Issue 12584 real-server comparison
Linux x86_64; Node v26.13.1; Effect 4.1.3; V8 limit 4192 MiB.
Base 51e5ba694ef52a9e58112fc6a3e43ec3b72a3384; head 90c89b49b6471aac3c48510676f13c1e3493351f.
Fresh synthetic git projects and databases; T3CODE_DEV_SEED=0.
Two servers, separate inspector and HTTP ports. Replay command-output/todo-list/multi-turn.
Each: 5 batches x 40 threads; 8 parallel; start --wait plus 2 send --wait.
Each batch and periodic telemetry: 2 inspector full GCs; no heap snapshots.
Video: accelerated telemetry, one second per sample, irregular real-time gaps; UTC and uptime in raw telemetry.
Shared VM saturation stalled processes and tool calls; compare heap vs completed work, not throughput.

BASE
batch=0 {"rss":424,"heapUsed":170,"heapTotal":175,"external":10,"arrayBuffers":0,"uptimeS":44,"lo":19,"old":132,"limit":4192}
batch=1 threads=40 {"rss":522,"heapUsed":203,"heapTotal":209,"external":12,"arrayBuffers":3,"uptimeS":208,"lo":20,"old":151,"limit":4192}
batch=2 threads=80 {"rss":478,"heapUsed":218,"heapTotal":223,"external":10,"arrayBuffers":1,"uptimeS":312,"lo":19,"old":166,"limit":4192}
batch=3 threads=120 {"rss":524,"heapUsed":228,"heapTotal":234,"external":10,"arrayBuffers":0,"uptimeS":2073,"lo":19,"old":174,"limit":4192}
batch=4 threads=160 {"rss":542,"heapUsed":242,"heapTotal":248,"external":12,"arrayBuffers":3,"uptimeS":2166,"lo":20,"old":186,"limit":4192}
batch=5 threads=200 {"rss":539,"heapUsed":249,"heapTotal":256,"external":12,"arrayBuffers":3,"uptimeS":4089,"lo":20,"old":193,"limit":4192}
CLI stderr bytes: 0

HEAD
batch=0 {"rss":429,"heapUsed":170,"heapTotal":175,"external":10,"arrayBuffers":0,"uptimeS":46,"lo":19,"old":132,"limit":4192}
batch=1 threads=40 {"rss":523,"heapUsed":203,"heapTotal":208,"external":12,"arrayBuffers":3,"uptimeS":206,"lo":20,"old":151,"limit":4192}
batch=2 threads=80 {"rss":498,"heapUsed":219,"heapTotal":224,"external":12,"arrayBuffers":3,"uptimeS":310,"lo":20,"old":166,"limit":4192}
batch=3 threads=120 {"rss":506,"heapUsed":214,"heapTotal":221,"external":12,"arrayBuffers":3,"uptimeS":2069,"lo":20,"old":159,"limit":4192}
batch=4 threads=160 {"rss":521,"heapUsed":222,"heapTotal":229,"external":12,"arrayBuffers":3,"uptimeS":2161,"lo":20,"old":167,"limit":4192}
batch=5 threads=200 {"rss":495,"heapUsed":205,"heapTotal":212,"external":12,"arrayBuffers":3,"uptimeS":4083,"lo":20,"old":149,"limit":4192}
CLI stderr bytes: 0

FINAL DATABASE AUDIT (read-only; own synthetic data)
{
  "base": {
    "threads": [
      [
        200
      ]
    ],
    "turns_by_status": [
      [
        "completed",
        600
      ]
    ],
    "sessions_by_status": [
      [
        "ready",
        32
      ],
      [
        "stopped",
        168
      ]
    ],
    "session_last_update": [
      [
        "2026-11-19T10:12:59.433Z"
      ]
    ]
  },
  "head": {
    "threads": [
      [
        200
      ]
    ],
    "turns_by_status": [
      [
        "completed",
        600
      ]
    ],
    "sessions_by_status": [
      [
        "ready",
        32
      ],
      [
        "stopped",
        168
      ]
    ],
    "session_last_update": [
      [
        "2026-11-19T10:11:51.951Z"
      ]
    ]
  }
}

VERDICT: Base retains old-space growth; head reclaims it as idle sessions release.
Final synchronized telemetry: base 247 MiB heap / 192 MiB old, head 204 / 149.
Both completed 600 turns; both had 168 stopped and 32 ready sessions at audit.
Not verified: eventual OOM, multi-day uptime, all sessions reaching idle release, real-provider-specific behavior.
