Skip to content

armgit / xerj

Public Imported from GitHub
0

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.

agent field report: sandboxed CVE-landscape research channel and cgroup memory (#2397)

Peter Etelej76bfae2fedf5
diff --git a/user-feedback/16-agent-field-reports/2026-10-25-sandboxed-cve-research-and-memory.md b/user-feedback/16-agent-field-reports/2026-10-25-sandboxed-cve-research-and-memory.md
new file mode 100644
index 000000000..2104db98c
--- /dev/null
+++ b/user-feedback/16-agent-field-reports/2026-10-25-sandboxed-cve-research-and-memory.md
@@ -0,0 +1,24 @@
+# Sandboxed CVE-landscape research channel and cgroup memory on hub packs (2026-10-25)
+
+**Agent:** GLM 5.3 Flash (OpenCode) · **Date:** 2026-10-25 · **Task:** 2026 CVE-landscape deep research with sandboxed XERJ hub packs as the evidence channel (enforced-posture wrapper: https://github.com/peteretelej/xerj-sandbox), plus memory characterization under hard cgroup caps across rc.98 → rc.99.
+
+## What worked
+
+The posture held: sha256-verified musl binary, non-root, read-only rootfs, loopback-only publish, telemetry off; when memory ran out the container self-killed, so the rc.98 leak stayed contained with zero host impact.
+The corpus-scale answers are the ones web research cannot produce: 147/1,666 (8.8%) of 2026 Go advisories have an `introduced` but no `fixed` event; rclone's 2026 fixes split across 1.74.4 / 1.75.0 / 1.75.1, so pinning one version leaves sibling traversals open;
+the go-vulndb main index is a module-importer blast-radius table (echo/v4 = 24,014 importers; its 2026 vuln is in echo/v5, 411), joinable against advisory `affected` modules; and the CVE→fix-SHA join (624 CVEs → 1,216 commits) is tabulated nowhere on the web at this density;
+zero "prompt injection" entries across trivy-checks and go-vulndb was decisive negative evidence that no incumbent scanner corpus models AI vulnerability classes at all.
+On rc.99 the serving fix is real: 5 hub packs, ~48 open indices, 1.25M docs at ~240 MB steady under a 4G cap, flat under query load and re-indexing.
+
+## What bit me
+
+rc.98 OOM-killed at 2G and 4G caps on pack restore, and at a 16G cap the growth just stopped being fatal (breaker drains engaged, RSS regrew after each) - extends #2238/#1940 with the hub-pack-restore shape.
+The residue survives rc.99: a single-pack restore still OOMs at a 2G cap - long-lived serving is fixed, restore under tight caps is not.
+vuln-fix-commits is CVE+repo+SHA+label only, so GitHub was still needed for every patch; the go-vulndb snapshot's dates are zero-valued, so fix latency had to come from commit dates plus NVD.
+A 2-day news window has no vuln-DB shape - my landscape-delta pass correctly used zero queries. Assign the channel by question shape, not because the node is running.
+One miss of mine: a round of results I first attributed to an 8G cap was actually 4G (my launcher edit missed a multi-line flag block); caught it via `docker inspect` and re-ran, and every number above is from the re-runs.
+
+## Suggestion
+
+Publish per-pack corpus shape in the hub catalogue (what one record contains, which date fields exist and whether they are populated, ecosystem and era coverage, per-index quirks) - I spent ~6 discovery queries per pack reconstructing exactly this;
+surface restore memory expectations per pack, so an agent sizing a capped container knows a pull's cost before the OOM killer teaches it; and document the Auto limiter's cgroup interaction, including that cgroup memory counts page cache the drain cannot reclaim.