← Back to Blogs
GH Issues

OpenClaw Issue Digest: Regression Trends and Runtime Stability in v2026.5.12

06:30–12:30 UTC May 16, 2026

OpenClaw Issue Digest: Regression Trends and Runtime Stability in v2026.5.12

Open Issues

The recent update to OpenClaw 2026.5.12 has introduced several critical regressions and stability challenges, primarily centered around configuration migration, runtime performance, and provider compatibility.

Critical Regressions and Stability Issues

  • Configuration Wipe during Upgrade: Issue #82533 reports a severe regression where the channels section of openclaw.json is silently dropped during the update to 2026.5.12. This is compounded by a failure in the pre-update backup mechanism, which writes the backup after the data is wiped, leaving users with no recovery path for their WhatsApp and other channel configurations.
  • Event Loop Starvation: Multiple reports (#82462, #82525) indicate that the gateway's event loop is being blocked for 25-30 seconds. This is occurring during large context processing for OpenAI-compatible providers and during getMe retries for Telegram bot accounts. This starvation prevents the gateway from processing heartbeats, /stop commands, and inbound messages, effectively freezing the system.
  • SSRF Protection Over-blocking: The introduction of default-on SSRF protection in 2026.5.12 has bricked self-hosted setups using private or CGNAT IPs (e.g., Tailscale, LMStudio, Ollama). Users are seeing SsrFBlockedError and generic network timeouts (#82010).
  • Codex Harness Registration Failures: There is a recurring pattern of "Requested agent harness 'codex' is not registered" errors (#82437, #82368). This appears to be a result of the update process disabling the Codex plugin or failing to migrate the runtime configuration to the new model-scoped policy.

Provider and Model Compatibility

  • Xiaomi MiMo 400 Errors: The change to add permissive items schemas to array tool parameters in 2026.5.12 has broken compatibility with the Xiaomi MiMo v2.5 provider, which returns 400 Param Incorrect when these extended schemas are present (#82447, #82500).
  • Kimi Thinking Mode Failures: Kimi models are rejecting requests with HTTP 400 when thinking is enabled but reasoning_content is missing in assistant tool-call messages during multi-turn sessions (#82161).
  • OpenRouter Signature Poisoning: A regression in the OpenRouter passthrough path is persisting literal field names (e.g., "reasoning_details") as thinkingSignature, causing subsequent turns to fail with HTTP 500 (#82335).

Session and Memory Management

  • Unbounded Backup Growth: The session-file-repair mechanism is writing unbounded .bak snapshots without rotation, leading to gigabytes of wasted disk space on stuck sessions (#80960).
  • Session Model Pinning: A persistent bug exists where a session pinned to a fallback model during a failure never reconciles back to the primary model, even after the primary recovers (#82544).
  • Context Pollution in Telegram: Telegram sessions are duplicating history by attaching a conversation context block that overlaps with the existing session transcript, wasting tokens and degrading response quality (#82040).

Key Themes

1. The "Update Trap"

There is a strong theme of destructive updates. From the silent wiping of channel configs (#82533) to the breaking of custom provider schemas (#81923) and the disabling of the Codex plugin (#82368), the migration path to 2026.5.12 appears unstable. Users are reporting a pattern where every minor version upgrade breaks previously working custom configurations.

2. Runtime Performance Bottlenecks

Performance issues are shifting from raw LLM latency to gateway overhead. The combination of event-loop starvation (#82462) and expensive per-run auth/model preparation (#82476, #80131) is significantly increasing Time-to-First-Token (TTFT), sometimes by 10-20 seconds.

3. Codex vs. Pi Parity

As OpenClaw moves toward Codex as the default runtime, a significant effort is underway to ensure parity. Current issues highlight structural drift in transcripts (#80395) and gaps in tool-call shapes between the two runtimes (#80171, #80173).

Action Required

Immediate Attention (High Severity)

  • Fix openclaw.json migration: Resolve the silent deletion of the channels block and fix the pre-update backup timing to prevent data loss (#82533).
  • Address Event Loop Blocking: Implement setImmediate yields or worker threads for SSE parsing and Telegram getMe calls to prevent gateway freezes (#82462, #82525).
  • Refine SSRF Guard: Document and simplify the allowPrivateNetwork bypass for self-hosted providers to restore connectivity for LAN/Tailscale users (#82010).

Blocked or High-Priority Fixes

  • Codex Harness Registration: Fix the update path to ensure the Codex plugin remains enabled and the runtime policy is correctly migrated (#82368, #82437).
  • Tool Schema Compatibility: Make the items schema addition optional or configurable to support providers like Xiaomi MiMo (#82447).
  • Session File Repair: Implement backup rotation or deletion after successful repair to stop unbounded disk growth (#80960).

Contributor Opportunities

  • Accessibility: Implement the requested "Linear Persistent Workspace Mode" for blind users to reduce cognitive load from session fragmentation (#82450).
  • UI Enhancements: Develop the proposed sessions overview with peek and inline reply capabilities (#80901) and dynamic browser tab titles (#80942).

References

Issues