cao-learning

Call report_outcome once per completed step or work item with a task label, success flag, workflow name, and optional score and friction notes. Report after validation, keep friction notes to 1-3 conclusion-only sentences, and never paste logs or secrets.

The retrospector reads outcomes with list_outcomes and stores worker-craft lessons via store_lesson with a target agent profile, not memory_store. It stores 0-3 lessons per retrospection, each under 400 characters and ending with an Applies when clause.

If report_outcome or a memory tool returns disabled: true, skip it silently and continue the task. Learning is opt-in and often deliberately off for control runs, so a disabled response is expected behavior, not an error.

Workers scan the injected cao-memory block and any Learned Patterns section before starting, applying lessons whose Applies when clause matches the current task. New discoveries are stored immediately with memory_store under a stable key, and corrections overwrite the same key.

Dispatch the retrospector at natural boundaries such as a completed package, feature, or review cycle, not after every step. If no retrospector profile is available, skip the handoff entirely.