loose

Review the review. If the review standard is too weak, the system is not really reviewing. If it is biased, it is reviewing the wrong thing. This stage protects the quality of the review system itself.

Verification - confirm that reality matches the claim

Verify whether the fixes really closed the review findings. This stage produces verificationResult and closeFindings. If the fix did not actually close the finding, go back and repair it before verifying again. This is the most honest gate in the system.

Evolution - write capability gaps and reusable patterns back into the system

Convert experience into structural upgrades: reusable patterns go into memory, failures become learning artifacts, capability gaps are handed to Scout, and agent boundaries are written back into canonical sources. Every run must end with a writebackDecision: either write back something concrete or explicitly explain why there is nothing to persist. A run that does not preserve learning is wasted work.

The 8 stages together form the execution spine.

Why are they only "relatively" fixed? Because some stages can be skipped in simple cases - but the system must explicitly record why they were skipped. Nothing is skipped silently.

11 phases = a business workflow built on the skeleton

If the 8-stage workflow is the skeleton, then the 11-phase business workflow is the run-packaging progression that grows on top of it:

direction -> planning -> execution -> review -> meta_review -> revision -> verify -> summary -> feedback -> evolve -> mirror

It is not a second system. It is derived from the 8-stage skeleton. The difference is:

• The 8 stages focus on execution logic - "what order should work happen in"
• The 11 phases focus on business governance - "what each phase must deliver, how completion is defined, and when mirrors must be refreshed"

flowchart TB
    subgraph spine["8-stage workflow (hidden skeleton)"]
        direction LR
        C1[Critical] --> F1[Fetch] --> T1[Thinking] --> E1[Execution] --> R1[Review] --> MR1[Meta-Review] --> V1[Verification] --> EV1[Evolution]
    end

    subgraph workflow["11-phase business workflow"]
        direction LR
        D2[direction] --> P2[planning] --> EX2[execution] --> RE2[review] --> MET2[meta_review] --> REV2[revision] --> VER2[verify] --> SUM2[summary] --> FB2[feedback] --> EVO2[evolve] --> MIR2[mirror]
    end

    C1 -.-> D2
    F1 -.-> P2
    T1 -.-> P2
    E1 -.-> EX2
    R1 -.-> RE2
    MR1 -.-> MET2
    V1 -.-> VER2
    EV1 -.-> EVO2

    style spine fill:#1e1b4b,stroke:#7c3aed,color:#e0e7ff
    style workflow fill:#14532d,stroke:#22c55e,color:#dcfce7

The 11-phase business workflow adds revision, summary, feedback, and mirror, so the process is not only about "getting it done" but also about getting it done well, closing the loop correctly, and keeping runtime projections aligned.

Contracts = what each node must deliver

Workflow alone is not enough. Each stage also needs to define what it must output. That is what the contracts do.

Meta_Kim contracts are not verbal agreements. They are structured packets:

flowchart LR
    subgraph packets["Contract artifact flow"]
        direction LR
        IP[intentPacket<br/>Intent lock] --> DP[dispatchBoard<br/>Dispatch board]
        DP --> WTP[workerTaskPacket<br/>Task packet]
        WTP --> RP[reviewPacket<br/>Review findings]
        RP --> RR[revisionResponse<br/>Revision response]
        RR --> VR[verificationResult<br/>Verification result]
        VR --> SP[summaryPacket<br/>Final summary]
        SP --> EW[evolutionWriteback<br/>Learning writeback]
    end

    IP ~~~ C2["Critical"]
    DP ~~~ T2["Thinking"]
    WTP ~~~ E2["Execution"]
    RP ~~~ R2["Review"]
    RR ~~~ REV2["Revision"]
    VR ~~~ V2["Verification"]
    SP ~~~ S2["Summary"]
    EW ~~~ EV2["Evolution"]

    style packets fill:#1a1a2e,stroke:#e94560,color:#fff

These artifacts are not optional documents. They are the system’s source of truth. Without contracts, the next node is not "handing off" - it is guessing what the previous node meant. That is why so much AI collaboration falls apart on complex work.

The current implementation carries these artifacts explicitly: taskClassification before execution, cardPlanPacket before dealing, dispatchEnvelopePacket before dispatch, reviewPacket.findings after review, revisionResponses + verificationResults + closeFindings between revision and verification, summaryPacket before external publication, and writebackDecision before evolution.

npm run meta:validate:run checks whether these artifact chains close completely.

Gates = stage reached does not mean stage passed

Contracts define what each node must deliver. Gates define whether that delivery is good enough to move forward.

