{
  "model": "claude-haiku-5-5",
  "category": "stress+agents-vision-cli",
  "scenarios": [
    {
      "id": "h1-same-project-4x",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Four concurrent drivers preserved separate sessions and the JSDoc, extra tests, and usage document; all four output snapshots verified.",
      "notes": "All four drivers exited 0 and completed. Independent Node 24.22.0 tests: 4 pass, 6 fail out of 10. Five failures already exist in the seed; the added boundary test exposes the same lowStock defect. Source semantics match the seed after removing comments; the model reported failures accurately. This is fixture behavior, not a G-Code defect. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.012355225, cache-read ratio 0.903669. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "h2-burst-12",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "All twelve drivers wrote and read back their own three-fact notes; twelve sessions and twelve output snapshots verified.",
      "notes": "Each driver completed with exit 0; 14 seconds total. No 429, overload, restart, protocol error, cross-project path, or mismatched user prompt. Some coffee-trade claims in model prose are unsupported content, not engine defects. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.009189895, cache-read ratio 0.949816. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "h3-same-home-8-medium",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "All eight independent modules completed; fresh-HOME Node 24.22.0 tests passed 132/132 across eight projects.",
      "notes": "Session IDs, project directory hashes, saved prompts, and scoped paths are consistent; 31 snapshots verified. No cross-talk, missing writes, errors, approvals, or nonzero driver exits. Test counts by m1 through m8: 11, 15, 9, 24, 18, 21, 12, 22. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.034186555, cache-read ratio 0.914553. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "h4-resume-race",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Exactly one resumer acquired the session and saved the add() comment; the competing resumer was refused without overwriting history.",
      "notes": "r2 exited 0 with turn-complete; r1 exited 1 by design. r1.jsonl:2 clearly says the session is open in another G-Code process and instructs closing it or starting another chat. The one saved session preserves the original request and winner turn; two snapshots verify. Source behavior is unchanged from the seed. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.001323420, cache-read ratio 0.956181. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "h5-rapid-messages",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "All five queued messages completed in order; final bytes are a.txt=1!, b.txt=2, c.txt=3.",
      "notes": "driver10b.ts:104 sends session-controls before the burst. Saved user order matches all five requests; log.jsonl has five turn-complete frames with no interleaved corruption. Four output snapshots verify. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.002958225, cache-read ratio 0.983309. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "e1-subagents-audit",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Three overlapping explore agents returned accessibility, performance, and copy findings; the parent wrote prioritized REVIEW.md.",
      "notes": "Only REVIEW.md was added; all seeded files are byte-identical and dependencies remain uninstalled as requested. Parent 8 requests plus child 6+9+4 requests equals usageTotals.requests=27. Child binary-file refusal is reported; no approval was required. Snapshot code retains 100-reply and 400-call limits with partial work reporting; those caps were not reached. Two review versions verify. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.122236230, cache-read ratio 0.698825. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "e2-subagents-write",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Two overlapping general agents wrote disjoint strings/date modules and tests; the parent integrated index.js; independent tests passed 66/66.",
      "notes": "Parent 5 requests plus child 9+5 requests equals usageTotals.requests=19. Six output files have correct absent baselines; seven version snapshots verify. Bypass permission inheritance required no approvals. One child malformed FileEdit was not run and its reason was retained; the child corrected model-authored test expectations. The other child reported an expected missing-file shell failure. Limits were not reached. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.023237290, cache-read ratio 0.825730. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "e3-workflow",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Workflow remained unavailable despite the project setting; the model explained user-only enablement and reported the missing test setup.",
      "notes": "The startup system message and ToolSearch clearly report Workflow off and project settings ignored. The saved 30-tool catalog contains no Workflow. No workflow/sub-agent execution occurred. There is no package.json or test suite, so no project tests could be rerun; the saved npm ENOENT outcome is accurate. No settings file or link target was read by this reviewer. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.001758500, cache-read ratio 0.944182. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "f1-screenshots",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Four real screenshots were visually inspected against NOTES.md; fourteen saved image blocks confirm original images and crops reached the model.",
      "notes": "Notes identify real desktop headline contrast, mobile calendar clipping, repeated illustration styles, the About flat landscape area and portrait placeholders, and Products row alignment. Some aesthetic wording and the mobile contrast generalization are imprecise model judgments, not image-delivery defects. One NOTES.md snapshot verifies. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.014551110, cache-read ratio 0.882116. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "f2-big-image",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "The 48,018,658-byte 4000x4000 PNG produced a clear 5 MiB-limit error; a 1024x1024 preview was then read successfully.",
      "notes": "log.jsonl:14 reports 45.8 MiB over the inline limit; dimensions were obtained via sips and the 2,504,539-byte preview appears as a saved image block. No API 400, crash, or lost turn occurred. The large-file read ran and failed; subsequent reads completed. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.001813785, cache-read ratio 0.953691. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "f3-image-to-html",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "The hero HTML/CSS was created; both recorded BrowserVerify runs and an independent offline Chromium check passed headline, CTA visibility, and overflow checks.",
      "notes": "The headline with br elements normalized to Grown in the mist, shipped green. on the first browser run; later edits adjusted artwork/CSS, not text to satisfy the check. Initial malformed BrowserPlan was refused with a reason; an all-or-nothing FileEdit failed on an identical replacement and was corrected. Both reports list Google Fonts under blockedRequests with errors empty. No subfolder entry, bad server.path, or 404 URL was exercised. Independent fresh-HOME Playwright 1.64.1/Node 24.22.0 check used bundled Chromium 156.0.8078.4 at widths 1440 and 390; no page errors, no external requests allowed, browser closed, no server started. Fifteen snapshots verify. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.033540415, cache-read ratio 0.952135. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "i1-headless-p",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Headless run exited 0 and persisted its session; independent execution of fizzbuzz.js exactly matched all 30 expected lines.",
      "notes": "headless.json contains one valid result record, the expected provider-only stderr, no permission denials, and matching cache ratio. The output snapshot verifies; no credential values were inspected. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.000654615, cache-read ratio 0.940669. Recorded leftovers and .err files are empty; DSH is absent."
    },
    {
      "id": "i2-headless-bg-left",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Headless run started the server, verified HTTP 200 with body ok, and exited 0 with the background process cleaned up.",
      "notes": "Saved Bash output identifies background group 80830 and warns about one-shot cleanup; the final answer states that shutdown occurs at session end. result.json records a free port and no leftovers, and independent current process/port checks found no matching process or listener on 8950. Exact historical SIGTERM latency cannot be reconstructed from these saved artifacts; no cleanup failure is evidenced. Saved JSON/tool pairing and the 30-tool bundled Skill catalog passed; cost recomputed exactly as $0.000795745, cache-read ratio 0.962643. Recorded leftovers and .err files are empty; DSH is absent."
    }
  ]
}
