cao-learning

CAO Self-Learning

CAO workflows can improve as they repeat: outcomes you report feed a retrospector agent that distills durable lessons into memory, and those lessons reach future sessions automatically. Your job depends on your role.

All of this is opt-in infrastructure. If report_outcome or a memory tool returns disabled: true, skip it silently and continue your task — learning is off for this run (often deliberately, e.g. a control run) and that is expected, not an error.

A bare error with no disabled key is different: say so, don't skip it. That shape means the tool could not reach a verdict — cao-server is unreachable, or its settings.json could not be read — so learning may well be ON while nothing is being recorded. Mention it in your response and carry on with the task.

If you are a SUPERVISOR

Report an outcome after each meaningful unit of work

One report_outcome call per completed step, delegated task, or work item — after validation/review, not before:

report_outcome(
    task_label="convert package CustomerETL (iteration 2)",
    success=false,
    workflow_name="ssis-migration",
    agent_profile="transformer",           # who did the work (defaults to you)
    score=40,                              # optional 0-100 metric if you have one
    friction_notes="Lookup with partial cache emitted an invalid join; "
                   "improver patched the cache-mode mapping."
)

Rules for friction_notes:

• 1–3 sentences, conclusions only — the root cause, not the story.
• NEVER paste transcripts, logs, stack traces, file contents, or secrets.
• Empty string on a clean pass is fine; the success flag already carries signal.

Report failures faithfully — failed iterations are the most valuable learning signal. Do not skip reporting because a step went badly.

Dispatch the retrospector at natural boundaries

After each completed work item (a package, a feature, a review cycle) — not after every step — hand off to the retrospector agent: