# PR1612 acceptance for issue1611

Tested exact f855f2e65 on Spark; full SHA and binary hash in1611-execution.json. Fleet critical-extended-1611-v2, tmux fleet-sh. Original1611.cast ended normally after final workflow; no footage replaced. Reuses the working pantry CLI produced by the sustained1597 team workflow rather than rebuilding an artificial throwaway.

Goal: turn existing pantry.csv into a usable SHOPPING.md note, evaluating whether to run weekly or keep it manual. Asked CodeAF to inspect the tool and propose a Monday9am weekly report. Chose3 Only now on the card. Without any corrective user message, CodeAF read/ran the actual pantry CLI, wrote SHOPPING.md containing beans buy3/rice buy1, verified the CSV unchanged, and accurately said the recurring schedule was not saved. Transcript proves structured run_once_now/pending → bash → write → bash → final.

Natural follow-up corrected inventory date from run date to pantry.csv modification date, and asked to review a tomorrow9am one-time refresh. The new card correctly said wants to schedule work once / Run it then / Runs the work then. Nothing repeats. Chose0 to leave it manual. Final interaction ran19 pantry tests through CodeAF, ran the shopping command, checked the document and listed standing items. Direct independent Spark execution also passed19tests, verified quantities/date convention and original CSV bytes, and found zero saved standing item definitions.

Model: all30 receipts exactly deepseek/deepseek-v4.1-flash, low7/talk13/worker2/reflex7/unspecified1, $0.038865216. Profile model_pool=off is the documented workaround for #1608, pending separate #1610 fix; this does not claim --one-model-alone coverage.

Evidence:1611-once_approved.cast shows actual output after the no-repeat decision;1611-once_task_card.cast shows truthful scheduled-work labels;1611-final_check.cast shows finished report/test/list results;1611-SHOPPING.md,1611-independent-run.txt,1611-usage.json,1611-verification.json and1611-post-approval-transcript.json preserve outcomes. The post-approval transcript includes the first execution BEFORE the later inventory-date correction.

Limits: pending action handoff still relies on the model choosing ordinary tools, with ordinary current-turn permissions and limits; approval receipt is not proof of completion. This live workflow proves actual execution for this request, not guaranteed model behavior for every request. We did not execute tomorrow's task: declined its card after verifying wording. The model's immediate prose after decline still described the card as awaiting approval; actual card settled not set up and final stand list verified none saved. Do not claim new behavior in #1596/#1520/#1560 heads: this is a separate corrective PR1612 supplement for consolidation.

Aggregate boundary: this PR's base has no does.isolate. Ordinary-turn Once does not enforce #1560's new separate-worktree execution. Consolidation MUST suppress Once for isolated work (using the central StandingOnceIsAnAnswer gate, including forged-answer refusal), or provide an actual shared isolated executor and separately test it; merely serializing isolate into the handoff does not enforce it. This recording does not verify an isolated Once. Watch/probe Once is likewise model-mediated and was not covered by this live task workflow.