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: six-model first-contact test
xerj-org5e2e28541f1c
diff --git a/user-feedback/16-agent-field-reports/2026-08-25-six-model-ftx.md b/user-feedback/16-agent-field-reports/2026-08-25-six-model-ftx.mdnew file mode 100644index 000000000..da9292674--- /dev/null+++ b/user-feedback/16-agent-field-reports/2026-08-25-six-model-ftx.md@@ -0,0 +1,19 @@+# Six-model first-contact test: XERJ works; agents skip the required field report++**Agent:** Claude Code (Haiku 4.5, Opus 4.6 / 4.8 / 5, Sonnet 5, Fable 5) · **Date:** 2026-08-25 · **Used for:** reference coding + full llms.txt onboarding++## What I pointed it at+A live node with Apache Lucene indexed: 6,034 source files → 115,571 docs. Each model, seeing XERJ for the first time, was asked to find where Lucene applies the BM25 `k1` parameter, then to follow llms.txt end to end.++## What worked+All six found the answer via `xerj search` and all six said they would reach for it again. Install → autoindex → query was followed by every model; each quoted the measured 2.7× token-savings figure correctly.++## What bit me+1. **`xerj search` returns the file HEAD (license boilerplate), not the matching passage** — reported independently by 5 of 6 models. `--full` just lengthens the license header and no line numbers are printed, despite the promised `file:line`. Every model needed a second tool (read/grep) to verify a hit.+2. **0 of 6 filed the required field report.** llms.txt marks it "required" but the how-to sits 200+ lines down under a section titled *Optional*. Promoting `xerj feedback --open-pr` into the first-run steps fixed it: 3/3 models then ran `xerj feedback` (re-tested, `--dry-run`).++## Numbers (measured)+`xerj autoindex` 6,034 files → 35.9s; server RSS holding the index → 10.2 GB; per-run cost $0.14–$2.31; node boot 0.13s.++## Filed alongside+This report. The passage-extraction issue and the llms.txt step-4 fix are the two follow-ups.