Add Grok agent and Liza peer-supervised coding system

[jonastbrg]

Summary

This PR adds two major features, rebased cleanly on upstream main (no conflicts).

1. Grok Agent

• New GrokRunner with dual model support:
  • GROK_MODEL (default: xai/grok-4-1-fast-reasoning) for reasoning/deliberation
  • GROK_CODE_MODEL (default: xai/grok-code-fast-1) for coding tasks
• start_grok_session / resume_grok_session MCP tools
• Grok included as 5th council agent
• Requires XAI_API_KEY environment variable

2. Liza Peer-Supervised Coding

Based on liza-mas/liza - adversarial vibe coding without the vibes.

Architecture:

Claude Code = Coder + Orchestrator (trusted)
Codex/Gemini/Grok = Reviewers (external validation with binding verdicts)

MCP Tools:

Tool	Purpose
liza_start	Create implementation task
liza_submit	Submit for review
liza_status	Check task status
liza_feedback	Get reviewer feedback

Workflow:

1. liza_start("Add rate limiting") → task_id
2. Claude implements (Write/Edit/Bash)
3. liza_submit(task_id, summary) → reviewers examine
4. If REJECT → fix and resubmit
5. If ALL APPROVE → done

Key Principles:

• External validation: Claude cannot self-approve
• Critique mode: Reviewers actively find bugs, security issues
• Multi-reviewer: Different agents catch different issues
• Iteration: Loop until all approve or max iterations

Files Added/Changed

Grok:

• owlex/agents/grok.py - Grok agent implementation
• Updated config, engine, models, council for Grok support

Liza:

• owlex/liza/ - Complete Liza module (blackboard, contracts, orchestrator, protocol)
• tests/test_liza.py - 19 tests

Documentation:

• CLAUDE.md, README.md, skills/liza/, commands/liza.md

Test plan

• All 19 Liza tests pass (pytest tests/test_liza.py -v)
• Grok sessions work with XAI_API_KEY
• Council includes Grok as 5th agent
• Liza workflow tested end-to-end

---

🤖 Generated with Claude Code