0049: Replan exhaustion is attributed to its cause

Status: accepted (2026-07-26) · Scope: product tier (core/governor, governor CLI, dashboard, public package APIs)

Context

A bounded goal spends its frozen replan budget on planner attempts. Two of those attempts fail in ways that look identical once the budget is gone. In one, the planner produces no proposal at all: it returns an attention decision, or its run is interrupted before any candidate exists. In the other, the planner produces a proposal and independent review rejects it across its bounded revision rounds.

Both outcomes leave the goal quiescent, and both consume a charged replan. The projection reported the exhausted goal with a single undifferentiated reason (goal exhausted its N replans, or the latest attention plan’s own text). An operator reading that could not tell a planner that never proposed anything from review-round exhaustion, even though the two call for different responses: the first points at the planner or the goal framing, the second at the reviewers or the acceptance bar. Misreading a no-proposal streak as review-round exhaustion sends the operator to the wrong place.

The distinction was already latent in durable state. A plan carries a review receipt exactly when a proposal was produced and independently reviewed, so the receipt’s presence separates the two causes. Nothing surfaced it.

Decision

Attribute each consumed replan through the existing plan record, with no new record field, event, or planner mechanism:

• A plan with a review receipt whose status is not passed is a reviewer rejection, meaning a proposal existed and review rejected it (revision-round exhaustion, reviewer attention, or review failure).
• A charged plan with no review receipt that ended in attention, interrupted, failed, or retry_requested is a planner no-proposal outcome: the run consumed a replan slot without producing a reviewable proposal. retry_requested is a no-proposal attention the operator asked to replan: the original run still produced no proposal and still spent a charged replan, so it counts here even though a later plan superseded it.
• Uncharged environment/upstream planner failures (ADR 0046) do not spend the replan budget and are attributed to their outside precondition, not here.

Projection exposes an optional replan_consumption breakdown ({ reviewer_rejection, planner_no_proposal }), present only when at least one consumed replan was a planner no-proposal outcome. When present, the exhausted-goal projection.reason (the field itself, not only the rendered row) names the planner no-proposal share, and harn governor list / harn governor show surface the breakdown. The latest plan’s own per-plan reason still leads: a reviewed-and-rejected latest plan reports its review-round exhaustion first, then the reason names the planner no-proposal share of the budget. This is deliberate: a goal whose budget was mostly spent by planners that never proposed must not read as pure review exhaustion just because its final, reviewed plan was rejected. The projected reason a programmatic consumer reads and the operator-visible row must tell the same truth.

A goal with no planner no-proposal history omits the field, keeps the reason it had before, and renders identically in list and show. Budget accounting, cumulative counters, and append-only record authority are unchanged; the attribution is a read-time projection over records that already exist.

Alternatives considered

• Add a new plan event or record field marking the cause. Rejected because the review receipt already distinguishes the two outcomes, and a new field would leave records written before it ambiguous.
• Run a separate planner or reviewer path for the no-proposal case. Rejected because a parallel mechanism duplicates the frozen planner contract and can drift from it; the distinction is an observation, not a new behavior.
• Charge the two causes differently. Rejected because budget semantics are a frozen bound; attribution must not change what a replan costs.
• Infer the cause from the replanning policy (treat any review-configured goal’s exhaustion as review-round exhaustion). Rejected because a review-configured goal can burn its whole budget on planners that never reach review, which is exactly the case this ADR must not mislabel.

Result and gotchas

• Exhaustion by planner no-proposal is never displayed as review-round exhaustion; the reason and the CLI both name the planner.
• The breakdown counts only non-advancing consumed replans. Applied plans that advanced the graph are not counted as either cause.
• The field is additive and optional. Consumers that never read it are unaffected, and goals without planner no-proposal history project unchanged.
• Attribution is derived at read time from durable plan records, so it stays correct after process loss with no migration.