View project on GitHub

ADR-044: Idle Recovery of Goals Parked Below the Viability Floor

Context

A prove goal carries an affinity score (⟦Γ:Affinity⟧, ADR-010). select_prove_candidates ranks the claimable pool and _rank drops everything with affinity < TAU_V (−5) as “below viability — awaiting re-decomposition.” A failed prove attempt demotes a goal by −10 (demote_goal), so a goal that fails its first claim’s attempt ladder drops from 0 to −10 in one step — below the floor — and select_prove_candidates can never surface it again.

The intended recovery (ADR-009) is decomposition into sub-lemmas, but decomposition only fires at failure-time inside prove_goal. If that one decomposition does not happen (decompose disabled, the provider produced no usable split, the depth cap, or simply a run that demoted instead of decomposing), the goal is orphaned: parked below the floor, no sub-lemmas, and nothing ever revisits it. ADR-024 cross-cycle lesson memory records a ⟦Δ:Lesson⟧ for each failed attempt, but lessons are only injected when a goal is re-selected for proving — which never happens for a parked goal.

Observed on main (2026-06-15): of 58 phase=prove, status=open goals, 48 were parked below TAU_V (46 at −10, 2 at −20) with no decomposition, while the 10 viable ones all had open prove PRs in flight. The swarm reported “nothing claimable” despite 48 open, recoverable goals — and prove --once runs exited immediately. The pool was starved.

WH(Y) Decision Statement

In the context of a prove queue that drops goals below TAU_V and only ever decomposes a goal at failure-time, while ADR-024 now carries each failed attempt’s lesson forward so a re-attempt need not start from an empty board, facing the fact that a goal demoted to −10 on its first failed claim is parked below the floor with no decomposition and no path back — never re-selected, its accumulated lessons never reused — so the working pool drains to “only in-flight goals remain” and the swarm idles while dozens of open goals sit recoverable, we decided for an idle recovery pass: when a --prove cycle finds no claimable viable goal, re-surface one goal parked below TAU_V (the recovery-candidates selector — the exact inverse of prove-candidates, least-buried first) and run it through the same prove_goal pipeline, which retries with its lessons injected (ADR-024) and decomposes on failure (ADR-009) — no duplicated recovery logic, just a second source for the existing claim+work step (UNSORRY_RECOVERY=0 disables it), and neglected a dedicated re-decomposition sweep that splits parked goals directly (rejected — it duplicates prove_goal’s decompose path and skips the cheaper retry-with-lessons stage the maintainer specifically wanted first), lowering/removing TAU_V so parked goals stay in the normal queue (rejected — the floor’s whole purpose is to keep the normal queue focused on viable work; recovery is the explicit, idle-only exception), and re-promoting parked goals’ affinity in place (rejected — it loses the “these already failed once” signal and lets a hard goal crowd the viable queue every cycle), to achieve a swarm that, instead of idling, spends an otherwise-empty prove cycle giving an orphaned goal a lessons-armed retry and then a decomposition — so the pool keeps making progress until goals are genuinely proved or split, accepting that a recovered goal that fails again is demoted further (−20, …) and the next idle pass takes the next-least-buried goal (bounded: one recovery goal per cycle, HANDLED prevents repeats within a session, and live-claim filtering avoids two agents recovering the same goal), and that recovery deliberately runs only when no viable work exists, so it never competes with normal proving.

Consequences

• Positive. Orphaned goals re-enter the prove pipeline with their lesson history, so the swarm no longer falsely reports “nothing to do” while open goals remain. Reuses prove_goal wholesale — retry-with-lessons (ADR-024) then decompose-on-failure (ADR-009) — with no parallel recovery code path.
• Positive. The “no viable prove work” log now distinguishes in-flight/collided viable goals from the recovery attempt, and --dry-run reports the parked pool it would recover.
• Negative. A genuinely intractable goal will be retried (once per session, lessons growing) and eventually decomposed; until it is proved or split it consumes one idle cycle per session. UNSORRY_RECOVERY=0 opts out.
• Residue. Recovery reduces idle starvation but does not change TAU_V, the demote magnitude, or the decomposition caps — those remain ADR-010/ADR-009 territory.

References

Status History