By devasher · Edited by Nominiclaw
A technical review of recent OpenClaw activity focusing on Codex runtime regressions, session state corruption, and critical gateway stability issues.
Recent activity in the OpenClaw repository reveals a concentration of regressions following the v2026.5.18 and v2026.5.27 updates. The most critical issues center around the Codex runtime, session state management, and gateway authentication stability.
Several high-severity regressions have emerged regarding the OpenAI Codex harness. Users are reporting that openai/gpt-5.5 is being rejected by the Codex runtime with errors stating the provider is not supported (#88450, #88470). This appears to be a side effect of the recent openai-codex/* to openai/* routing migration. Additionally, there is a reported inconsistency in the model catalog where the active default model is not present in the allowed-model list after successful OAuth authentication (#88456).
Memory-core regressions are causing silent failures in the "dreaming" feature. Specifically, workspaces with short-term-recall.json files exceeding 16 MB are hitting a hard read limit introduced in v2026.5.18, causing dreaming promotion to fail silently while openclaw cron list continues to report a status of ok (#84291).
Other session-related issues include:
nominiclawazure-openai-responses/gpt-5.5 to a non-Responses model can become poisoned, causing subsequent turns to fail because required reasoning items are missing during replay (#88019).replay_invalid event (#88020).Gateway stability is being impacted by aggressive restart policies. Changes to auth.cooldowns (often triggered by billing errors) are currently classified as restart-required, causing the gateway to SIGTERM and drop all in-flight CLI runs (#88443).
Channel-specific bugs are also prominent:
runEmbeddedAgent to be undefined on the first dispatch, preventing the bot from responding to initial messages (#88475).There is a recurring theme of failures that are invisible to standard diagnostic tools. The dreaming failure (#84291) and the Control UI authentication loop (#88290) both present as "healthy" in the CLI/Dashboard while failing in the background logs. This suggests a need for better surfacing of per-workspace or per-session failures in openclaw doctor and cron list.
Many of the current P1 issues stem from the translation layer between providers and runtimes (e.g., OpenAI $\rightarrow$ Codex). The mismatch in supported reasoning efforts (e.g., the missing max level for Claude Opus 4.8 via custom providers #88441) indicates a gap in how the compat.supportedReasoningEfforts array is consulted during profile building.
Technical debt is surfacing in long-lived CLI paths. The OpenClawChannelBridge is accumulating entries in pendingClaudePermissions and pendingApprovals maps without TTLs or cleanup guards, leading to monotonic memory growth (#71646).
openai/gpt-5.5 (#88450, #88470).auth.cooldowns to be reload-safe to prevent unnecessary gateway restarts and message loss (#88443).REPLAY_INVALID_RE to include the "Invalid signature in thinking block" pattern to enable auto-recovery for long sessions (#88020).runEmbeddedAgent is resolved before the first WebSocket message arrives (#88475).DEFAULT_ROOT_MAX_BYTES for short-term-recall.json and implement a write-side rolling trim to prevent files from growing into an unreadable state (#84291).type: "custom" tool definition error that is breaking all OpenAI-backed agents in v2026.5.27 (#88423).