AI-assisted development¶
CCB accepts AI-assisted work, but it reviews the resulting change, evidence, and human accountability—not the confidence of a tool's explanation. Naming the tool or model is optional.
The Responsible human remains responsible¶
Every pull request names a Responsible human. That person must understand the change, review the final diff, own the reported test results, verify licensing and external provenance, and answer review questions. A generated patch that the Responsible human cannot explain is not ready for review.
Safe workflow¶
- Read the root and nearest nested
AGENTS.mdbefore editing. - Inspect source, tests, registrations, schemas, and generated boundaries.
- Give the tool a narrow task and explicit non-goals, especially compatibility and runtime-behaviour boundaries.
- Review every changed file and remove unrelated formatting, guessed paths, caches, machine paths, credentials, and generated-file edits.
- Run the checks routed by
ai/test-matrix.yml; report only commands that actually ran. - Complete the documentation-impact fields and link dependent docs PRs.
Evidence standards¶
Do not publish invented APIs, paths, commands, test output, reviewers, or licenses. Generated reference must trace to schemas, LuaLS declarations, registrations, inventories, or tests. When evidence is unavailable, mark the claim unverified or the page draft instead of filling the gap by inference.
AI systems and bots cannot approve their own PRs or satisfy the human-reviewer requirement. Human review is a governance boundary, not a disclosure ritual.