In one sentence:

A stage tells you where you are; a gate tells you whether you are allowed to move on.

flowchart LR
    A["Reach a stage"] --> B{"Gate decision"}
    B -->|Pass| C["Release: move forward"]
    B -->|Fail| D["Revision: add evidence / fix output"]
    B -->|Hold| E["Pause: wait for conditions to mature"]
    B -->|Escalate| F["Higher-level intervention"]

    style A fill:#dbeafe,stroke:#2563eb,color:#000
    style B fill:#7c3aed,stroke:#4c1d95,color:#fff
    style C fill:#dcfce7,stroke:#16a34a,color:#000
    style D fill:#fee2e2,stroke:#dc2626,color:#000
    style E fill:#e0f2fe,stroke:#0284c7,color:#000
    style F fill:#fef3c7,stroke:#f59e0b,color:#000

Key gates in the system:

The most important one is the publicDisplay gate. If verification has not passed, the summary is not closed, or the deliverable chain is broken, the system cannot pretend that the work is finished.

The relationship between gates and contracts:

• Contracts answer "what must this node deliver" - they are about delivery obligations
• Gates answer "is this good enough to move forward" - they are about release decisions
• Without contracts, gates have nothing to judge
• Without gates, contracts are just a ceremony

Dynamic dealing = flexibility layered on top of the skeleton

The 8-stage skeleton is relatively fixed, but real tasks vary too much to be handled by one rigid path. That is why Meta_Kim introduces dynamic dealing.

Dealing corresponds to the 8 stages, but not as a simple 1:1 map. The 10 cards are:

The important part is that some cards are dynamic:

• When three high-attention cards are dealt consecutively, the system forcibly inserts Pause - it does not wait for the user to notice
• When security risk appears, Risk preempts the current flow
• When the user already knows something, the corresponding card is skipped
• When task iteration exceeds the upper bound, the system escalates to Warden adjudication

Dynamic dealing gives the fixed skeleton some breathing room: strict where it must be strict, flexible where flexibility helps.

flowchart TD
    START[Current card completed] --> SKIP{Check next card<br/>skip_condition}
    SKIP -->|Satisfied, skip| NEXT[Continue to the next card]
    SKIP -->|Not satisfied| INTR{Check interrupt queue}
    INTR -->|Security risk preempts| RISK[Risk card<br/>highest priority]
    INTR -->|No preemption| PAUSE{Three or more<br/>high-cost cards?}
    PAUSE -->|Yes, force a break| P[Pause card<br/>zero attention]
    PAUSE -->|No| DEAL[Deal by priority]
    RISK --> DEAL
    P --> DEAL
    DEAL --> COUNT{Iterations over<br/>max_iterations?}
    COUNT -->|Yes| WARDEN[Escalate to Warden adjudication]
    COUNT -->|No| START

    style RISK fill:#dc2626,color:#fff
    style P fill:#1e3a5f,color:#93c5fd
    style WARDEN fill:#7c3aed,color:#fff
    style DEAL fill:#16a34a,color:#fff

Closed loop = iterate, generate, improve

Once the skeleton, progression workflow, contracts, and dynamic dealing are in place, the system forms a closed loop:

Request arrives -> skeleton starts -> dealing decision -> dispatch execution -> review and verify -> preserve lessons -> upgrade agents -> next run starts stronger

The loop is not one-and-done. Each round can:

1. Generate the missing agent - if a capability gap appears, the system can create a new agent through the Type B pipeline
2. Improve agent capability - Evolution writes back changes to SOUL.md, skill loadouts, and toolchains
3. Clarify every agent’s boundary - each agent owns one class of work; boundary violations are intercepted by Sentinel

flowchart TD
    INPUT[Request arrives] --> SPINE[Hidden skeleton starts]
    SPINE --> CARD[Dynamic dealing decision]
    CARD --> DISPATCH[Dispatch to specialist agent]
    DISPATCH --> REVIEW[Review + verification]
    REVIEW --> |Pass| EVOLVE[Preserve lessons]
    REVIEW --> |Fail| FIX[Fix + review again]
    FIX --> REVIEW
    EVOLVE --> UPGRADE[Upgrade agent capability]
    UPGRADE --> |Capability gap found| CREATE[Type B pipeline<br/>auto-create new agent]
    UPGRADE --> |Boundary needs adjustment| BOUNDARY[Adjust agent boundary]
    CREATE --> INPUT2[Next run starts stronger]
    BOUNDARY --> INPUT2

    style INPUT fill:#fbbf24,color:#000
    style EVOLVE fill:#34d399,color:#000
    style CREATE fill:#f87171,color:#fff
    style INPUT2 fill:#fbbf24,color:#000

