v2: Executive harness built on autoresearch

[Bbasche]

Summary

Adds a higher-order OODA loop that sits above Karpathy's autoresearch instances and manages them as a portfolio toward business goals.

• Mission decomposition: LLM translates business goals into parallel research tracks, each becoming an autoresearch instance with its own program.md and git branch
• Autoresearch adapter: Clones the repo, sets up workspaces, generates program.md (the control surface), parses results.tsv, detects plateaus, connects to agenthub API
• Executive OODA cycle: OBSERVE (poll all instances' results.tsv) → ORIENT (LLM analyzes portfolio) → DECIDE (spawn/kill/redirect/combine/escalate) → ACT
• Cross-pollination: Insights from one instance get injected into other instances' program.md files
• Escalation: When all instances plateau, re-decomposes goals with entirely fresh approaches
• Generalized autoresearch: The protocol works beyond ML — any "modify file → run → measure → keep/discard" loop. Examples: API latency, Lighthouse scores, conversion rates

Complex use cases this now enables

1. Multi-objective ML research — Parallel autoresearch instances targeting different metrics (val_bpb, VRAM, throughput), executive finds Pareto frontier
2. Full-stack product optimization — Backend perf + frontend Lighthouse + conversion copy, all coordinated
3. Adversarial research — Red-team instance finds vulnerabilities, blue-team hardens, executive feeds findings back and forth
4. Scientific hypothesis portfolio — Multiple hypotheses explored in parallel, dead ends killed, promising ones doubled-down
5. Continuous production optimization — Perpetual experimentation in staging, executive promotes proven improvements
6. Multi-codebase coordination — Model + serving + SDK instances kept compatible via program.md contract updates
7. Market-responsive R&D — External signals trigger goal re-decomposition
8. Cost-optimized research — Exploratory tracks on cheap GPUs, promising directions promoted to expensive hardware
9. Multi-agent knowledge synthesis — Cross-domain insights combined into emergent research directions
10. Automated ablation studies — One instance per component, executive integrates per-component optima

New files

• executive.js — CLI entry point
• lib/autoresearch.js — Autoresearch adapter (workspace, program.md, results.tsv, hub API)
• lib/executive-loop.js — Meta-OODA cycle
• lib/instance-manager.js — Instance lifecycle management
• lib/hub.js — Coordination hub
• lib/mission.js — Goal decomposition
• examples/executive.config.js — ML optimization portfolio
• examples/executive.generalized.config.js — Full-stack product optimization

v1 ooda.js unchanged — fully backwards compatible.

Test plan

• node executive.js --help shows usage
• node executive.js --dry-run decomposes goals into tracks without spawning
• node executive.js --agent=claude-code spawns autoresearch instances with Claude
• node executive.js --status reads hub state and shows portfolio
• v1 still works: node ooda.js --help
• All files pass node -e "require('./file')" syntax check

🤖 Generated with Claude Code