mirror of
https://github.com/garrytan/gstack.git
synced 2026-10-31 14:06:50 +01:00
Garry Tanandcapy-ai[bot]
b0f1e65e0d
* fix(qa-callers): the caller scheduling contract no longer calls a probe read-only discovery (#3153) Census 38212265208 (main @e23f7b807, 2026-10-12) failed ship-exploratory-late-input: "probe preceded resource read: exploratory / system-functional / system-functional". The native events show capture 001 issued in the same provider response as the Reads of ship/sections/plan-completion.md and ship/sections/review-army.md, before the Step 9 section (which holds the "never batch a probe with a Read" STOP) returned and before any QA section was read. The Oct 7 capture (mtb1007a) is the same class. Diagnosis: fixture defect, not product prose. /ship's prose never names the QA capture helper (only qa/sections/exploratory.md does); the fixture prompt supplies it up front, and the fixture's appended scheduling contract (7697fb2, 2026-10-06) told the session to batch "Reads and read-only discovery" while its never-group list omitted Read-result dependencies. The case passed on Oct 3 and Oct 5, before that contract, and failed on Oct 7 and Oct 12 after it. Repair: the contract defines read-only discovery (captures, guard starts/runs and diagnostic-client runs are probes, never discovery), which removes the fixture's own invitation to batch them. It deliberately does not restate the product's Step 9 STOP: the case exists to detect that STOP being lost. The detector (callerProbeReadErrors) is unchanged. Test: the Oct 12 native tool order is replayed inline (no new fixture file, so the paid-runner global inputs stay untouched) with a reordered negative control, and the actual runner callback's contract is checked. Fails on47d639aat the discovery-definition assertion; passes after. Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> * fix(review-army): the simplification case drops its false premise and stages only its section (#3153) Census 38212265208's pass-rate gate flagged review-army-simplification at 159 s of its 180 s session budget (88.1%, above the 85% cap; main-only panel 142/91/146 s). The three Oct 12 transcripts show where the time went: - the prompt claimed "the diff is over 100 lines", but the fixture diff is 68; every trial reconciled that (scope tool, specialist stats, force-include reasoning, coverage-gap prose in the report and final message); - each session read the whole extracted review-SKILL.md and the 18.7 KB checklist, and grepped for a "## Step 4.5" heading the carved fixture no longer has; - the report was unbounded (30-45 s of output composition); trial 3 also dispatched a subagent and re-verified its findings. Repair (fixture work only, budget unchanged at 180 s / 20 turns): the case uses the same stageReviewArmySession/reviewArmyScope staging as perf-n-plus-one and consensus (recorded scope, stats, learnings and diff; only review-army.md and simplification.md staged), selects the specialist with the real /review --simplification force flag, applies the checklist in-session as the original prompt asked, and bounds the report to 30 lines with a brief final acknowledgement. The planted-structure and disavowed-frame assertions are unchanged. Test: review-army-budget.test.ts drives the actual registered callback with a mocked runner (pass control plus two negative controls); it fails on47d639a("Expected to contain: DIFF_LINES: 68") and passes after. Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> * docs(evals): record the #3153 census diagnoses (review-army-simplification headroom, ship-exploratory-late-input) (#3153) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> * chore(release): tentative v1.91.108.0 stamp and CHANGELOG for the #3153 census repair (#3153) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> --------- Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>