OpenClaw Development Digest: Orchestration Primitives, LLM Core Refactoring, and Stability Fixes

The latest development window for OpenClaw has been marked by significant architectural shifts and a concerted effort to harden the system's reliability. From the introduction of complex orchestration primitives in the Workboard to a deep refactoring of the LLM runtime, the focus is on scalability and ensuring a seamless user experience across diverse communication channels.

Merged PRs

• fix: show chat errors as visible messages Original PR
• refactor: extract LLM core packages Original PR
• fix(ios): guard websocket ping continuation Original PR
• fix(cli-runner): write-side flush gate + orphan-tool-use invalidator Original PR
• fix(whatsapp): retry QR login 408 timeouts Original PR
• perf(test): fix explicit-file Vitest wrapper hangs Original PR
• fix(agent): preserve media task success on delivery miss Original PR
• Refactor task state onto shared SQLite Original PR
• feat(ci): autoscrub dependency lockfile-only PR changes Original PR
• ci: extend platform checkout fetch timeout Original PR
• Fix restart sentinel internal continuations Original PR
• fix(imessage): preserve SMS approval reply routes Original PR
• Fix Codex raw image generation media projection Original PR
• Fix subagent DM completion delivery after yield Original PR
• fix(agents): extend terminal outcome projections Original PR
• feat(workboard): add orchestration primitives Original PR
• fix(agents): classify embedded provider business denials for fallback Original PR

Key Changes

Architectural Refactors & New Features

• Workboard Orchestration: The Workboard now supports advanced orchestration primitives, including dependency links, idempotent child creation, and structured lifecycle operations (complete/block). This allows for complex agent fan-out and dependency promotion, all backed by a plugin SQLite store.
• LLM Core Extraction: To improve modularity, shared LLM contracts, event streams, and tool validation have been extracted into packages/llm-core, while the provider-neutral API registry and runtime helpers now reside in packages/llm-runtime.
• State Management: Task and task-flow registry persistence have been migrated to a shared OpenClaw SQLite database, replacing legacy sidecar files and hardening snapshot pruning.

Stability & Reliability Fixes

• Session Recovery (Claude-CLI): Addressed critical context-loss failures in the Claude-CLI runner. A new "binding-flush gate" ensures session IDs are only persisted if the transcript is successfully flushed to disk, and an "orphan-tool-use invalidator" prevents the agent from stalling when a session crashes mid-tool.
• Message Delivery & Routing:
  • iMessage/SMS: Fixed a bug where SMS approval replies and follow-ups would fail or fall back to iMessage, ensuring they remain SMS-qualified.
  • Subagents: Resolved an issue where subagent completion announcements were lost after a sessions_yield because the outbound adapter was temporarily unavailable.
  • Codex: Fixed a projection issue where Codex raw image generation results were being dropped, now correctly delivering them as managed media artifacts to channels like Telegram.
• Restart Sentinel: Fixed a security and UX leak where restart-sentinel turns could auto-deliver plain assistant output to source chats. These are now internal by default.

Developer Experience & CI

• CI Optimization: Introduced an autoscrubbing mechanism for lockfile-only PR changes to reduce accidental dependency noise. Additionally, the platform checkout fetch timeout for Windows/macOS runners was extended to 90s to prevent index-pack failures.
• Testing Performance: Fixed Vitest wrapper hangs by routing explicit-file invocations through a project router, significantly reducing startup time for single-file tests.

Impact

These changes directly address several high-impact pain points for users and maintainers:

• Reduced "Amnesia": The fixes to the CLI-runner binding-flush and orphan-tool-use logic prevent the agent from "forgetting" prior conversations after a crash or restart, which was a primary source of frustration for power users.
• Improved Reliability in Messaging: By hardening the iMessage SMS routes and subagent delivery, OpenClaw ensures that critical approval flows and agent handoffs are no longer silently dropped.
• Enhanced Control: The new Workboard primitives transform the board from a simple view into a powerful orchestration tool, allowing users to manage complex, multi-agent workflows with explicit dependencies.
• Better Error Visibility: The UI now surfaces gateway chat errors as visible assistant messages, ending the "silent failure" behavior where the chat would simply hang when a provider authentication failed.