goal-condition-synthesis

To prevent agent drift, you need a falsifiable goal completion condition grounded in observable artifacts. This synthesizes a testable /goal from plan artifacts to stop multi-session amnesia by referencing specific transcript signals like provenance logs and task statuses.

Generating a testable goal completion condition involves synthesizing a falsifiable /goal from AA-MA plan artifacts. The process validates banned vague terms, ensures references to at least two observable artifacts like git tags or test exit codes, and computes a deterministic condition-hash for auditability.

Yes, you can bound execution costs by applying deterministic turn-cap budgeting. This validates a capped number of turns within a 1 to 30 range based on pending milestones, restricting cost for unattended execution or verify-iterate loops.

Goal synthesis requires active task directories containing AA-MA plan artifacts. It validates references to at least two observable artifacts such as provenance log lines, tasks status fields, git tags, and test exit codes to ensure the completion condition is falsifiable.

No, goal condition synthesis applies only to active task directories containing the required AA-MA files. It needs existing plan artifacts to ground the falsifiable completion condition and prevent vague intent from causing agent looping.