Frequently Asked Questions (FAQ)

Does CDAL send any code to external servers?

No. All commands (cdal validate, cdal codegen, cdal audit, cdal trace, cdal selftest) execute 100% locally on your machine. @cdal/core has zero runtime dependencies and makes zero network requests.

How much overhead do the PreToolUse hooks add?

Hook performance is strictly benchmarked at < 150ms p95 per invocation. Because hooks run directly via Node.js standard library with compiled regex patterns, agents do not experience noticeable latency when editing files.

Why not just put rules in CLAUDE.md or AGENTS.md?

LLMs are probabilistic. Empirical benchmarks demonstrate that in multi-turn sessions with large diffs, agents ignore markdown instructions in 20% to 40% of complex tasks—especially when tests fail and the agent faces pressure to "make it pass". CDAL couples markdown instructions with physical exit-2 hooks and git phase diff gates.

How does CDAL prevent agents from bypassing @ts-ignore rules?

The CDAL guard engine includes gate-evasion detection. If an agent tries to construct forbidden tokens using string splitting (e.g. ['@ts', 'ignore'].join('')) or double casting (as unknown as Target), the probe engine catches the evasion immediately.

Can CDAL be used without Claude Code?

Yes. CDAL supports Cursor, Codex, OpenHands, and generic IDE environments via AGENTS.md, .cursor/rules/, and the cdal phase <contract|red|green> --base origin/main gate in your CI pipeline.