← Back to Blogs
GH Issues

OpenClaw Issue Digest: Session State Corruption, Auth Regressions, and Runtime Stability

18:30–00:30 UTC May 19, 2026

OpenClaw Issue Digest: Session State Corruption, Auth Regressions, and Runtime Stability

Open Issues

Recent activity in the OpenClaw repository reveals a series of high-severity regressions and architectural gaps, particularly concerning session state persistence, authentication security, and the stability of the Codex runtime.

Critical Stability and State Issues

Several reports highlight systemic failures in session management. Issue #66377 describes a critical scenario where agent session corruption persists even after agent deletion because state is fragmented across the agent directory, the LCM database, and cron job configurations. Similarly, #65983 identifies a leak where background PTY exec runs survive gateway restarts, becoming untracked orphan process trees that consume system memory.

In the Codex runtime, users are reporting severe stability issues. Issue #83968 reports a crash loop on macOS with an uncaught AssertionError [ERR_ASSERTION] assert(!this.paused), while #84305 details a failure where the contextEngine fails after turns exceed the model's context window without triggering compaction, leading to a "poisoned" session state.

Authentication and Security Vulnerabilities

Two high-severity security issues have emerged. Issue #84337 reveals a critical vulnerability where a hook ingress token can unlock password-mode gateway authentication if the secrets match, effectively upgrading a limited hook token to full operator access. Additionally, #65624 reports that Mattermost slash commands default to cleartext callback URLs, exposing reusable command tokens to on-path attackers.

Channel and Provider Regressions

Integration stability has seen several setbacks:

  • WhatsApp: Issue #65774 reports a critical failure where cron jobs ignored configured schedules and sent messages at 1 AM, and could not be stopped even via openclaw gateway stop.
  • Signal: Issue #66119 describes a catastrophic update to v2026.4.12 that set registered: false in account files, leading to the permanent deletion of Signal accounts from the servers.
  • Codex/OpenAI: Issue #84038 reports that doctor --fix silently migrates openai-codex/ configs to openai/, breaking the PI+OAuth runtime and causing 3-4x token inflation.
  • Telegram: Issue #80520 highlights a silent failure where messages are dropped without any sendMessage logs.

Key Themes

1. Fragmented State Persistence

There is a recurring theme of "ghost" state. Whether it is orphaned PTY processes (#65983), persistent corruption across agent deletions (#66377), or heartbeat-spawned sessions capturing user inbounds and causing conversation forks (#84332), the system struggles to maintain a single, authoritative source of truth for active runtimes.

2. The "Silent Failure" Pattern

Many reported bugs follow a pattern of silent degradation. Examples include the message_tool_only delivery mode in Telegram group chats silently dropping responses when the model forgets to call the tool (#84327), and the sqlite-vec extension failing to load on macOS, which silently degrades memory search to FTS5-only (#66977).

3. Codex Runtime Fragility

The bundled Codex harness is currently a primary source of instability. From startup delays and model catalog registration failures (#66251) to the delayed mirroring of inbound user transcripts in WebChat (#83528), the Codex integration requires significant stabilization work to be viable for production use.

Action Required

Immediate Security Fixes

  • Gateway Auth: Resolve the hook-token-to-password-auth escalation (#84337) and enforce HTTPS for Mattermost slash callbacks (#65624).
  • Signal Account Safety: Investigate the account deletion trigger in v2026.4.12 to prevent further data loss (#66119).

High-Priority Stability Work

  • Codex Stabilization: Address the assert(!this.paused) crash (#83968) and the context engine failure (#84305).
  • Session Cleanup: Implement a robust openclaw agent reset command to clear all fragmented state across LCM and cron configs (#66377).
  • Process Supervision: Fix the PTY orphan leak to ensure background workers are reaped on gateway restart (#65983).

Blocked Features/Bugs

  • Mattermost Slash Commands: Fix the 503 "not yet initialized" error in v2026.4.15 (#68113).
  • Control UI: Resolve the blank screen issue caused by the unresolved markdown-it-task-lists module import (#67680).

References

Issues