0062: Governor plans infer their final outcome root

Status: accepted (2026-08-08) · Scope: product tier (core/governor planning)

Context

A governor work graph points from later work to its earlier prerequisites. Its root is therefore the final outcome item: traversing that item’s dependencies must reach every other proposed item. Planner output also had to name that item in a separate root field.

That extra choice was redundant and easy to reverse. In ordinary planning language, “root” often means the first foundation. A planner could produce the right ordered dependency graph, name its first item as root, and have the whole proposal rejected because later items were not reachable from that key. The validator was enforcing Harnery’s graph correctly, but the model-facing representation made a valid decomposition fail.

Decision

Remove root from the schema for model-authored apply proposals. Harnery now infers the root after validating the ordered work array:

1. Dependencies still name active work or an earlier proposed key.
2. Harnery finds proposed keys that no other proposed item depends on.
3. The graph must have exactly one such final outcome.
4. That outcome must use a root-capable frozen template.
5. Traversing its proposed dependencies must reach every proposed item.

The planner prompt states the dependency direction and asks for one final outcome that depends transitively on the rest. Reviewers receive the inferred root in the normalized candidate and the same meaning in their goal context.

The durable GovernorPlanProposal remains schema version 1 and still stores root. Plan records, review receipts, digests, application, projections, and public readers therefore retain one canonical materialized shape. Only the model-authored input stops duplicating information already present in the DAG.

Alternatives considered

• Clarify the old root field with more prose. Rejected because the field still asks the model to encode the same relationship twice and permits the two answers to disagree.
• Rename root to final_outcome in a new durable schema. Rejected because inference fixes the model boundary without migrating plan records, review receipts, readers, and public types.
• Reverse dependency direction. Rejected because immutable work already uses “this item depends on these prerequisites” throughout Harnery.
• Repair a wrong root in a second model call. Rejected for this failure because deterministic inference is cheaper and cannot disagree with the DAG. Semantic retries remain appropriate for graph defects that cannot be derived.
• Choose the last array item unconditionally. Rejected because array order alone does not prove that parallel branches converge on one outcome.

Result and gotchas

• A prerequisite-first plan no longer fails merely because the planner calls its foundation the root; the model does not author that field.
• Multiple terminal branches still fail closed. A planner must add a bounded final outcome that joins them or return attention.
• The inferred final item must use a root-capable template. Template authority remains frozen by the governor and is not widened by inference.
• Terminal complete and attention decisions keep their existing empty root and work fields. They do not carry an executable graph.