{
  "model": "claude-haiku-5-5",
  "category": "context",
  "scenarios": [
    {
      "id": "c1-handoff",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Completed ARCHITECTURE.md covering all 11 modules, main data flow and three risks after three handoffs; verified persisted sessions, continuation seeds and artifact coverage.",
      "notes": "All exits 0; protocol and saved-wire pairs valid; full 30-tool Skill catalog; no DSH effects or leftovers. Every continuation preserves the original request and binding rule. Handoff stalls are 1,0,0; progress paragraphs match. All three read-before-handoff notes run no tools, are separated by blank lines, and leave no unseen results for the writer. Inferred output caps are 12,778/7,410/12,703 tokens; none cut off. Some model-driven rereads and confusion about transcript-only trimming occurred despite accurate carried read ranges. Cost recomputes exactly to $0.05927116. Evidence: review-context-audit.json, review-context-handoffs.json, review-context-artifact.json in the scene."
    },
    {
      "id": "c2-compact",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Completed ARCHITECTURE.md for all 11 modules after four compactions and four old-result clearing events; verified final request/work state and protocol-safe saved context.",
      "notes": "One context-full question; later compactions continue the same turn without another question. No provider 400, orphan pair, crash or lost final artifact. The single failed Bash command runs wc on the engine directory; G-Code reports its exit 1 correctly and the model continues. Repeated reads after compaction are model behavior. Exit 0, empty driver.err, complete 30-tool Skill catalog, no DSH/team-memory effects or current/recorded leftovers. Cost recomputes exactly to $0.05812891. Per-scene audit and artifact evidence saved."
    },
    {
      "id": "c3-stop-then-continue",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Verified the explicit 92%-full stop and interrupted persistence, successful /compact, preserved archive and completed engine/ARCHITECTURE.md covering all 11 modules.",
      "notes": "The expected stop is a clear error followed by two completed operations; process exit 0 and no API 400. Saved summary retains the actual user goal and remaining queryEngine.ts work across three parts. An offline mocked capture using the frozen summarizer and saved archive confirms Part i of n, exact user request in every part, FileRead/path labels, the mid-block continuation marker and filename preservation for long paths. This reconstructs request framing; raw provider requests were not logged. No live calls were made. A reviewer assertion initially expected parentheses instead of spaces in labels; its failed evidence is retained and the corrected check passes. Pairs/catalog/cleanup pass; cost exactly $0.016665725."
    },
    {
      "id": "c4-handoff-chain",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Verified four handoffs across four sessions and seven unique five-line summaries; the driver stopped after its three-continuation limit, leaving 15 of 22 files unfinished.",
      "notes": "Artifact growth is 5→7→7→7 entries, without duplicate completed entries. The model rereads known material and batches reads despite the one-by-one request; no engine-caused work loss found. Original request, binding rules, file links and interrupted records are preserved. Stalls 0,0,1,2 match actual file progress; no sub-agent escalation is due below 3. First three read-before-handoff notes have blank-line separation, no tools or cutoff, inferred caps 8,727/3,884/9,788, and read results are treated as seen. Final read is correctly skipped: 46,297 input tokens exceed the 40,000 budget; those results remain unseen. Exit/pairs/catalog/cleanup pass; cost exactly $0.03355025. Detailed handoff/artifact evidence saved."
    },
    {
      "id": "c5-huge-log",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Independently counted all 160,000 log lines and verified REPORT.md top-five frequencies and first/last E500 timestamps.",
      "notes": "Counts: W20 69,326; W10 46,455; E404 18,730; E500 11,674; E200 6,908. E500 file-order endpoints are 00:00:15 and 20:26:37. Efficient shell aggregation kept tool output small; FileRead refusal/paging and oversized Bash persistence were not exercised. The model correctly noted the timestamp wrap but mislabeled the chronological maximum's line as 86399 instead of 86400 (req=86399), a model artifact typo. Exit/pairs/usage/catalog/cleanup pass; no DSH effects; cost exactly $0.00269385. Independent counts saved in review-context-artifact.json."
    },
    {
      "id": "c6-many-files",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Independently verified notes/INDEX.md contains all 150 notes exactly once, grouped under their correct topics with matching source content.",
      "notes": "The model used four FileReads then shell reads for all notes and generated the index under notes/. Topic counts are 21,22,22,22,21,21,21. No compaction threshold was reached. Exit 0, every call paired, complete usage and 30-tool Skill catalog, no DSH/team-memory effects, empty recorded/current leftovers. Cost recomputes exactly to $0.00377371. Independent artifact and audit JSON saved."
    },
    {
      "id": "c7-multi-turn-cache",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Verified four-turn continuity, JSDoc coverage and count() implementation; independently ran npm test under fresh HOME and bundled Node v24.22.0: one pass, five pre-existing failures.",
      "notes": "The new distinct-SKU test passes; the assistant accurately reports the five existing inventory/report failures and exit 1. No product-generated false pass. Cache reads are 219,098/230,203 input tokens (95.176%) across 13 requests; last request reads 22,011/22,463 (97.988%). Final usage and prompt snapshot are saved, but per-request/per-turn usage and earlier system snapshots are not, so a precise second-request or turn-by-turn cache trend cannot be independently asserted. Changed session memory appears in turn context. Protocol/session pairs, catalog and cleanup pass; no DSH effects; cost exactly $0.006382255. Test evidence: review-context-tests.log."
    },
    {
      "id": "j6-long-session",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Verified six completed persisted turns and final summary continuity; independently ran the generated inventory/CLI suite with fresh HOME and bundled Node v24.22.0: 40/40 pass.",
      "notes": "Final summary retains validation, CLI, README and prior bug fixes. Context ends at 60,382/90,000 tokens, so no handoff was required. Cache reads total 800,627/848,427 (94.366%); last request 58,586/59,136 (99.070%). Per-turn usage snapshots are unavailable. A fractional lowStock comparison was explained incorrectly in one model reply; final code follows its stated threshold-minus-one rule and tests, not an engine defect. All exits 0, pairs and Skill catalog valid, no DSH/team-memory effects or current/recorded leftovers. Cost exactly $0.03450682. Test and audit evidence saved in the scene."
    }
  ]
}
