Why AI Context Packs Must Adapt Dynamically to Target Model Economics

Different LLMs have wildly different context windows and cost economics. Providing a generic 20k-token context pack is wasteful for smaller models and insufficient for massive 200k+ models.

In TokenCap, we implemented model-aware budgeting in src/bridge/generate.js.

How Token Allocations Adjust

1. gpt-4o (Target: ~16k to 32k tokens): Balanced span sampling, key symbol definitions, priority diffs, and compressed helper functions.
2. claude-3-5-sonnet (Target: ~32k to 64k tokens): Extended call-chain depths, full component prop interfaces, and complete ADR documentation.
3. gemini-1.5-pro (Target: ~100k+ tokens): Deep repository breadth, full implementation bodies across related clusters, and complete test suites.
4. llama-3-8b / Local Models (Target: ~4k to 8k tokens): Aggressive even-span folding, signatures only, zero diffs, highest Louvain cluster focus.

Selecting Profiles via CLI or Bridge

# Generate context tailored specifically for Claude 3.5 Sonnet
tokencap make --profile claude-3-5-sonnet

CLI Output:

Loaded model profile: claude-3-5-sonnet
Token budget ceiling: 64,000 tokens
Allocated:
  - Architecture and graph overview: 2,400 tokens
  - Core targeted symbols: 18,200 tokens
  - Transitive caller bodies: 14,800 tokens
  - Active ADRs and memory: 1,600 tokens
Total Generated: 37,000 tokens (58% of ceiling, 0% overflow risk)

The bridge passes these model hints automatically when queries arrive from our Chrome companion or MCP server.

Read the model profiling guide at tokencap.vansharora.app