Agent boundaries + skill integration

The 9 meta roles each own a different domain:

Each agent can load powerful skills and commands as needed. Meta_Kim ships with 9 community skills and supports custom extension.

flowchart TD
    WARDEN[meta-warden<br/>Coordination / arbitration / synthesis] --> CONDUCTOR[meta-conductor<br/>Workflow / rhythm]
    WARDEN --> GENESIS[meta-genesis<br/>Agent design]
    WARDEN --> ARTISAN[meta-artisan<br/>Skill / tool matching]
    WARDEN --> SENTINEL[meta-sentinel<br/>Security / permissions / rollback]
    WARDEN --> LIBRARIAN[meta-librarian<br/>Memory / continuity]
    WARDEN --> PRISM[meta-prism<br/>Quality review]
    WARDEN --> SCOUT[meta-scout<br/>External capability discovery]
    WARDEN --> CHRYSALIS[meta-chrysalis<br/>Evolution writeback]

    GENESIS -.-> |SOUL.md| ARTISAN
    ARTISAN -.-> |Skill loadout| GENESIS
    CONDUCTOR -.-> |Task board| WARDEN
    SENTINEL -.-> |Security interception| WARDEN
    PRISM -.-> |Review report| WARDEN
    SCOUT -.-> |Capability candidates| ARTISAN
    LIBRARIAN -.-> |Context memory| WARDEN
    CHRYSALIS -.-> |Scar / writeback proposal| WARDEN

    SKILLS[9 community skills<br/>+ custom extensions] --> ARTISAN
    HOOKS[Hook automation<br/>intercept / format / check] --> SENTINEL

    style WARDEN fill:#7c3aed,color:#fff
    style CONDUCTOR fill:#60a5fa,color:#000
    style GENESIS fill:#fbbf24,color:#000
    style ARTISAN fill:#34d399,color:#000
    style SENTINEL fill:#f87171,color:#fff
    style LIBRARIAN fill:#a78bfa,color:#fff
    style PRISM fill:#fb923c,color:#000
    style SCOUT fill:#2dd4bf,color:#000
    style CHRYSALIS fill:#84cc16,color:#000

Hook automation

In Claude Code, Meta_Kim uses hooks for automation:

• Dangerous command blocking: operations like rm -rf and DROP TABLE are blocked automatically
• Git push reminder: remind you to check before pushing
• Formatting: automatically format JS/TS files after edits
• Type checking: run TypeScript checks after edits
• console.log warning: remind you to remove console.log
• Session-end audit: check for leftover issues before the session ends
• Session-end memory save: write session summaries to MCP Memory Service on session end
• Subagent context injection: automatically inject project context into subagents

These hooks are not optional polish. They are the execution-layer guardrails of the governance system.

Cross-platform mapping

Public runtime support follows three tiers: Primary, Beta, and Compatibility.

Meta_Kim currently owns two primary formal projection targets, two opt-in beta compatibility adapters, and two non-default compatibility projection targets:

The canonical source layer is canonical/agents/, canonical/skills/meta-theory/, config/contracts/, and config/capability-index/. The repository mirrors that layer into platform-specific projections through npm run meta:sync.

Open-source boundary: Generated runtime projection directories are local outputs, gitignored, and not GitHub source. That includes .claude/, .codex/, .agents/, .cursor/, openclaw/, .mcp.json, and codex/. The nine governance agents live in canonical/agents/; Meta_Kim sync does not generate execution-layer Codex agents such as worker, explorer, frontend, backend, test, review, analysis, verify, or docs.

flowchart TB
    CANONICAL["canonical/ + config/<br/>(single source layer)"]

    CANONICAL --> |npm run meta:sync| CLAUDE[".claude/<br/>Claude Code<br/>agents + skills + hooks"]
    CANONICAL --> |npm run meta:sync| CODEX[".codex/ + .agents/<br/>Codex<br/>governance agents.toml + skills + hooks"]
    CANONICAL --> |npm run meta:sync| OPENCLAW["openclaw/<br/>OpenClaw<br/>workspaces + skills + internal hooks"]
    CANONICAL --> |npm run meta:sync| CURSOR[".cursor/<br/>Cursor<br/>agents + rules + skills + hooks + MCP"]

    BETA["Beta adapters<br/>ZCode / DeepSeek Harness / Qoder / Trae"] -.-> |packaged compatibility adapters| CANONICAL

    style CANONICAL fill:#7c3aed,color:#fff
    style CLAUDE fill:#fbbf24,color:#000
    style CODEX fill:#34d399,color:#000
    style OPENCLAW fill:#60a5fa,color:#000
    style CURSOR fill:#f87171,color:#fff
    style BETA fill:#b45309,color:#fff

Meta_Kim has two Primary targets, four Beta adapters, and two Compatibility projections. Claude Code and Codex are the default primary path; ZCode, DeepSeek Harness, Qoder, and Trae are Beta; OpenClaw and Cursor are Compatibility.

