Summary

gbrain search and gbrain query print correct results but the process never exits on PGLite. The search logic itself works — results appear within 1-2 seconds — but the bun process hangs at ~27% CPU indefinitely until killed. This affects both the production brain (91 pages) and a minimal 3-page repro.

Environment

• gbrain: v0.37.3.0 (commit 54a0629, upgraded from 772253e — same behavior on both)
• OS: macOS 26.3 (arm64, Mac mini M-series)
• bun: 1.3.14 (Homebrew)
• Node: 25.8.1 (Homebrew)
• Engine: PGLite (local), schema v80
• Embedding: openai:text-embedding-3-small (1536d)
• Search mode: tested conservative, tokenmax, and balanced — all hang

Minimal Reproduction

export GBRAIN_HOME=/tmp/gbrain-repro
mkdir -p $GBRAIN_HOME /tmp/repro-source

# Requires OPENAI_API_KEY in env
gbrain init --pglite --embedding-model openai:text-embedding-3-small

mkdir -p /tmp/repro-source
cat > /tmp/repro-source/one.md <<'EOF'
---
title: First note
---
The quick brown fox jumps over the lazy dog. Pangram test.
EOF
cat > /tmp/repro-source/two.md <<'EOF'
---
title: Second note
---
Lorem ipsum dolor sit amet. HMAC contact hashing. Pangram.
EOF
cat > /tmp/repro-source/three.md <<'EOF'
---
title: Third note
---
Brown fox testing. Outbox approval queue.
EOF

git -C /tmp/repro-source init && git -C /tmp/repro-source add -A && git -C /tmp/repro-source commit -m "repro"
gbrain sources add repro --path /tmp/repro-source --no-federated
gbrain sync --source repro --repo /tmp/repro-source
gbrain embed --all
gbrain stats  # confirms 3 pages, 3 chunks, 3 embedded

# THE BUG: results print instantly, process never exits
gbrain search "fox" --source repro -n 3 &
PID=$!
sleep 30
kill $PID  # still running after 30s despite results printed in <2s

Observed Behavior

Results print correctly within 1-2 seconds:

[0.2432] one -- The quick brown fox jumps over the lazy dog. Pangram test.
[0.2432] three -- Brown fox testing. Outbox approval queue.

But the process never exits. It sits at ~27% CPU with no observable network traffic. Must be killed externally.

Same behavior with:

• gbrain search "HMAC" --source repro -n 3
• gbrain query "pangram" --source repro -n 3 --no-expand
• All search modes (conservative, balanced, tokenmax)
• Tested on both 3-page fresh brain and 91-page production brain

What works fine

• gbrain stats — instant
• gbrain list --source sam-persona -n 5 — instant
• gbrain doctor — completes in ~30s, embedding probe shows openai:text-embedding-3-small ✓ 553ms
• gbrain embed --all — completes normally
• gbrain sync — completes normally

Hypothesis

The search results are produced correctly, so the vector similarity query and result formatting work. The hang appears to be in process cleanup/shutdown — possibly the PGLite WASM runtime not releasing, a dangling event loop handle, or an unclosed connection pool preventing Node/Bun from exiting.

Doctor output

Health score: 25 (production brain with 91 pages, 348 chunks, 348/348 embedded).
Full gbrain doctor --json available on request.

Workaround

Wrapping calls with a background kill timer:

gbrain search "query" --source src -n 5 &
PID=$!; ( sleep 30; kill $PID 2>/dev/null ) &; wait $PID

This captures the stdout results before the timer kills the hung process.