meta-theory

Use the machine-governed Meta_Kim entry for this request:

$ARGUMENTS

Default product runtime path:

When this command is installed globally, run the installed Meta_Kim package root directly:

node "__META_KIM_PACKAGE_ROOT__/scripts/run-meta-theory-governed-execution.mjs" --runtime claude_code --emit-conversation-notice "$ARGUMENTS"

If this command file has not been rendered by global sync and the placeholder is still present, fall back only when the current project is the Meta_Kim source checkout or provides the package script:

npm run meta:theory:run:notice -- --runtime claude_code "$ARGUMENTS"

Then reopen the report with:

node "__META_KIM_PACKAGE_ROOT__/scripts/run-meta-theory-governed-execution.mjs" --read latest

This command is not a prose-only reminder. For explicit /meta-theory requests, project/repo understanding, architecture analysis, commercialization strategy, market or competitor reasoning, governance review, release work, runtime work, or any non-trivial durable task, run the machine entry above or state the exact blocking reason.

Minimum route:

1. Classify the request with the Meta_Kim entry rules.
2. Run Fetch before Thinking. Project understanding must inspect local project evidence such as README, AGENTS, package scripts, canonical/config contracts, capability index, MCP/runtime configs, and Graphify navigation when present.
3. If the answer depends on current market, competitor, platform, provider, pricing, dependency, version, or official documentation facts, prove an available retrieval path or return blocked to Fetch.
4. Use the run artifact and report as the evidence surface. Do not replace the machine route with a short manual summary.

Two host-native phases are mandatory:

• Phase 1 — route and handoff: run the governed Node command only to produce route compatibility and a handoff state. ready_for_host_handoff means Claude Code must perform the next native action; awaiting_native_choice means it must call AskUserQuestion; blocked means it must stop. The runner never authorizes task execution.
• Phase 2 — current-host action: after any required AskUserQuestion answer, use the active Claude Code host's real Agent / Task, tools, and permission surface. Persistent acceptance records are advisory compatibility evidence only. No callback, environment value, or JSON can complete the native choice or authorize this task.

Claude Code execution rule:

HOST-NATIVE FAN-OUT PREFERRED. The main thread is the dispatcher, never the execution worker. The governed runner produces route evidence, workerTaskPackets, and hostInvocationRequestPacket; it does not by itself prove live Claude Code Agent / Task execution. Use Claude Code's native Agent/Task tool directly to fan out independent worker lanes — the runner only records evidence, discovers capabilities, and suggests lanes; it does not enforce dispatch.

• This /meta-theory invocation is explicit user authorization to use Claude Code Agent/Task delegation and parallel worker lanes when Thinking proves the lanes are independent and safe.
• Use the authoritative stage DAG plus native Agent/Task as the sufficient fan-out route. agent-teams-playbook is optional: select it only when it improves orchestration, record not_required for fewer than two ready lanes, and record optional_adapter_not_selected when native fan-out is sufficient. Resolve it from the first available skill root (~/.claude/skills/agent-teams-playbook/SKILL.md, .claude/skills/agent-teams-playbook/SKILL.md, .agents/skills/agent-teams-playbook/SKILL.md, or a configured dependency root). Treat it as a selected fan-out adapter unless a live Skill/Agent Team/Task call is attached.
• Then call the active Claude Code host's real Agent / Task surface for each selected execution worker lane. Every Agent/Task prompt must cite the corresponding workerTaskPackets[].taskPacketId, selected owner/role, output schema, write scope, collision boundary, verification owner, and expected evidence shape.
• Record the returned Agent/Task tool-call id, completion status, and worker-task-to-agent mapping as trusted host invocation evidence. If no callable Agent/Task surface is available, record the checked tool/capability names and blocked reason; do not silently continue as main-thread execution.