Token Optimization & Budget Policy

In agentic coding workflows, token overheads are the single largest source of API latency and cost. MultiModel Dev OS integrates tools and policies to optimize context inputs.

---

1. Caveman Mode

• Goal: Reduce baseline context token footprint by ~79%.
• CLI Flag: node bin/multimodel-dev-os.js init --caveman
• Behavior: Instantiates minimal .caveman.md files for AGENTS.md, MEMORY.md, TASKS.md, and RUNBOOK.md, omitting explanatory files and dense markdown guidelines.

---

2. Context Budget Auditing

• Define strict context bounds inside [.ai/context/context-budget.md](file:///F:/multimodel-dev-os/.ai/context/context-budget.md).
• CLI Command: node bin/multimodel-dev-os.js doctor --tokens
  • Checks if build outputs or caches (node_modules/, dist/, .next/, build/) are missing from .gitignore or are being exposed to LLM scan scopes.

---

3. Policy Guidelines

Keep Checklist Files Short

Use TASKS.md exclusively for active sprint tasks. Move completed or historical backlogs to archive directories or logs.

Model Handoff Compression

When switching from reasoning models to quick-fixing models, use [.ai/prompts/compress-context.md](file:///F:/multimodel-dev-os/.ai/prompts/compress-context.md) to distill findings into high-density prompts.