Eval harness for sandboxed code-orchestration agents (Cloudflare Code Mode and similar "write code, don't call tools" patterns).
v0.1 quantifies the core claim: describing tools as JSON schemas grows linearly with tool count, while a single code tool with on-demand search is a near-constant baseline.
python -m codemode_eval token-budget tools.json
# tools=2500 schema_dump_tokens=... codemode_tokens=1000 reduction=99.x%- Network-isolation probe (assert generated code cannot
fetch()) - Cross-model code-generation reliability benchmark
MIT.