OpenClaw Digest: Improving Session Persistence, Channel Integrations, and Tool Transparency

Recent updates to OpenClaw have focused on strengthening the reliability of long-running agent sessions, refining the user experience across various messaging channels, and improving the observability of tool executions. These changes ensure that agents can maintain state across context window compactions and that users receive cleaner, more accurate feedback on mobile platforms.

Merged PRs

• [AI-assisted] fix(status): accept expected Gateway bind listeners #78128
• [AI-assisted] fix(status): treat CLI runtime aliases as selected route #79082
• [AI-assisted] fix(qqbot): translate skill descriptions #78153
• [AI-assisted] fix(tasks): prune stale cron session registry entries #74023
• fix(gateway): restore runtime-postbuild sync in watch mode #70805
• [codex] Map ACP thinking to advertised effort key #79926
• Stop preserving stale whole-agent Codex runtime pins #79651
• [codex] Mirror Codex tool calls in transcripts #79952
• fix(agents): preserve active exec references across compaction #79307
• fix(imessage): wire reply attachments through send-rich --file (with feature gate) #79864
• test: sandbox audit-exec-surface under HOME tempdir #79885
• [AI-assisted] fix(whatsapp): delta repeated tool preambles #79120
• fix(slack): wake interactive reply sessions #79836

Key Changes

Agent Session & Context Management

One of the most critical fixes addresses a failure in subagent session management. Previously, when a subagent underwent context compaction, references to active background exec processes were lost, effectively orphaning the processes.

Subagent compaction loses running exec process references... the process session reference is lost. The actual process may continue running but becomes orphaned — the subagent can no longer poll/log/write to it.

To resolve this, OpenClaw now maintains a bounded runtime-owned snapshot of active background exec sessions, ensuring these references are preserved in rebuilt system prompts after compaction.

Channel Integration Enhancements

Several updates improve the reliability and clarity of messaging integrations:

• WhatsApp: Fixed a bug where tool-call preambles were sent as cumulative, growing messages. The system now sends only the delta (the new suffix) of the preamble, reducing noise for the user.
• iMessage: Implemented a feature gate to support attachments in threaded replies via send-rich --file. This allows agents to deliver files in a single threaded send rather than dropping them or throwing errors.
• Slack: Resolved an issue where interactive reply buttons were logged but failed to wake the active session. Slack block actions now request an immediate heartbeat to ensure the conversation continues visibly.

Codex & ACP Refinements

• Tool Transparency: Codex tool calls and results are now mirrored into the canonical OpenClaw session transcript. This ensures that consumers like ClawBench can see the full trajectory of a run, including bash commands and file edits, rather than just the final assistant message.
• ACP Mapping: Added a mapping layer for the Agent Client Protocol (ACP) to resolve canonical OpenClaw options (like thinking) to backend-advertised keys (like effort), preventing internal errors when updating session configurations.
• Legacy Cleanup: Removed stale whole-agent Codex runtime pins and normalized legacy openai-codex/* routing in cron stores to align with the current openai/* canonical references.

Infrastructure & Diagnostics

• Gateway Status: The status and doctor commands now correctly recognize healthy LAN-bound Gateway listeners, eliminating false-positive port conflict warnings for users with internal reverse-proxy deployments.
• Development Workflow: Restored runtime-postbuild synchronization in watch mode to prevent the Gateway from launching with incomplete runtime trees when generated files are missing.

Impact

These changes significantly improve the robustness of OpenClaw for professional use cases. The preservation of exec references allows agents to handle long-running data collection or training tasks without fear of context window limits interrupting their control. The improvements to WhatsApp and iMessage make the agent feel more native and less intrusive on mobile devices.

Furthermore, the mirroring of Codex tool calls into transcripts provides essential observability for developers benchmarking their models. By treating the transcript as the single source of truth, OpenClaw ensures that the actual behavior of the agent—including its internal tool-use trajectory—is fully auditable and reproducible.