The decision-authority boundary is narrower than projection support. Current Codex app-server and Claude SDK/CLI callbacks can provide exactly correlated, non-authorizing observations, but public host surfaces do not prove Codex Desktop UI, human identity, or a human answer. That trusted-host authority remains parked. Legacy governance-gate parity and cutover are separately deferred, so 3.0 does not claim that shadow/read-only results have replaced production gates. OpenClaw still lacks a Meta_Kim typed-plugin enforcement adapter for tool blocking, and Cursor arbitrary native-choice popup authority remains unverified.

The point is format discipline, not a ranking: each formal tool target keeps its own agent, skill, hook, MCP, choice, and config surface instead of pretending one host's format is universal.

Choice surfaces are tool-specific. Claude Code should use AskUserQuestion; Codex should use request_user_input when ~/.codex/config.toml has [features].default_mode_request_user_input = true; Cursor uses an alwaysApply project rule to trigger a chat decision card plus official preToolUse / failClosed hooks for tool gating; OpenClaw uses workspace/chat cards unless a typed plugin approval hook is explicitly installed and strictly tested.

Four-layer repository structure

Three state layers (project / global / local)

These three layers are easy to mix up, so they must stay separate:

What lives inside .meta-kim/?

.meta-kim/ is Meta_Kim's local save file. It does three things:

1. Remembers your choices — local.overrides.json

When you run node setup.mjs for the first time and pick "I want Claude Code and Codex", that choice is saved here. Next time you run setup, you don't have to choose again.

Example: You have Claude Code, Codex, and OpenClaw installed, but only want the first two. This file stores that preference — all scripts read it to know which runtimes to install skills for.

2. Records work history — state/{profile}/run-index.sqlite

When you run a governed workflow (e.g. "use the 8-stage spine to review some code"), the result can be indexed into a SQLite database. Later you can query "what did I review last time, what was found, what's still unresolved?"

Example: Last week you asked meta-prism to review the auth module. This week you changed the auth module again. The system checks .meta-kim/state/ and finds "last review found 3 issues, 2 were fixed, 1 is still open" — you don't have to repeat yourself.

3. Cross-session recovery — state/{profile}/compaction/

When you're halfway through a conversation and your token budget runs out, the compaction packet saves your current progress (which step you're on, what's still pending) so you can pick up where you left off in a new session.

Example: You ask Meta_Kim to do a complex multi-file refactor. You get through step 6 before the session ends. Next session, the system reads the compaction packet: "at step 6, step 7 hasn't started" — picks up from step 7, no need to start over.

Other files: doctor-cache/ stores npm run meta:doctor:governance results (written after each run), migrations/ tracks schema upgrades between Meta_Kim versions, profile.json stores profile metadata. All managed by scripts — you never edit them by hand.

Quick reference:

What works globally vs. in-repo only

Meta_Kim separates reusable global capability from directory-authorized governance. After global installation (node setup.mjs), global skills, agents, commands, and MCP entries can be discovered from any project when the runtime supports them. Global hooks require explicit --with-global-hooks; project-governed behavior applies only where this directory has confirmed context/config/state or project-specific overrides. This table separates reusable capability from checks that require the Meta_Kim source repo:

The first three layers are the primary defense once a directory is enabled. Code validation is a final safety net that requires running from the Meta_Kim repo (or pointing to its scripts).

Three-Layer Memory

Meta_Kim does not use a single memory layer. It uses three, each with a different job, so agents can keep improving while becoming more familiar with the project.

