feat: jason-openclaw example — ClawLaw operating system

[bengii]

Summary

• Adds examples/jason-openclaw/ — a real-world autonomous agent running on OpenClaw under the ClawLaw operating system
• Two disciplines encoded as gitagent rules: atomic work packets (scope-locked units reviewed before any side effect executes) and the compute ladder (model selection by provider health, not task complexity; local-first, paid break-glass only)
• Includes full implementation checklist, 42-test validation harness, and narco-check memory integrity audit skill

What's Included

File	Purpose
agent.yaml	Manifest with ClawLaw config flags, compute ladder chain, budget constraints
SOUL.md	Voice, ClawCode operator loop, scope discipline, budget limits
RULES.md	Five hard ClawLaw rules — packets, ladder, exec integrity, narco-check, scope
skills/packet-factory/SKILL.md	Atomic work packet gatekeeper — DRAFT→APPROVED→DISPATCHED lifecycle
skills/compute-ladder/SKILL.md	5-tier model routing with health checks and cost guard
skills/narco-check/SKILL.md	Memory integrity audit with 5-4-3-2-1 grounding check
workflows/clawlaw-architecture.md	Full architecture doc + phased implementation checklist
knowledge/clawlaw-test-harness.md	42-test suite covering every ClawLaw rule

Why ClawLaw

The two failure modes this solves, verified in production:

1. Scope creep → session death: one instruction expands to 5 tasks → context balloon → compaction failure → model timeout → session dead. Fixed by atomic packets with immutable scope.
2. Paid model overuse: fallback chain routes to expensive models when primary is slow, not broken. Fixed by compute ladder (fallback = 429/timeout only, never "complexity").

Test Plan

• examples/jason-openclaw/README.md renders correctly on GitHub
• agent.yaml validates against gitagent spec v0.1.0
• All skill files have correct frontmatter format

🤖 Generated with Claude Code