Artifact Set — The Cache Wars

← Back to the paper · PDF

Raw per-request billing counters, proxy logs, and unmodified runner scripts backing every number in Tables 2–4 of The Cache Wars: An Empirical Study of Prompt-Cache Efficiency in Five LLM Agent Harnesses (AGNT Labs Technical Report v1.1, July 7 2026). Benchmark executed 2026-07-06/07 against claude-sonnet-4-5-20250929 (funded API key) and the Codex default backend.

Integrity. SHA256SUMS.txt lists a checksum for every file below. Sanitization. Seven runner scripts had operator-local credential paths rewritten to environment-relative equivalents (marked [sanitized for publication] inline); no counters, logs, or measurement code were altered. The discarded INVALID- run is retained deliberately: runs that violate guards G1–G3 are kept, not hidden.

Replication. Procedure, pinned versions, and guard definitions: paper §10 Reproducibility. Expect ~$2–5 in API spend for a full five-harness replication.

FileSizeSHA-256Description
Traces & receipts (raw counters — verbatim)
results-e4b-agnt-v2.json2.0 KB9d50ff1f2d8e70a4…AGNT v0.6.4 — Anthropic usage blocks, 5 turns + 390 s idle
results-e1-claude-code.json2.4 KB415df20c94b5f5c7…Claude Code CLI 2.1.126 — JSON telemetry per turn
results-e2-codex.json2.2 KB54378ceab77f669b…Codex CLI 0.130.0 — session-log last_token_usage extraction
results-oc-real.jsonl2.5 KB186b9a3e1dff8336…OpenClaw 2026.5.12 — embedded-agent lastCallUsage (OpenRouter + Anthropic-direct sessions)
results-hermes-anthropic-direct.json1.6 KBfa25aee06f6071c2…Hermes 0.18.0 — Anthropic-direct turns (proxy-verified)
results-omp-default.jsonl1.4 KBad03999639a9fc10…oh-my-pi (OMP) 16.3.11 — Anthropic-direct, API-key default (5 m TTL), agent_end usage
results-omp-1hbest.jsonl0.7 KB6fa54343c69a32d1…oh-my-pi (OMP) 16.3.11 — Anthropic-direct, PI_CACHE_RETENTION=long (1 h TTL), incl. post-pause turn
results-hermes-real-final.json1.5 KB460651698b9ab85a…Hermes 0.18.0 — OpenRouter-path final run
results-hermes-real-turns.json0.4 KBfacfcff9e941b40c…Hermes 0.18.0 — OpenRouter-path per-turn records
results-e4-strategies.json9.3 KB1d417fad3b870052…E4 — controlled identical-content replay arms (marker-strategy comparison)
results-e5-openai.json2.6 KBbdcfaab187600915…E5 — OpenAI automatic-caching observation runs
anth-proxy-log.jsonl3.9 KB84a91d074e6bcfe6…Ground truth: per-call cache_control markers + usage counters, Anthropic path (incl. the §8 subscription-OAuth rejections)
proxy-log.jsonl11.0 KB8466f308ce87ba8a…Ground truth: per-call markers + usage counters, OpenRouter path
INVALID-results-hermes-real-v3.jsonl0.6 KBb09bd2c66fb0bdbd…Discarded Hermes v3 run — failed guard G1 (history not accumulating); retained for transparency
Instruments (proxies, runners, inspectors)
anth-proxy.js3.2 KB3dbf088cd30363b4…Transparent logging proxy: 127.0.0.1:18081 → api.anthropic.com; one JSONL line per /v1/messages call
or-proxy.js3.1 KB0bbe403b1afb31e3…Transparent logging proxy for the OpenRouter path
e1-claude-code.js3.3 KBde27926cbd8a473c…Runner: Claude Code CLI, 4 rapid turns + idle turn
e2-codex.js (credential paths sanitized)4.0 KBf6e9015bbbceb2e7…Runner: Codex CLI via exec/resume, session-log extraction
e4-strategies.js (credential paths sanitized)9.1 KB81fe88d00010134e…Runner: identical-content replay arms (marker strategies)
e4b-agnt-v2.js (credential paths sanitized)6.6 KB0f1984bb66491337…Runner: AGNT arm, Anthropic usage blocks
e5-openai.js3.8 KB890afbe53900ca37…Runner: OpenAI automatic-caching arm
oc-turn.js1.7 KBb4c2b7d6a63967cc…Runner: OpenClaw embedded-agent single turn
hermes-driver4.py (credential paths sanitized)3.6 KB5852c89578c0feaa…Runner: Hermes final driver (OpenRouter path) — growing-conversation protocol with fail-loud guards
hermes-anth-a.py (credential paths sanitized)3.4 KB39a544d85ce28f24…Runner: Hermes Anthropic-direct, Phase A (4 rapid turns)
hermes-anth-idle.py (credential paths sanitized)2.0 KB0686177c3c07bb08…Runner: Hermes Anthropic-direct, idle turn 5 (after 390 s)
hermes-session-inspect.js (credential paths sanitized)1.3 KB97a8958a8cc4c0d5…Inspector: Hermes session-file telemetry fields
package.json0.1 KB2555b0be5064d2c4…Node dependencies (xxhash-wasm for prefix-stability hashing)
AGNT Labs · Technical Report v1.1 artifacts · Published 2026-07-07