Each layer has different activation requirements:

• Layer 1 is built into Claude Code — requires Claude Code runtime (auto-memory at ~/.claude/projects/*/memory/)
• Layer 2 is installed automatically by node setup.mjs
• Layer 3 is installed by node setup.mjs; setup attempts a background HTTP start, and manual startup is the fallback when the health endpoint is unavailable (see Layer 3 activation below)

Layer 1: Memory (agent upgrade memory)

• Responsibility: agent upgrades and continuous learning
• Storage: .claude/projects/*/memory/
• Mechanism: before each run ends, the system reads memory and uses it to decide whether the agent should be upgraded or its boundary should change
• Core value: agents get smarter over time instead of restarting from zero each time
• Activation: automatic — AI reads and writes memory during each session
• Query: ask AI directly — "what did we learn from previous sessions about this project?"

Layer 2: Graphify (project-level LLM wiki)

• Responsibility: project-level code knowledge graph
• Storage: graphify-out/graph.json (NetworkX node-link format); humans and agents use it through query/path/explain slices, with graphify-out/GRAPH_REPORT.md reserved for broad architecture orientation
• Mechanism (data): node setup.mjs (optional Python step) installs graphify and idempotently runs python -m graphify claude install and python -m graphify hook install even if graphify was already installed via pip; git hooks rebuild the graph on commit/checkout in the current repo. npm run meta:graphify:install does the same (including hooks).
• Windows migration: if an existing Claude project still reports C:Users...graphify.EXE: command not found, run meta-kim doctor hooks --fix from that project. It backs up .claude/settings.json and repairs only the known unsafe Graphify Hook form; use --all only when you also intend to inspect user-level settings.
• Mechanism (usage): synced meta-theory dev-governance.md Fetch Step 0.5 defines how the model should detect and use the graph — not a background service. Claude Code subagents get a short hint via subagent-context.mjs, not automatic embedding of graph.json. Focused work should call graphify query, graphify path, or graphify explain to get candidate file anchors, then verify route-changing claims against source files. Codex / OpenClaw / Cursor share the same reference after meta:sync but have no SubagentStart hook; optional python -m graphify codex install or python -m graphify claw install in a target repo patches that repo’s docs per graphify CLI (python -m graphify --help).
• Core value:
  • Make memory increasingly familiar with the project - not by remembering raw code, but by understanding structure and relationships
  • Reduce hallucinations - agents start from graph-backed file anchors and verify claims against source instead of guessing
  • Cut token usage - subgraph extraction replaces raw file reads, with up to 71x compression
• Quality threshold:
  • Fuzzy nodes > 30% -> mark the graph as low quality and fall back to direct file reads
  • Total nodes < 10 -> the graph is too sparse and should fall back to Glob/Grep
  • A "god node" with too many incoming edges -> mark as a serial bottleneck
• Activation: node setup.mjs optional Python step or npm run meta:graphify:install — install/check, networkx, Claude-side registration, this repo’s git hooks; first graph build still depends on a hook run or a manual build command
• Query: python -m graphify query "your question" — natural language query against the code graph

Platform Automation Comparison

Key insight: Claude Code, Codex, and Cursor all have native hook configuration, but their schemas differ. OpenClaw uses its own internal/plugin hook model.

For multi-platform setups, run node setup.mjs — it loops through all selected platforms and runs graphify <platform> install for each one idempotently.

Layer 3: SQL (vector-level session retrieval)

