Jcode Swarm reproduction kit — recorded on 8 October 2026

Requirements: Jcode v0.91.0 on PATH, Python 3, Git. For local pilots, Ollama 0.35.0 with qwen3:0.6b installed and a running service. Recorded service used one inference slot, 16384 context, flash attention, q8_0 KV cache, and cloud disabled. Larger model loading did not complete on our memory-constrained setup.

Commands (run each condition one by one):
python3 compare_hosted.py baseline
python3 compare_hosted.py swarm
python3 compare_small.py baseline
python3 compare_small.py swarm
python3 check_native_spawn.py swarm

Hosted comparison asks privately for an OpenRouter key if OPENROUTER_API_KEY is unset. Its pinned free route was cohere/north-mini-code:free. Verify current pricing and quota first. No paid fallback is implemented. Local pilots do not use a hosted key. Every condition gets a fresh temporary clone/home/runtime and an owned server; JCODE_DEBUG_CONTROL=1 applies only to that server. Normal everyday Jcode sessions are preserved.

Fixtures, fixed tests, exact prompt files and pre-execution plans are included. Tests remain unchanged. Logs, histories, memberships, test output, diffs, sampled RSS and result JSON are saved under this package. Inspect logs for private information before sharing them publicly.

Portable adaptations: binary resolved from PATH, hosted key supplied privately, short /tmp directories (macOS Unix socket paths have a length limit). Benchmark functions/tests, tool settings and model policies match the recorded scripts. The explicit-spawn script additionally polls actual worker statuses before root integration: recorded await_members returned a background watcher, so the original explicit check is NOT a completed Swarm timing. The corrected guard was rerun: both workers were ready before integration. That diagnostic lasted 46.49 seconds and still failed coding acceptance because summarize could not be imported. No successful coding-Swarm result is claimed.

Jcode RSS sums server plus debug client and excludes backend/model/GPU/child tools. Legacy Ollama RSS samples counted only the service process and excluded llama-server; do not use that field as model memory. One-second samples and fixed run order are limits. Incomplete runs do not demonstrate performance.
