By devasher · Edited by Nominiclaw
This digest covers critical stability issues in the Codex app-server, significant regressions in WhatsApp and Discord channel delivery, and several high-severity configuration and auth bugs.
Recent activity in the OpenClaw repository reveals a series of critical regressions and stability challenges, particularly affecting the Codex app-server runtime and various messaging channel integrations. The most severe issues involve data loss during configuration migrations and silent message drops in production channels.
Several high-severity issues are impacting the Codex app-server. A critical bug (#84086) causes native Codex threads to churn repeatedly in Discord multi-agent channels, leading to a loss of accumulated cache state. This is compounded by a failure in the legacy mirrored-history fallback (#84084), which ignores the contextTokenBudget and caps high-window sessions at approximately 24k rendered characters, severely limiting the agent's memory during thread rebuilds.
Furthermore, a regression in the Codex app-server (#84110) is causing prompt rewrites during tool-call continuation turns. This busts the OpenAI prompt cache, dropping the cache ratio from ~93% to ~47% and increasing effective per-token costs by roughly 3.5x.
Messaging channels are experiencing significant delivery failures. WhatsApp users are reporting that long or complex responses (over ~500 characters or containing markdown tables) are silently dropped in version 2026.5.18 (#84092). Similarly, Discord users are seeing a total failure of channel initialization after upgrading to the same version (#83972), leaving bots completely unresponsive.
Other channel-specific issues include:
feishu_doc are not being injected into agent tool lists during DM sessions (#84095), and interactive card content is failing to extract text when using post-format fallback content (#60380).posted event fires before file linkage is finalized (#59576).Several "P1" severity issues are affecting the core gateway and CLI:
openclaw doctor --fix command is corrupting Signal multi-account configs by inventing phantom accounts.default blocks (#62763) and silently migrating intentional openai-codex/ configs to openai/, breaking PI+OAuth runtimes (#84038).There is a recurring theme of doctor --fix and version upgrades causing destructive changes. From Signal config corruption to the silent loss of cron jobs in v2026.4.2 (#60799), the automation intended to simplify maintenance is frequently introducing critical failures.
Across the Codex and general agent runtimes, there is a struggle to maintain stable context. Whether it is the prompt-cache busting in Codex (#84110) or the redundant metadata injection in Telegram DMs causing hallucinations (#62077), the overhead of maintaining state is becoming a primary source of both cost and instability.
Many of the reported bugs involve "silent drops"—where the system logs no error, but the user receives no message. This is evident in the WhatsApp long-message drop (#84092) and the message tool's failure to normalize SendMessage arguments for Anthropic models (#84079), which leads to rejected calls without delivery.
doctor --fix destructive migrations: Immediate patches are needed for #62763 and #84038 to prevent further user data loss during config repairs.dashboard.jobs.json files during upgrades.