OpenClaw Activity Digest: May 6, 2026

Open Issues

The six-hour window from 2026-05-06T12:30:26.000Z to 2026-05-06T18:30:26.000Z saw several significant issues reported and discussed within the OpenClaw repository.

#78581 · Sub-agent announce wake events dropped on FallbackSummaryError

This critical bug (#78581) prevents sub-agents from announcing wake events when the parent session encounters a FallbackSummaryError, typically due to provider cooldowns. The error classification logic fails to recognize FallbackSummaryError, leading to zero retries, silent failures, and manual intervention requirements for parent sessions. This is identified as the root cause of recurring user complaints about unresponsive sub-agents during peak provider load.

#78578 · WhatsApp agent reply MEDIA: directives dropped

Issue #78578 details a regression in the WhatsApp channel where agent replies containing MEDIA: directives are not dispatched as attachments. While direct CLI media sends work, the agent's rich output path for media is failing, impacting the ability of agents to send generated images or other media files via WhatsApp.

#78311 · Control UI Settings page shows wrong Exec Policy

Bug #78311 points out a discrepancy in the Control UI's Settings page, which incorrectly displays the Execution Policy as allowlist even when the configuration (tools.exec.security) is set to full. This is due to the UI reading from an incorrect configuration path (agents.defaults.exec.security instead of tools.exec.security).

#78478 · Treat toolResult entries as valid cut points for safeguard compaction

Feature request #78478 proposes enhancing the safeguard compaction mode by treating toolResult entries as valid cut points. Currently, sessions dominated by tool calls fail to compact meaningfully, leading to empty fallback summaries and preventing session extension. Including toolResult would resolve this for long-running agentic tasks and cron jobs.

#8441 · Add thinking and model config to skills.entries

Feature request #8441 suggests adding thinking and model configuration fields to skills.entries. This would allow per-skill overrides for thinking level and model selection, aligning with existing cron job payload capabilities and enabling more granular control over skill execution.

#78579 · Decouple OpenClaw core from channel plugins

Proposal #78579 advocates for a more modular update architecture by decoupling the OpenClaw core from channel plugins. This would allow independent updates, faster fixes for specific channels, and easier community contributions, improving overall system maintainability and stability.

#78577 · Control UI CSP/X-Frame-Options has no opt-in for same-origin framing

Regression #78577 highlights that OpenClaw 2026.5.5 hardcodes X-Frame-Options: DENY and frame-ancestors 'none' for the Control UI, preventing same-origin iframe embeds. A configuration option is requested to allow 'self' framing for custom dashboard integrations.

#9546 · Add option to disable 'new messages' notification banner in WebChat

Feature request #9546 aims to provide an option to disable the intrusive 'new messages' notification banner in the WebChat UI, improving the user experience when reading long responses.

#9987 · CLI command to list pending exec approval requests

Feature request #9987 proposes a CLI command (openclaw approvals pending) to list pending execution approval requests, including the command details. This would greatly improve visibility and debugging for agents requiring approvals.

#78463 · Stale running tasks after gateway restart

Bug #78463 reports that tasks remain in a running state after a gateway restart, marked as stale_running in openclaw tasks audit. Manual intervention (openclaw tasks maintenance --apply) is required to clean them up, impacting operational visibility.

#78571 · Telegram Connection Bot Bug

Bug #78571 describes issues with the Telegram connection bot on Windows, where various invocation failures (ENOENT, EINVAL, silent reply dispatch) prevent successful message processing due to problems with command resolution and shim handling.

#78570 · OpenClaw update/doctor rewrites OAuth model provider to API-key provider

Bug #78570 details how openclaw update and doctor --fix incorrectly rewrite OpenAI Codex OAuth model references (openai-codex/*) to direct OpenAI API key references (openai/*), leading to unexpected costs and broken configurations.

#78569 · exec host=node invoke-failed after approval (APPROVAL_CLIENT_MISMATCH) via Control UI

Regression #78569 indicates that exec host=node commands requiring approval fail with APPROVAL_CLIENT_MISMATCH when approved via the Control UI using token authentication. This prevents remote node execution approval workflows.

#9132 · Control UI: Add status indicator for sessions

Feature request #9132 suggests adding visual status indicators (active, completed, timed out) to sessions in the Control UI for better clarity.

#78509 · doctor --fix migrates Codex OAuth GPT-5.5 route to openai/gpt-5.5, causing missing OPENAI_API_KEY

Bug #78509 is similar to #78570, specifically noting that doctor --fix's migration of Codex OAuth models to direct OpenAI paths breaks the setup by requiring a missing OPENAI_API_KEY.

#8959 · Feature: Cross-platform multi-select component in message tool

Feature request #8959 proposes adding a multi-select capability to the message tool, allowing users to select multiple options in a single interaction, reducing latency and cost associated with multiple agent turns.

#78562 · Repeated tool-loop context overflows cause successive auto-compactions

Bug #78562 describes a scenario in OpenClaw 2026.5.5 where repeated tool-loop context overflows lead to successive auto-compaction cycles, causing the agent to get stuck in a compaction loop.

#78560 · Control UI: Standalone browser-native voice input (SpeechRecognition API)

Enhancement #78560 suggests adding a self-contained, browser-native voice input button to the Control UI using the Web Speech API, offering live transcripts and simpler integration without server-side audio processing.

#9658 · Feature: Support streaming or file-based tool arguments to bypass JSON serialization limits

Feature request #9658 addresses the limitation of JSON serialization for large tool arguments, proposing support for streaming or file-based arguments to handle large files and complex code generation reliably.

#78514 · [Bug]: WhatsApp plugin install fails on Windows native (Node 22, npm 10.9.3)

Bug #78514 reports that the @openclaw/whatsapp plugin fails to install via openclaw plugins install on Windows with Node 22 and npm 10.9.3, citing an npm reify phase error related to dependency resolution.

#78553 · Web UI Skills page missing multi-agent support

Bug #78553 highlights that the Web UI Skills page only displays skills from the default agent's workspace, failing to show skills from other configured agents, despite the backend API supporting multi-agent skill retrieval.

#78502 · Google Gemini models hang/timeout on main sessions but work via direct API & isolated subagents

Bug #78502 details a critical issue where Google Gemini models (3.1 Pro & 2.5 Pro) hang and timeout on main OpenClaw sessions but function correctly with direct API calls and isolated subagents, indicating a problem with OpenClaw's Google provider pipeline for full sessions.

#78538 · [Docs]: secrets apply --from <plan.json> accepts providerUpserts and providerDeletes, but docs only document targets

Documentation gap #78538 points out that the secrets apply command's plan contract accepts providerUpserts and providerDeletes, but these are not documented, making it difficult for users to create comprehensive secret management plans.

#78551 · WeChat (openclaw-weixin) channel login fails with