# PR1597 sustained team workflow

Tested fbb38632cc4e768f69de48364f53e02d2137c926 on Spark in Fleet session critical-extended-1597-v2 (tmux fleet-sh). Binary SHA256 and timestamp are in 1597-execution.json. Original 1597.cast is 433 seconds; recording ended normally with Ctrl-C after final handoff.

Goal: create a useful read-only pantry shopping-list CLI from a CSV inventory. Manager started a builder and an independent reviewer. The initial stock-report CLI passed nine tests. Actual use revealed that stock counts were less useful than shopping quantities, so the person requested --shopping and negative-stock rejection. Builder implemented the request, reviewer checked it and identified test-strength gaps. The person asked to close those gaps; builder added/pinned tests, reviewer independently verified the final suite and README examples, and manager gave the final shopping command. On request the manager committed exactly README.md, pantry.py and test_pantry.py locally, with no push. Resulting fixture commit: 4011fcf402051c4730cbecfa14ccc298b1b4e20f.

Independent execution on Spark: 19 subprocess tests passed; pantry.csv default output beans:0/rice:2; --shopping prints beans: buy 3 and rice: buy 1; --threshold 5 prints beans: buy 5 and rice: buy 3. Inventory remains byte-for-byte identical to its original Git version. Exact output is in 1597-independent-run.txt; deliverable source is preserved in 1597-artifacts/.

PR-specific coverage: manager creation, child creation, coherent builder/reviewer handles, directives to idle members, automatic member completion notifications and manager wake, multiple follow-up handoffs, approval inheritance persisted as allow for both builder and reviewer while title/model were preserved. 1597-team-traffic.jsonl captures actual team events; 1597-verification.json records metadata and receipt summary.

Model evidence: all182 usage receipts report deepseek/deepseek-v4.1-flash, including low16, talk77, worker27, reflex37, mastermind21 and4 receipts whose seat field is omitted. Total recorded cost $0.235423992. No fallback or alternate model appears in this fresh run. 1597-usage.json is complete receipt evidence.

Limitations: profile explicitly disables model_pool to work around separately tracked CodeAF #1608; this run does not prove --one-model alone pins all roles until that fix is verified. The original failed-model-policy recording and receipts are retained separately and are NOT counted as passing. Live approval coverage is allow/YOLO inheritance, not permission-denied or persistence-failure injection (covered by existing regressions). Manager initially polled impatiently and stopped its own turn before review completion; later notifications woke it and completed the workflow. Recording preserves those rough edges. No new CodeAF defect was identified in this lane. Full CodeAF suite was not repeated for an unchanged revision; prior CI/focused checks remain applicable. Only the small generated project's19 tests were rerun here, exclusively on Spark.

Media: use1597-final_review.cast screenshot for final review result,1597-shopping_request.cast for the natural revision,1597-commit_progress.cast for final local commit. Preserve complete1597.cast; any shortened/sped preview must disclose pacing changes. media_evidence owner publishes accessible media and updates PR body; this file itself is local handoff evidence.