• Responsibility: vector storage and retrieval for project sessions
• Storage: SQLite + vector extension (sqlite-vec)
• Mechanism: store key session information as vectors, then retrieve later by semantic similarity
• Core value:
  • Cross-session continuity - pick up where the last conversation left off
  • Vector-level retrieval - semantic understanding instead of keyword matching
  • Precise recall - find the most relevant context from historical sessions
• Activation: node setup.mjs installs and configures the MCP Memory Service (Layer 3), installs runtime memory hooks, then attempts to start the HTTP service in the background.
  • For Claude Code: SessionStart and Stop memory-save hooks are auto-registered during node setup.mjs; session-start writes project state via mcp_memory_global.py --mode session.
  • For Codex: ~/.codex/hooks.json receives SessionStart, UserPromptSubmit, and Stop bridges to meta-kim-memory-save.mjs, so start/prompt/end checkpoints are automatic.
  • For OpenClaw: ~/.openclaw/hooks/mcp-memory-service receives a managed hook for command:new, command:reset, session:compact:after, and command:stop.
  • For Cursor: ~/.cursor/hooks.json receives beforeSubmitPrompt and stop bridges to the shared memory hook.
• Start server: memory server --http (with MCP_ALLOW_ANONYMOUS_ACCESS=true on macOS/Linux, or $env:MCP_ALLOW_ANONYMOUS_ACCESS="true" in Windows PowerShell), then verify http://127.0.0.1:8000/api/health.
• Port: health-check examples use http://127.0.0.1:8000; shipped hooks default to http://localhost:8000 unless MCP_MEMORY_URL, META_KIM_MEMORY_PORT, or runtime memory config overrides the endpoint. Both loopback hosts are accepted.
• Hooks: auto-registered for Claude Code, Codex, Cursor, and OpenClaw; each runtime uses its native hook format while sharing the same MCP Memory HTTP endpoint.
• MCP registration vs writes: .mcp.json registers the MCP Memory server (memory server) for client access. Automatic session writes are separate lifecycle hooks: Claude Code uses stop-memory-save.mjs, Codex/Cursor use meta-kim-memory-save.mjs, and OpenClaw uses its managed mcp-memory-service hook.
• Evidence boundary: a configured .mcp.json, installed hook, HTTP health response, successful write, successful read, and cross-session recall are different evidence layers. Do not claim cross-session recall from setup or health checks alone.
• Query: npm run meta:query:runs -- --owner <agent> — find past runs by agent, or npm run meta:index:runs -- <artifact> for manual indexing of validated run artifacts
• Troubleshooting:
  • Python hook fails on Windows: If the SessionStart hook fails with exit code 49 or shows no output, the Python command may point to the Windows Store shim. Run node scripts/install-mcp-memory-hooks.mjs to auto-detect and fix. The installer now skips WindowsApps shims and prefers explicit Python executables from LOCALAPPDATA\Programs\Python*. Use --force flag to re-register even if current path appears valid.
  • Check installation: Run node scripts/install-mcp-memory-hooks.mjs --check to verify hook status and Python path validity.
  • Check HTTP health: Run curl -fsS --max-time 3 http://127.0.0.1:8000/api/health after startup. npm run meta:test:mcp checks Meta_Kim's runtime MCP server, not this external MCP Memory HTTP service.
  • Manual verification: Test your Python command with python --version or the detected path with "C:/Users/YOUR_USER/AppData/Local/Programs/Python/Python311/python.exe" --version.

How the three layers work together

flowchart TB
    subgraph memory["Layer 1: Memory"]
        M_IN[Before the run ends<br/>read memory] --> M_JUDGE[Decide whether the agent<br/>needs an upgrade]
        M_JUDGE --> M_OUT[Update boundary<br/>adjust capability]
    end

    subgraph graphify["Layer 2: Graphify"]
        G_IN[When source files > 20<br/>auto-generate graph] --> G_COMPRESS[Subgraph extraction<br/>up to 71x compression]
        G_COMPRESS --> G_QUERY[Agent answers from graph<br/>facts]
    end

    subgraph sql["Layer 3: SQL"]
        S_IN[Session key information<br/>stored as vectors] --> S_INDEX[SQLite + sqlite-vec<br/>vector index]
        S_INDEX --> S_RECALL[Semantic similarity<br/>precise recall]
    end

    memory <--> graphify
    graphify <--> sql
    sql <--> memory

    GOAL1[Reduce hallucinations<br/>answer from facts, not invention]
    GOAL2[Reduce token usage<br/>compression instead of full reads]

    memory --> GOAL1
    graphify --> GOAL1
    graphify --> GOAL2
    sql --> GOAL2

    style memory fill:#fbbf24,color:#000
    style graphify fill:#34d399,color:#000
    style sql fill:#60a5fa,color:#000
    style GOAL1 fill:#dc2626,color:#fff
    style GOAL2 fill:#dc2626,color:#fff

The three memory layers work together toward two core goals:

1. Greatly reduce hallucinations - agents answer from facts and context instead of inventing details
2. Greatly reduce token usage - use graph compression instead of full-file reads and vector retrieval instead of brute-force search

Ops Command Quick Reference

Daily use

Project directory updates only touch directories you select, pass with --project-dir, or save for reuse. Add --save-project-dirs with --project-dir to remember a script-provided list for later --all-projects runs. Existing local settings, MCP, and hook configuration files are merged or preserved instead of being blindly replaced.

Interactive update flow:

1. Run node setup.mjs --update.
2. If you already saved project directories, choose "Update all saved project directories".
3. To configure them for the first time, choose "Add or change saved project directories, then update them".
4. Enter the directories in one line, separated by semicolons or commas: D:/Project/a; D:/Project/b; D:/Project/c.
5. Remembered directories are stored in .meta-kim/local.overrides.json under this Meta_Kim checkout as projectDeployDirs; later runs can use node setup.mjs --update --all-projects.

Sync and validation

Additional governance and scope checks (meta:install-scope:verify, meta:project-cache:verify, meta:doctor:governance) are listed in AGENTS.md.

