0061: Mission planners prefer independently observable slices

Status: superseded by ADR 0064 (2026-08-09) · Scope: product tier (governor planning)

Context

A governor plan already carries a mission objective, milestone acceptance, bounded work items, dependencies, immutable workflow templates, independent review, and explicit approval. Those contracts make a plan durable and safe to execute, but they do not by themselves encourage a useful implementation order.

A planner can produce a valid graph that groups all storage work, then all service work, then all API work, and leaves the first observable integration until the final item. That shape delays feedback and makes a poor early design choice expensive to reverse. A plan can also leave module boundaries, key types, or call paths implicit, pushing consequential choices into the implementation turn where review is costlier.

Decision

• Governor planner prompts prefer the smallest observable end-to-end slice that can be verified before later work begins when work crosses system layers.
• Layer-only work remains valid when it has independent acceptance evidence or is a necessary prerequisite.
• Work objectives and acceptance criteria should expose consequential contracts, module boundaries, key types or call paths, and the observation that proves the slice.
• Independent plan reviewers treat deferred observable integration and implicit program-shape decisions as blocking when the candidate gives no reason for them. Revision prompts carry the same guidance.
• Harnery does not add a second design-document schema. Existing mission, milestone, work, review, and proof records remain the authoritative seams.

Alternatives considered

• Add required product, architecture, program-design, and delivery objects to every plan proposal. Rejected because many recovery plans and small missions do not need that ceremony, and field presence would not prove design quality.
• Enforce verticality from the dependency graph. Rejected because Harnery cannot infer semantic system layers or observability from work titles and edges. A deterministic check would produce false confidence.
• Leave implementation order entirely to each planner profile. Rejected because the failure shape is general and the governor owns the common planning contract. Host profiles may add domain detail without restating the baseline.

Consequences

• New planner and reviewer runs receive a compact, source-neutral quality rule; existing persisted proposals and schemas remain readable.
• The rule improves the odds of early feedback but is not advertised as a conformance guarantee. Review and acceptance evidence still decide whether a specific plan is good.
• No new state, migration, public type, or command is introduced.