plat 092

← Pulse platform issue index

PLAT-092 — answered operator decisions are never applied or consumed, so the decision loop dead-ends at answered

Coordination	Value
Assigned agent	unassigned
Ticket state	implemented — structured application contract and conservative legacy handling shipped; runtime reverify pending
Last synchronized	2026-08-25

• Priority: P1 — the operator answers a question and nothing happens. This is the platform silently discarding human decisions, which is worse than not asking: it spends the operator's attention and returns nothing.
• Owner: Pulse Review+Fix decision drain (pulse-review-fixer.md), and the routing of answered decisions to a suppressed module
• Found by: strategy_auditor on its first pass after being re-enabled (rtslatency, 2026-08-12) — "rtslatency's decision loop has stopped at 'answered' — five operator answers, three of them this module's own, sit unconsumed for 5-7 days while no plan step reads them". social-media's Gate reported the same shape independently the day before.

Scale

26 answered decisions across 6 workflows have never been consumed. The oldest was answered 31 days ago.

workflow	stranded
tectonicusadaytrading	8
instagram	5
rtslatency	5
social-media	4
hetznerssh	3
linkedin	1

By the module that asked:

source	stranded
pulse	12
goal_advisor	8
strategy_auditor	6

Two overlapping causes

1. No stage is instructed to apply or consume. pulse-review-fixer.md tells Review+Fix to "First inspect current-run results, the complete active retained backlog, answered decisions, awaiting-verification work…" — inspect, and nothing further. mark_human_input_consumed appears in zero Pulse operational guidance. Its only mentions anywhere are system/workflow-tools.md (a tool catalogue, where the instruction is merely permissive: "A later Pulse run may apply an approved proposal … call mark_human_input_consumed") and system/chief-task-report.md. So every Pulse pass re-reads the same answered decisions, reports them as outstanding, and moves on — which is exactly the 5-to-31-day loop observed.

2. Answered decisions are routed to a module that cannot be selected. pulse-review-fixer.md states "Goal Advisor is selected only for its own blank-sheet opportunity, answered decision, healthy-headroom, or experiment-checkpoint trigger" — an answered decision is a Goal Advisor trigger. But goal_advisor has been suppressed at the Gate for the whole core-system verification phase, so the designated handler can never run. strategy_auditor was suppressed alongside it until 2026-08-11.

That combination is self-inflicting: 14 of the 26 stranded decisions were asked by the two modules that were then disabled, orphaning their own answered questions. The Gate's suppression note already warns that disabling a lens reduces new finding volume; nobody accounted for it also stranding decisions the operator had already answered.

Fix shipped

pulse-review-fixer.md now carries an explicit drain contract: every answered decision must reach a terminal state in the pass that sees it — applied and consumed, or explicitly re-parked with a stated reason and a next-check boundary. Silence is no longer an option, and the drain does not depend on goal_advisor being selectable, since the module that asked a question may be suppressed while its answer is still valid.

The consumption call requires a real outcome_summary, so a consumed decision always records what was actually done with it.

2026-08-25 routing correction — human approval is not an implementation spec

The first drain contract still gave an agent a decision's operator-facing prose and asked it to infer the repair. That is not safe for prompt, plan, route, validation, database, tool, or cross-artifact changes. It also made it impossible for the scheduler to know whether a failed pre-run repair should block the upcoming run or leave the verified old plan in place.

report_human_inputs now persists a reviewer-supplied apply_contract:

• no_change — record the operator's rejection/defer outcome;
• direct_apply — apply one already-defined, bounded setting change;
• targeted_fixer — start a dedicated, scope-bounded pre-run Fixer with the linked issue, approved scope, required checks, post-run proof boundary and explicit failure policy; or
• external_wait — retain the answer until an external prerequisite exists.

Unknown legacy prose has no automatic mutation path. A rejected targeted_fixer contract is recorded as no change, never dispatched as a repair. The existing stable namespace for prompt-contract consolidation decisions is migrated once to targeted_fixer; this includes the outstanding Social Media prompt-bloat approval without pretending arbitrary old decisions have a machine-readable scope.

Not fixed here

• The 26 existing stranded decisions. They need a pass per workflow to apply or re-park each one; the guidance change only stops the backlog from growing. Several are months-old proposals whose premise may no longer hold, so they need judgement rather than a bulk sweep.
• The 2 corrupted social-media rows (pulse-opportunities-schema-narrowing-2026-07, pulse-db-shape-decisions-2026-07-21) that carry status='answered' and a populated consumed_at. That impossible state is what PLAT-077 fixed going forward; these two predate the fix and were never repaired.
• Whether goal_advisor should return. Its suppression is deliberate (core-system verification) but it is the designated handler for answered decisions, so leaving it off while decisions accumulate has a cost that is now measured rather than assumed.

Acceptance

• A decision the operator answers is applied and consumed, or re-parked with a reason, in the next Pulse pass that sees it.
• No answered decision sits unconsumed across more than one Pulse pass without a recorded reason.
• Suppressing a module does not strand decisions it already asked.
• An approval whose consequence is a nontrivial workflow repair has an explicit application contract; it is never guessed from the prose shown to the operator.