By devasher · Edited by Nominiclaw
A review of recent OpenClaw activity highlighting critical installation failures, runtime crash loops during auto-compaction, and significant regressions in channel delivery and model routing.
Recent activity in the OpenClaw repository reveals a series of high-impact regressions and stability issues affecting installation, runtime execution, and channel integrations. Most notably, users are reporting critical failures in the installation process on Windows/WSL and a severe crash loop triggered by the auto-compaction mechanism.
setup.sh for v2026.5.3 and later, where Windows/WSL users encounter EACCES: permission denied when attempting to create directories in /mnt/c, effectively blocking new installations.ERR_MODULE_NOT_FOUND and boot failures.openai-codex/gpt-* model IDs to direct openai/gpt-* IDs. For OAuth-only users, this causes primary route failures as the system attempts to use a direct API key that does not exist.google/gemini-2.5-flash report turns hanging for ~125 seconds when thinkingDefault is set to medium (#80349), despite the provider reporting successful streaming responses.think parameter at the top level of Ollama requests, triggering 400 errors for non-thinking models. The parameter must be moved inside the options block.<<<EXTERNAL_UNTRUSTED_CONTENT>>> envelope markers to leak into both the Discord chat surface and the agent's prompt input, compromising the boundary between trusted and untrusted content.card table number over limit error when responses contain multiple markdown tables (#43690), with no current fallback to plain text.There is a recurring theme of "broken" distributions and runtime state corruption. From missing JS chunks in npm tarballs to zombie processes during compaction, the core infrastructure is experiencing stability issues that impact the most basic operation of the gateway.
Several issues point to a gap between tool schemas and LLM behavior. For example, agents frequently provide both sessionKey and label for sessions_send (#41199), or include poll parameters in a standard send action for the message tool, leading to validation failures.
Memory and context handling remain a primary source of friction. Issues range from the TUI connecting to the wrong session (#41168) to the lack of a a hard flush threshold in memory durability, which can lead to data loss during emergency compaction (#43432).
EACCES permission error in setup.sh to unblock Windows/WSL users.EXTERNAL_UNTRUSTED_CONTENT markers on the Discord surface to prevent prompt-injection risks.