Where the active rules live

When in doubt, edit canonical sources first, then run npm run meta:sync and npm run meta:check.

How to tell which scripts are useful

Start with package.json scripts. The supported maintenance paths are the meta:* commands documented above; most helper files under scripts/ exist because those commands call them. Use npm run meta:status, npm run meta:check, and npm run meta:verify:all for normal operation. Inspect an individual helper only when a package.json script or test points to it.

Skills and dependencies

Global dependency install/update commands refresh .meta-kim/state/{profile}/capability-index/global-capabilities.json after they modify runtime homes, so newly installed agents, skills, commands, MCP providers, hooks, plugins, and runtime tools are available to capability-first routing without a separate manual scan.

planning-with-files is a core external dependency, not a project-local .agents/skills/ mirror. After dependency install, check runtime home directories such as ~/.codex/skills/planning-with-files/, ~/.claude/skills/planning-with-files/, ~/.cursor/skills/planning-with-files/, or ~/.openclaw/skills/planning-with-files/. Do not conclude it is missing from the absence of .agents/skills/planning-with-files/ alone.

Native dependency installs (Superpowers, ECC, cli-anything)

Superpowers has native plugin entry points in Claude Code, Codex, and Cursor. Meta_Kim no longer treats the old Codex / Cursor skills/superpowers fallback as a correct plugin install; update runs remove the legacy fallback written by older Meta_Kim versions and tell users to use the host-native plugin entry point.

ECC uses the upstream affaan-m/ECC package and plugin identity. Meta_Kim delegates ECC-specific target support to the upstream ECC installer and keeps only the cleanup/hand-off boundary here, so dependency-owned target lists do not become Meta_Kim platform support claims.

For plugin bundles without a native host plugin entry point, the installer still falls back to a sparse-checkout of the upstream bundle's runtime-specific subtree:

Sparse-checkout fallback trees land in ~/.<runtime>/skills/<id>/; native ECC installs do not. The default install/update path selects Claude Code + Codex when the user presses Enter; run npm run meta:deps:install:claude-plugins for the Claude marketplace path only, or npm run meta:deps:install:all-runtimes to cover Claude Code, Codex, OpenClaw, and Cursor explicitly. Upgrading from an older install? Legacy full-repo clones are auto-detected by the .claude-plugin/ marker at the target root and re-extracted on the next run; old Codex/Cursor skills/superpowers, skills/ecc, and skills/everything-claude-code fallbacks are removed or replaced with native-install instructions.

Advanced ops

More advanced commands (meta:validate:run, meta:eval:agents, meta:eval:agents:live, meta:index:runs, meta:query:runs, migrate:meta-kim) are in AGENTS.md.

FAQ

Q: I installed via npx, where are my files?

Meta_Kim uses two distinct scopes; a normal global install does not populate the current project with durable runtime mirrors:

1. Your home — ~/.claude/, ~/.codex/, ~/.cursor/, and ~/.openclaw/ hold globally reusable assets selected for those runtimes.
2. Global manifest — ~/.meta-kim/install-manifest.json tracks managed global files for safe update and rollback.
3. Stable execution package — a global-writing install/update first stores the exact package under ~/.meta-kim/runtime/projection-packages/<package>/<version>/<packed-sha256>/. Claude Code and Codex Commands, Hook registrations, and merged settings/config reference this stable root instead of an npx cache or temporary extraction directory.

Project runtime mirrors are created by an explicit project install/bootstrap or by governed runtime sedimentation. In global_only, installation itself may retain only the minimal project Hook dependency closure required by the host contract. If a later governed run creates or iterates an Agent, Skill, or Command, Meta_Kim copies it into the current project with independent ownership so dependency updates cannot replace that project version.

One exception is intentional: if a global install/update detects a project that already has a valid Meta_Kim bootstrap manifest, it refreshes that existing project with the project's own saved runtime targets and merge/delta policy while also updating the global installation. It does not create a new project projection, and it preserves runtime-sedimented capabilities and user-owned files.

From a global install, run meta-kim status from any directory to see the full footprint. With npx, repeat npx --yes github:KimYx0207/Meta_Kim meta-kim status and replace status with check, doctor, update, or uninstall as needed. These commands resolve scripts from the package rather than the current directory. Help, status, and doctor remain query/diagnostic entrypoints and do not create the immutable package store merely because they were invoked; check is read-only and validates the current version's manifest-bound package authority. Install/update may materialize the stable package before writing global projections. Uninstall removes a stored bundle only when the manifest proves exact Meta_Kim ownership and the bundle has not drifted; unknown, changed, and user-owned content is preserved. Repository maintainers can keep using the equivalent npm run meta:* commands.

