llms.txt · git:20261003.02b13ba · 2026-10-03 · sha256 a6349f9124d7127d
llms.txt git:20261003.02b13baA
Immutable. This exact content never changes and is served at /api/v1/blob/a6349f9124d7127d.
# Meta_Kim Meta_Kim is a general agent governance framework: it turns user goals into evidence-backed decisions, capability discovery, bounded ownership, orchestration, outcome verification, and writeback across domains. It coordinates research, operations, learning, writing, software and other tasks through a repeatable spine: Critical -> Fetch -> Thinking -> Execution -> Review -> Meta-Review -> Verification -> Evolution ## Core Idea Runtime agents provide execution capabilities. Meta_Kim governs how they serve the current user's outcome: - clarify the real intent before execution - search available capabilities before picking an owner - compare feasible routes under the user's constraints and state the limits of the recommendation - route work to bounded agents, skills, scripts, MCP providers, or runtime tools - review and verify results with artifacts - distinguish structural checks, routing, native invocation and actual task outcomes - write back useful lessons only after approval ## Start Here - Project overview: README.md - Chinese overview: README.zh-CN.md - Maintainer guide: AGENTS.md - Claude Code guide: CLAUDE.md - First-run proof: examples/first-run/README.md - Launch metadata draft: examples/launch-kit/github-metadata.md Some maintainer worktrees also contain local PRD and research notes under docs/. Treat those as local evidence only when the files are present in the checkout. ## Source Of Truth Durable behavior should be edited in: - canonical/agents/ - canonical/skills/meta-theory/ - canonical/runtime-assets/ - config/contracts/ - config/capability-index/ Runtime folders such as .claude/, .codex/, .cursor/, openclaw/, and .agents/ are projections or adapters unless the task explicitly targets runtime wiring. ## Important Commands ```bash npm run meta:theory:run npm run meta:theory:report -- --run-id latest npm run meta:delivery:bundle npm run meta:release:smoke npm run meta:check npm run meta:verify:all ``` Use `npm run meta:release:smoke` for routine prompt, docs, changelog, version, and narrow governance wording changes. Upgrade to `npm run meta:verify:all` for install, runtime, hook, dependency, package, security, or release-grade changes. ## Public Claims Boundary Safe claims: - Meta_Kim has default formal projections for Claude Code and Codex, plus non-default formal projections for OpenClaw and Cursor. - The local complete-product MVP has repeatable scripts, tests, reports, and AI-readable delivery artifacts. - Capability-first dispatch is the default governance route: define the capability, search owners/tools, then execute. Do not overclaim: - Do not say every formal projection has the same native-live completeness or default install behavior. - Do not turn projection smoke, fixture pass, or generated reports into live runtime proof. - Do not describe the product as course, teacher, classroom, student, or lesson tooling. The product boundary is AI-readable governance and evidence review. ## External Reviewer Prompt When reviewing Meta_Kim, inspect whether a run produces: - explicit intent and success criteria - capability discovery evidence - selected owner, loadout, and verification owner - bounded worker tasks - review and meta-review findings - fresh verification evidence - writeback decision or none-with-reason Prefer artifact evidence under .meta-kim/state/default/ over chat summaries.