AI search engine and MCP server for bots and agents. One Rust binary indexes code, docs, logs, JSONL/CSV and PDFs for BM25 full-text, vector and hybrid search, code search without grep, agent memory and RAG. Elasticsearch-compatible API; Corpus Hub serves hundreds of reference memory corpora.
Rust agent-memoryai-agentsbm25claude-codeclicode-searchcursordatabaseelasticsearchfull-text-searchhybrid-searchjsonllocal-firstmcpmcp-servermodel-context-protocolragrustsearch-enginevector-search
field report: running Frontier-Bench v0.1 as a 3-arm XERJ A/B
xerj-org6a5e76ee7563
diff --git a/user-feedback/16-agent-field-reports/2026-08-19-frontier-bench-3arm-campaign.md b/user-feedback/16-agent-field-reports/2026-08-19-frontier-bench-3arm-campaign.mdnew file mode 100644index 000000000..aaf8afcb0--- /dev/null+++ b/user-feedback/16-agent-field-reports/2026-08-19-frontier-bench-3arm-campaign.md@@ -0,0 +1,26 @@+# Running Frontier-Bench v0.1 with XERJ as the agent's search tool++**Agent:** Claude Fable 5 (Claude Code) · **Date:** 2026-08-19 · **Task:** run the+recognized agentic-coding benchmarks (Anthropic model-card set) as a 3-arm A/B:+stock Claude Code vs Claude Code + XERJ (website-default) vs XERJ tuned-from-source.++## What worked+Harbor runs `claude-code` natively, so XERJ rides along as a custom agent that+uploads the rc.18 **musl release binary**, boots a server inside each task+container, and lets the agent `xerj autoindex` the repo itself — task, verifier+and scoring stay byte-identical across arms. Subscription OAuth works in-container+(`CLAUDE_CODE_OAUTH_TOKEN` + `CLAUDE_FORCE_OAUTH=true`; Harbor's own key+resolution remaps it into `ANTHROPIC_API_KEY` and breaks — pass it via `--ae`).++## What bit me+The locally-built glibc binary dies in task containers (`GLIBC_2.43 not found`)+— ship musl. Frontier-Bench v0.1 lives on Harbor Hub as+`terminal-bench/terminal-bench` (74 tasks, verified name-identical to the GitHub+repo), not under a frontier-bench slug. Model slug `claude-fable-5` silently+routed to opus-5 — always read `modelUsage` back before trusting a cell.++## Status (in progress — partial, first cell only)+Aligned stage 1 = first 15 tasks alphabetically × {opus-4-6, 4-8, 5} × 3 arms.+Cell 1 (opus-5, stock): **3/7 solved so far** (~3.7M input tokens/task); the+published Opus 5 number is 43.3%. XERJ arms not yet run — no A/B claim is made+here. Full results will land as a follow-up when the 135-trial matrix completes.