Q: What is different about Meta_Kim compared with a normal AI coding assistant?

A normal AI coding assistant does what you ask, with no governance layer in between. Meta_Kim inserts several layers between "ask" and "do": first it confirms what you actually want, then it plans who should do it, then it reviews the result, then it verifies the fix, and finally it preserves the lesson. It is not another AI; it is engineering discipline for AI.

Q: I only need to change one file. Do I need Meta_Kim?

No. Meta_Kim is for cross-file, cross-module, and multi-capability tasks. If you are only changing one function inside one file, plain Claude Code is enough. Do not use a cannon to hit a mosquito.

Q: What is the relationship between the 8-stage workflow and the 11-phase business workflow?

The 8-stage workflow is the execution skeleton (Critical -> Fetch -> Thinking -> Execution -> Review -> Meta-Review -> Verification -> Evolution) and stays relatively fixed. The 11-phase business workflow is a run-packaging workflow defined in config/contracts/workflow-contract.json (direction -> planning -> execution -> review -> meta_review -> revision -> verify -> summary -> feedback -> evolve -> mirror) and focuses on deliverable flow, closure, and runtime mirroring. It does not replace the 8 stages; it adds governance discipline on top.

Q: What does dynamic dealing mean?

The 8-stage workflow is fixed, but real tasks vary too much. Dealing gives the system flexibility inside that fixed path - for example, after three high-intensity actions in a row, the system can auto-pause with Pause; when a security issue appears, Risk can preempt the current flow. The fixed skeleton protects the baseline, and dynamic dealing provides adaptability.

Q: Will the three-layer memory be too heavy?

No. The three layers have separate jobs:

• Memory is very light - just a few markdown files
• Graphify only activates when source files exceed 20, and the graph can be reused after generation
• SQL uses local SQLite and does not require an extra database service

Together, they cost far less than asking AI to reread the entire project from scratch every time.

Q: Which platforms are supported?

Claude Code and Codex are Primary. ZCode, DeepSeek Harness, Qoder, and Trae are Beta. OpenClaw and Cursor are Compatibility. The exact support boundary lives in config/runtime-compatibility-catalog.json.

Q: Is the installation complicated?

One command is enough:

npx --yes github:KimYx0207/Meta_Kim meta-kim

Or clone the repository and run:

git clone https://github.com/KimYx0207/Meta_Kim.git
cd Meta_Kim
npm install
node setup.mjs

The wizard will guide you through language, platform, and installation scope.

Q: Why is it called "Meta"?

In Meta_Kim, meta means the smallest governable unit. A valid meta unit must:

• Own one clear class of responsibility
• Define what it refuses
• Be reviewable on its own
• Be replaceable
• Be safe to roll back

Not everything deserves to be called meta. Only what meets that bar counts.

Q: How does this project relate to MCP?

Meta_Kim uses MCP (Model Context Protocol) to expand the capability boundary of agents. Through the .mcp.json configuration, agents can call external tools and services. But Meta_Kim itself is not an MCP server - it is a governance framework, and MCP is only one of its integrated tools.

Further Reading

• README.zh-CN.md
• AGENTS.md
• CLAUDE.md
• config/contracts/workflow-contract.json
• canonical/runtime-assets/cursor/rules/meta-enforcement.mdc

Third-party Dependencies

Meta_Kim itself is licensed under Apache License 2.0. The following optional skill repositories are installed separately via node setup.mjs — each repo's license applies independently.

npm Dependencies

Optional Skill Repositories

meta-skill-creator is installed only for its currently declared targets: Claude Code at ~/.claude/skills/meta-skill-creator, and Codex at ~/.agents/skills/meta-skill-creator with a synchronized compatibility copy at ~/.codex/skills/meta-skill-creator. Cursor and OpenClaw are not current installation targets for this skill.

Optional pip Packages

License

This project is licensed under the Apache License 2.0.

Commercial Use and Attribution

Commercial use is allowed. If you redistribute Meta_Kim or substantial portions of it, keep the LICENSE and NOTICE files with your distribution.

Recommended attribution:

Meta_Kim by KimYx0207 — https://github.com/KimYx0207/Meta_Kim

Attribution must not imply endorsement by KimYx0207 or the Meta_Kim project. Third-party dependencies and optional skill repositories keep their own licenses.