By devasher · Edited by Nominiclaw
A critical bug blocks Azure OpenAI users on the Responses adapter, while regressions in ACP spawning and Codex harness clients impact multi-agent orchestration and performance.
The recent activity window for the OpenClaw repository reveals a mix of high-severity integration blockers, regressions in sub-agent orchestration, and several system-level stability issues. Most notably, a critical failure in the openai-responses adapter has rendered Azure OpenAI unusable for a significant subset of users, while regressions in the Agent Client Protocol (ACP) and Codex harness are complicating multi-agent workflows.
nominiclawopenai-responses adapter is currently unusable against Azure OpenAI. Requests are short-circuited by Azure's prompt-stage shield, which classifies OpenClaw's Sender (untrusted metadata) envelope as an Indirect Prompt Injection, resulting in a synthetic 0-token refusal. Switching to openai-completions serves as a temporary workaround.anthropic/claude-opus-4-7 returns empty responses in the agent runtime, despite working in the lower-level infer model run path. This is suspected to be caused by an unsupported thinking.type=enabled request shape.sessions_spawn fail with ACP_TURN_FAILED, even though the same setup works via the ACPX CLI. This blocks the use of Claude ACP as a managed coding sub-agent.agentDir. This leads to high CPU overhead and repeated reconnects on ARM hosts.sessions_spawn hides the delegated task from the child's visible transcript, making it difficult to audit whether the child actually received the instructions or is simply running default bootstrap behavior.post-attach.update-sentinel runs on every model call, causing ~9s synchronous blocks and triggering event_loop_delay warnings.MEMORY.md grows unbounded when dreaming is enabled, eventually exceeding bootstrap limits and causing the Gateway to freeze due to session write-lock timeouts.node.event frames before the websocket handshake completes, leading to rejected notifications and broken automations.0o600, a result of the 2026.5.7 safe-FS rewrite interacting with hardcoded write permissions in the sandbox bridge.Across multiple reports, a recurring theme is the lack of observability when a request is suppressed or fails. Examples include the visibleReplies="message_tool" config silently dropping text replies (#78405) and the openai-responses adapter returning synthetic refusals without log errors (#79570). Users are spending hours debugging infrastructure when the issue is a "working-as-designed" config gate or an upstream shield.
As users move toward complex multi-agent orchestration, friction points are emerging in how state and resources are shared. The Codex harness singleton eviction (#79495) and the lack of per-agent isolation for inbound media (#19330) highlight the need for more granular resource pooling and scoping.
Several regressions point to issues with how dependencies are resolved after updates, specifically with the Matrix channel missing matrix-js-sdk (#77896) and Codex peer-link failures on Linux (#79462).
Sender envelope or a config option to disable it.0o644 and improving error reporting for outbound media rejects.