control-metalayer-loop（控制元层循环）

Control Metalayer Loop

Use this skill to initialize or upgrade a repository into a control-loop driven agentic development system.

What To Load

• references/control-primitives.md for the control model and minimal control law.
• references/rules-and-commands.md for policy/rules and command governance.
• references/topology-growth.md for repository topology and scale path.
• references/wizard-cli.md for command usage.

Primary Entry Point

Use the Typer wizard:

python3 scripts/control_wizard.py init <repo-path> --profile governed

Profiles:

• baseline: minimal harness and command surface.
• governed: baseline + policy/commands/topology + control loop + metrics + git hooks.
• autonomous: governed + recovery/nightly controls + web and CLI E2E primitives.

Workflow

1. Baseline current repo workflows and constraints.
2. Initialize baseline metalayer artifacts.
3. Add control primitives and governance rules.
4. Audit and close gaps.
5. Iterate based on run outcomes and metric drift.

Step 1: Baseline

• Identify canonical test/lint/typecheck/build commands.
• Identify high-risk actions requiring policy gates.
• Identify required observability IDs for agent runs.

Step 2: Initialize Metalayer

Run:

python3 scripts/control_wizard.py init <repo-path> --profile baseline

This creates stable operational interfaces:

• AGENTS.md, PLANS.md, METALAYER.md
• Makefile.control and scripts/control/*
• docs/control/ARCHITECTURE.md and docs/control/OBSERVABILITY.md
• CI workflow for control checks

Step 3: Add Control Primitives

Run:

python3 scripts/control_wizard.py init <repo-path> --profile governed

This adds the core control plane:

• .control/policy.yaml
• .control/commands.yaml
• .control/topology.yaml
• docs/control/CONTROL_LOOP.md
• evals/control-metrics.yaml

For a fully self-sustaining loop:

python3 scripts/control_wizard.py init <repo-path> --profile autonomous

Adds:

• scripts/control/install_hooks.sh + .githooks/*
• scripts/control/recover.sh
• scripts/control/web_e2e.sh
• scripts/control/cli_e2e.sh
• .github/workflows/web-e2e.yml
• .github/workflows/cli-e2e.yml
• tests/e2e/web/* + playwright.config.ts
• tests/e2e/cli/smoke.sh
• .control/state.json
• .github/workflows/control-nightly.yml

Step 4: Validate

Run:

python3 scripts/control_wizard.py audit <repo-path>
python3 scripts/control_wizard.py audit <repo-path> --strict

Treat audit failures as blocking until corrected.

Step 5: Operate And Grow

• Keep command names stable (smoke, check, test, recover).
• Keep E2E command names stable (web-e2e, cli-e2e).
• Keep policy and command catalog synchronized with actual behavior.
• Track control metrics and adjust setpoints deliberately.
• Prune stale rules/scripts/docs to prevent entropy growth.

Adaptation Rules

• Do not overwrite existing project conventions without explicit reason.
• Prefer wrappers and policy files over ad-hoc command execution.
• Make every major behavior observable and auditable.
• Keep human escalation rules explicit and easy to trigger.

Related Skills

• agent-consciousness — Architectural synthesis of how the control metalayer, knowledge graph, and conversation logs form a persistent consciousness for agents.
• knowledge-graph-memory — Bridge script that transforms Claude Code conversation logs into Obsidian-compatible session documents, creating episodic memory for the knowledge graph.