Write agent-tooling scripts and tests in TypeScript and execute them with Bun; do not add .mjs or Node entrypoints. Existing SDK languages and package-management commands keep their repository ownership.
Use the verification map to select the existing checks and preserve a task's goal, authority, next step and blockers across sessions.
- For substantial work, initialize one task with the requested outcome and a matching profile. Keep the actual user request authoritative.
- Inspect
statuswhen resuming, updatenextStepandblockerin its localtask.json, and refresh external issue, PR and CI state before acting. - Run the profile after implementation. Read the receipt and its boundary before reporting what passed. A passing profile establishes only its named checks.
- Treat a recurring intervention as a candidate for an existing behavioral test, type or lint check. Prove the check rejects the original mistake. Preserve one-off judgment locally instead of expanding global instructions.
Do not create task records for trivial edits. Private task data and command output
stay under ignored .artifacts/; never attach them wholesale to a PR. This tool
records context and evidence, and does not grant commit, push, merge or release
authority.