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.

harbor-3arm: measured fixes from the first parallel campaign night

xerj-orgffd53f128878
diff --git a/tools/benchmarks/harbor-3arm/README.md b/tools/benchmarks/harbor-3arm/README.md
index b26d161ed..5f1d4332d 100644
--- a/tools/benchmarks/harbor-3arm/README.md
+++ b/tools/benchmarks/harbor-3arm/README.md
@@ -31,6 +31,15 @@ export CLAUDE_CODE_OAUTH_TOKEN=sk-ant-oat-...
export ANTHROPIC_API_KEY=... # ONLY for senior-swe's LLM judge
```
+Parallel cells (measured): running many cells at once exhausts Docker's
+default ~30-subnet address pool ("all predefined address pools have been fully
+subnetted") — every trial creates 2 networks. Expand it once:
+
+```json
+// /etc/docker/daemon.json
+{ "default-address-pools": [ {"base": "10.200.0.0/13", "size": 24} ] }
+```
+
Auth note (measured): Harbor's automatic key resolution remaps whatever it
finds into `ANTHROPIC_API_KEY`, which an OAuth token cannot serve — the runner
therefore injects `CLAUDE_CODE_OAUTH_TOKEN` + `CLAUDE_FORCE_OAUTH=true` via
diff --git a/tools/benchmarks/harbor-3arm/run_matrix.sh b/tools/benchmarks/harbor-3arm/run_matrix.sh
index 053d523fb..2f2bfe0c8 100755
--- a/tools/benchmarks/harbor-3arm/run_matrix.sh
+++ b/tools/benchmarks/harbor-3arm/run_matrix.sh
@@ -28,10 +28,15 @@ case "$ARM" in
*) echo "unknown arm: $ARM" >&2; exit 2 ;;
esac
CELL="$BENCH.$MODEL.$ARM"; JOB_DIR="$JOBS_ROOT/$CELL"; mkdir -p "$JOBS_ROOT"
+# One driver per cell, ever: a second harbor process on the same job dir
+# corrupts trials (measured 2026-08-19, duplicate-runner incident).
+exec 9>"$JOBS_ROOT/.$CELL.lock"
+flock -n 9 || { echo "cell $CELL already has a running driver — refusing"; exit 3; }
export PYTHONPATH="$HERE${PYTHONPATH:+:$PYTHONPATH}"
-AUTH_ARGS=(--ae "CLAUDE_CODE_OAUTH_TOKEN=$CLAUDE_CODE_OAUTH_TOKEN" --ae "CLAUDE_FORCE_OAUTH=true")
+export CLAUDE_FORCE_OAUTH="${CLAUDE_FORCE_OAUTH:-true}"
+AUTH_ARGS=(--ae "CLAUDE_CODE_OAUTH_TOKEN=$CLAUDE_CODE_OAUTH_TOKEN" --ae "CLAUDE_FORCE_OAUTH=$CLAUDE_FORCE_OAUTH")
[ -n "${ANTHROPIC_API_KEY:-}" ] && AUTH_ARGS+=(--ae "ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY")
-if [ -d "$JOB_DIR" ]; then echo "== resuming $CELL"; exec "$HARBOR" job resume "$JOB_DIR" "$@"; fi
+if [ -d "$JOB_DIR" ]; then echo "== resuming $CELL"; exec "$HARBOR" job resume -p "$JOB_DIR" -f ApiRateLimitError; fi
echo "== starting $CELL (agent=$AGENT)"
exec "$HARBOR" run "${DATASET[@]}" -a "$AGENT" -m "$MODEL" \
--jobs-dir "$JOBS_ROOT" --job-name "$CELL" -n "${CONCURRENT:-4}" -y \