[gmickel]

Why

After a fix, QA no longer re-runs every scenario. In discovery, before it plans scenarios, QA prints any earlier QA receipt for the spec. When the spec already has a QA receipt from an earlier head on the same branch, QA diffs the trees between that head and HEAD. It re-drives scenarios that failed or were blocked, scenarios on touched surfaces (including their shared dependencies), and one smoke scenario, and carries every other pass with the head and evidence it was captured at. flowctl stamps live rows with the current head and copies carried rows from the prior receipt, so a carried result can never claim the new head.

What changes for a user or operator

Second and later QA passes on a spec can skip scenarios a fix could not have changed. The receipt now lists each scenario as live at this head or carried from an earlier one, and coverage gains a `carried (@​sha8)` state. Without a usable prior receipt, or after a broad shared-code diff, a changed acceptance contract, a failing smoke run or unclear dependencies, QA runs in full as before.

Scope

13 files changed; +437/-27 lines; 4 generated, 3 mechanical.

1. Why re-check

A fix after a QA pass used to trigger a full re-run of every scenario. Check that the carry rules match the spec's R1-R4 and that unproven R5 sits in Open items.

2. What re-runs and what carries

Check that QA looks for the prior receipt in Phase 1, before planning scenarios, and that recheck.md re-drives everything a fix could reach and falls back to a full run whenever in doubt.

• added plugins/flow-next/skills/flow-next-qa/references/recheck.md : New gated reference: prior-receipt match, tree diff, full-run triggers, what re-runs, smoke first, how to record carried rows. [R1, R2, R4]
• plugins/flow-next/skills/flow-next-qa/workflow.md : Adds §1.4 (print the prior receipt, then read recheck.md before deriving scenarios), the carried coverage state, SHIP counting carried R-IDs, and the scenarios payload. [R1, R3]

3. Receipt guards in flowctl

Check that live rows always get the current head, carried rows are copied (never trusted from the payload), and every refusal case errors.

• plugins/flow-next/scripts/flowctl.py : qa receipt accepts scenarios; _qa_receipt_scenarios stamps live rows, copies carried ones; _qa_prior_scenarios refuses missing, other-spec or other-branch receipts. [R3]
• plugins/flow-next/tests/test_artifact_writers.py : Two behaviour tests: carried rows keep their original head and evidence; failed, absent, same-head and cross-branch carries are refused. [R3]
• plugins/flow-next/docs/flowctl.md : Documents the scenarios field, carry copying and refusals. [R3]

4. Changelog and measured result

Check that the entry states the measured result and its limit, not a general speed-up.

• CHANGELOG.md : Unreleased entry describing the re-check and the study numbers, including that the saving did not clear the noise bar. [R5]

5. Mirrors and flow state

Generated Codex copies and manifest, plus the spec and task records; skim.

3 mechanical files; 4 generated files

Coverage: R1 → groups 1, 2; R2 → groups 1, 2; R3 → groups 1, 2, 3; R4 → groups 1, 2; R5 → groups 1, 4

Blast radius

Read recheck.md first: it decides what carries. Then the qa receipt writer in flowctl.py (`_qa_receipt_scenarios`, `_qa_prior_scenarios`) and its two tests. The reference is gated, so QA without a prior receipt reads nothing new. Codex mirror copies and MANIFEST.json are regenerated. Unproven: the time saving (see Open items); QA did not run on this PR (pipeline.qa is off).

Verification

• Implementation review: Codex: NEEDS_WORK (2 kept findings), both fixed, SHIP; Phase 1 change reviewed separately, SHIP
• Full suite: 4054 tests, 0 failures at adf78d6; focused QA and prose tests green after 1237d52
• Lint: ruff 0.16.0 clean
• Codex mirror: sync-codex --check clean
• R5 regression recall: Planted regression caught in 26/26 valid draws, both arms, three passes
• fail: R5 time, shared-file fix: 133 s vs 147 s mean (N=3); not confirmed
• fail: R5 time, one-file fix: 113 s vs 150 s mean (N=5); spread 38 s > gap 36.8 s: INCONCLUSIVE
• fail: R5 time, Phase 1 check: 125 s vs 145 s mean (N=5); gap 20.2 s < 29 s bar: NOT CONFIRMED
• unverified: QA: not run: pipeline.qa is off

Tradeoffs

Decisions taken unattended: (1) R1 against R2: a changed acceptance contract now always runs QA in full, even when the changed criterion maps cleanly to one scenario (review finding, fixed in 7e4ec8d). (2) The R5 study's pre-registered 120 s gap assumed 15-30 min QA runs; a full pass took about 2.5 min, so a secondary 20%-of-baseline gap was logged as an amendment before any arm result. The primary rule stayed the verdict. (3) After pass 1 failed on a fix that touched shared code, a second pass was pre-registered on a one-file fix, the shape where carrying should help most. Pass 1's result stands. (4) The re-reviewer marked R5 met from the changelog numbers; I record it as not met, because the pre-registered rule says so. (5) On the maintainer's call after pass 2, the prior-receipt check moved from the end of Phase 2 to Phase 1 (1237d52) and a third pass was pre-registered and run; a first pass-3 run that loaded no plugin was discarded as a harness error. (6) QA skipped: pipeline.qa is off. Completion review skipped: one task covering R1-R5 that reached SHIP. Rejected alternative: carrying on a changed-file match alone, which misses shared dependencies.

Open items

R5 is NOT MET by its pre-registered rule in any of three passes, so landing is a product call. Private study, Sonnet, a synthetic storefront where a full QA pass takes about 2.5 min, 26 valid draws. Recall held in every draw: the re-check never missed the planted regression. Time: shared-file fix 133 vs 147 s (N=3), not confirmed; one-file fix 113 vs 150 s (N=5), inconclusive because one draw skipped the re-check; one-file fix with the Phase 1 check 125 vs 145 s (N=5), gap 20 s below the 29 s bar. The Phase 1 check fixed the skipped step (5/5 draws used the prior receipt); two draws still chose a full run after finding a new P1, which the reference allows. The scenarios that carry here are the cheapest to drive, so the saving is small on this app; larger apps are unmeasured. Options: land with the measured result as stated, or park.

https://claude.ai/code/session_01JByA3xCfmV1eQumDVW2R3J

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 35a32fb. Configure here.}