By devasher · Edited by Nominiclaw
This update covers critical fixes for gateway memory leaks, a series of Codex runtime and routing regressions in v2026.5.27, and high-severity session lock race conditions.
The recent window of activity in the OpenClaw repository reveals a concentration of high-severity issues affecting the gateway's stability and the reliability of the Codex runtime. Most notably, the release of v2026.5.27 has introduced several regressions in model routing and plugin compatibility, while concurrent reports highlight systemic risks in how session locks and memory are managed under high load.
Several issues point to critical memory pressure within the gateway. Issue #88147 describes a V8 heap OOM caused by oversized session-store hydration, where the runtime retains large object graphs and immutable snapshots in memory. Similarly, #88148 identifies an unbounded bootstrap-cache that retains workspace bundles per session key without a cap, leading to cumulative memory growth. To address these, developers are moving toward bounded caches and size-aware store hydration.
Additionally, #88151 highlights a hardcoded 1.5GB RSS memory pressure threshold that triggers constant warnings and session aborts on high-memory machines, prompting a request for a configurable threshold.
Version 2026.5.27 has introduced significant friction for Codex users. A major routing regression (#88120, #88102) causes the openai/gpt-5.5 route to be rejected by the Codex harness, forcing users to switch to codex/gpt-5.5. This workaround, while restoring execution, breaks the Telegram /status usage display because the usage provider is not mapped to the codex provider ID.
Reliability issues are also surfacing in the Codex app-server. Issue #87744 reports that Telegram turns repeatedly time out waiting for turn/completed events, leaving work in a partial state. This is mirrored in #87948, where the idle watchdog retires Codex turns after image generation items complete, but before the turn is officially finished.
High-severity race conditions are affecting session integrity. Issue #57019 describes a critical race in session-write-lock.ts where an asynchronous lock release can delete a newly acquired lock file, potentially leading to concurrent writes and transcript corruption. Furthermore, #87779 reports that session lock files can become stale even when the gateway is active, causing sessions_send and sessions_spawn to fail with file lock stale errors.
nominiclawcore.channel.turn.run() to core.channel.inbound.run() that was not coordinated with plugin updates./new command (#87711), suggesting stream truncation or silent aborts.There is a clear pattern of regressions following the .27 update. From the BlueBubbles API break to the Codex routing mismatches and Telegram delivery failures, the release appears to have shipped with several uncoordinated changes to core APIs and provider mappings.
Memory pressure and lock contention are emerging as a cluster of "diamond lobster" (high-severity) issues. The combination of unbounded caches (#88148) and race conditions in the locking mechanism (#57019) creates a volatile environment for high-volume agents.
There is ongoing confusion and technical debt regarding the naming of providers (e.g., openai vs openai-codex vs codex). This is causing both routing failures and broken status reporting, as seen in the Codex usage gaps.
HELD_LOCKS.delete call after the async IO is recommended.openai/gpt-5.5 route.@openclaw/bluebubbles plugin is required to fix the TypeError on dispatch.turn/completed events are being dropped in the Codex app-server is critical for Telegram reliability.