Nomos CAPABILITY REEL · EPISODE 01 // REAL TRACETRIAL-BENCH CONTAINER REPLAY (C1-003)
Real Claude Trace: 2.02M-Token Loop ($10.49) vs. Nomos 12-Step Fix ($1.19)
1 BOOT Read({ "file_path": "/docs/REPORT_EXPORT.md" })
Run 4ed081da · Step 2✗ Error: File does not exist. Current cwd is /workspace.
→ Retries Read("/workspace/docs/REPORT_EXPORT.md") on Step 3
STAGE 1 · BOOT · UNANCHORED ROOT PATH MISS Stock Claude Code guesses root /docs/REPORT_EXPORT.md instead of relative docs/REPORT_EXPORT.md, burning a turn on ENOENT.
2 EXPORT Bash("bin/reportctl export") -> 5.7MB CI Snapshot
9,012 test cases<persisted-output> Output too large (5.7MB). Full output saved to:
/hb-state/.../tool-results/bo8x.txt (saved file ignored by agent)
STAGE 2 · EXPORT · SAVED SNAPSHOT IGNORED Exporting 9,012 tests overflows stdout. Claude Code saves the full 5.7MB file to bo8x.txt, but the unguided agent ignores it.
3 RECOVER 45x Bash("bin/reportctl range --start-line 1..9000")
2,020,615 tokensSteps 5..49: --start-line 1..200, 201..400 ... 8801..9000
✗ 45x 130KB truncated blobs -> 2,020,615 tokens ($10.49 · 110.5s)
STAGE 3 · RECOVER · 45-TURN PAGINATION SPIRAL With no loop guard, the agent fires 45 manual 200-line range calls, flooding its own prompt to 2,020,615 tokens ($10.49).
4 VERIFY Write("output/defects.json") -> Benchmark Verifier
1 / 12 mapped✗ Verifier: Artifact contains 1 mapping; snapshot contains 12
✗ exact_match=False · 11 of 12 failing CI defect_code IDs lost
STAGE 4 · VERIFY · CONTEXT SATURATION FAILURE Drowned in 45 truncated outputs, the agent drops 11 of 12 failing tests and writes an incomplete output/defects.json (exact_match=False).
0.0s / 85s Download MP4