OpenClaw #113434 — Codex sessions.reset ID Reuse & Gateway RAM Exhaustion

1 · Problem Summary

Two independently actionable regressions collide in one incident:

2 · Root Cause Analysis

F-1: O(n) scans with no coalescing

Every Control-UI tick issues a fresh catalog/file listing. Each call is a full scan over the Codex session store. With a large store, scans overlap — a new scan starts before the previous finishes. Result: monotonic CPU/RAM growth → Gateway-wide unresponsiveness → OOM crash. Overlaps with #113171 / #110796 but is independently fixable.

F-2: Reset retires the generation, not the ID binding

sessions.reset invalidates the current Codex generation but hands the client back the same session ID. The next turn presents the old generation token bound to that ID and is rejected:

Codex session generation is no longer current: <session-id>

This is a classic idempotency-boundary bug: the reset operation has a side effect (generation bump) that is not reflected in the handle the caller keeps using.

3 · Fix Plan (with code)

F-1 → Cache + coalesce scans

// pseudo-patch: sessions catalog service
const cache = new TTLCache({ ttl: 2000 });   // 2s TTL
let inflight = null;

async function listCatalog() {
  if (cache.has('catalog')) return cache.get('catalog');
  if (inflight) return inflight;             // coalesce concurrent callers
  inflight = scanStore().then(r => {
    cache.set('catalog', r); inflight = null; return r;
  });
  return inflight;
}

F-2 → Mint a new generation handle on reset

// sessions.reset should return the NEW binding
function reset(sessionId) {
  retireGeneration(sessionId);
  const next = mintGeneration(sessionId);    // new generation counter
  return { sessionId, generation: next };    // client rebinds explicitly
}

F-3 → Debounce UI polling

// Control UI: debounce catalog refresh to ≥2s, pause when tab hidden
const refresh = debounce(fetchCatalog, 2000, { leading: true });

4 · Stopgap Checklist (until patched)

• ☐ Restart Gateway between heavy Codex sessions to clear scan accumulation
• ☐ Cap the Codex session store size (archive old sessions out of the scanned dir)
• ☐ Watch RSS: if memory climbs while the UI is idle, you are hitting F-1
• ☐ After any sessions.reset, start a new session instead of continuing the old handle (avoids F-2 entirely)
• ☐ Close unused Control-UI tabs — each one multiplies the polling load

5 · Prevention — the ARK pattern

All three findings belong to failure classes ARK guards against in agent systems:

pip install ark-trust        # Python
npm install @feilunxitong/arkit   # TypeScript

6 · Community Validation (updated 2026-07-28)

After this report was published, upstream work independently confirmed both critical findings — with merged code, not opinions:

Credit: PollyBot13's current-main update mapped these merged fixes back to the findings above. The fix shapes match this report's recommendations (coalesce + cache for F-1; explicit generation-handle rebind for F-2).