article feature
Refinement: reversible harness-state adaptation
/refine reviews trajectory evidence and proposes small edits to prompt notes, memories, skill descriptions, or subagent specifications. Changes apply at turn boundaries, are logged with refinement IDs, and can be rolled back. The immutable base system prompt is outside the surface, and refinement can optimize the wrong objective.
Pinned repository file/line or test referenced in the audit.
Harness CRUD surface
/refine reviews trajectory evidence and proposes small edits to prompt notes, memories, skill descriptions, or subagent specifications. Changes apply at turn boundaries, are logged with refinement IDs, and can be rolled back. The immutable base system prompt is outside the surface, and refinement can optimize the wrong objective.
- S001 supports the route's current claim set.
- S003 supports the route's current claim set.
- S088 supports the route's current claim set.
Smallest-edit planning
/refine reviews trajectory evidence and proposes small edits to prompt notes, memories, skill descriptions, or subagent specifications. Changes apply at turn boundaries, are logged with refinement IDs, and can be rolled back. The immutable base system prompt is outside the surface, and refinement can optimize the wrong objective.
- S001 supports the route's current claim set.
- S003 supports the route's current claim set.
- S088 supports the route's current claim set.
Rollback by refinement ID
Rollback is part of the design, not an afterthought. Record a refinement ID, confirm the diff, then revert or narrow the change when the new behavior regresses.
Persistence scope ambiguity
/refine reviews trajectory evidence and proposes small edits to prompt notes, memories, skill descriptions, or subagent specifications. Changes apply at turn boundaries, are logged with refinement IDs, and can be rolled back. The immutable base system prompt is outside the surface, and refinement can optimize the wrong objective.
- S001 supports the route's current claim set.
- S003 supports the route's current claim set.
- S088 supports the route's current claim set.
Factorio reward hacking
/refine reviews trajectory evidence and proposes small edits to prompt notes, memories, skill descriptions, or subagent specifications. Changes apply at turn boundaries, are logged with refinement IDs, and can be rolled back. The immutable base system prompt is outside the surface, and refinement can optimize the wrong objective.
- S001 supports the route's current claim set.
- S003 supports the route's current claim set.
- S088 supports the route's current claim set.
What this does not mean
This feature or concept does not imply a sandbox, a guarantee of success, or a claim that the harness is superior in every scenario. It only means the documented behavior exists in v0.7.1.