Built-in workflows

Discover issues analyzes only selected traces. Normalization and signal extraction supply evidence to the managed brain, which diagnoses and groups findings with confidence and uncertainty. Each issue has an identity independent of audits, with multiple occurrences and linked repair tasks. Repeated analysis does not duplicate an occurrence. Discovery ends with issues for selection.

Fix accepts an issue, trace or description. Confirm ownership and expected behavior; ask when ambiguous. Prepare a regression check, reproduce the original failure, author a repair in an isolated worktree, and independently review reference and candidate evidence. Direct execution uses at most three repair attempts and one bounded ledger. Existing accepted regression goals remain required gates. Missing reproduction or verification cannot establish a resolved issue.

Optimize searches within an accepted goal, evaluator, baseline, scope and budget. Keep frozen comparisons, failed and dominated attempts, independently reviewed final measurements, and user choice among verified alternatives. Evaluation preparation and baseline measurement use the same capabilities as Fix. Full Audit preserves the entire audit scope and rubric when requested.

The runtime owns one task contract and durable lifecycle for every workflow. Managed brain sessions reason and edit; validated Python code authorizes execution and accepts evidence. Improvement memory is advisory. A verified result records the tested revision and explicitly distinguishes production recovery. No workflow automatically publishes, merges or deploys.

Production feedback

Assess project scans bounded code and selected traces, activates detected agent bindings, discovers issues, and recommends next actions. Observe production acquires a saved monitor window, measures accepted criteria, and optionally diagnoses new evidence. Both use ordinary tasks; neither starts repairs. See the production guide for limits, comparison rules and scheduling.