OpenClaw merged a major P1 correctness fix for node-backed sessions this morning: PR #164647, "fix(nodes): give node sessions the same prompt and bootstrap context as local sessions". The change closes a gap between Gateway-local sessions and sessions executed on a paired node.
The issue was direct: node sessions could miss selected agent instructions, persona, memory guidance, and conversation context. For an agent runtime, that is not cosmetic. It changes how the model understands the user, the workspace, the rules, and the tools available to the turn.
What Changed
The PR says node turns now use OpenClaw's existing bootstrap, system-prompt, prompt-hook, and context-engine owners instead of assembling a smaller generic coding prompt. The node still contributes execution facts such as workspace path, host, operating system, shell, and process details, but the selected instructions come from the same preparation path as local sessions.
That preserves the distinction between Gateway-owned policy and node-owned execution facts. In practical terms, a remote node can describe where it is running without losing the human's configured agent context.
The change also keeps prepared instructions authoritative when tools change, and it preserves runtime-context entries through transcript commit and replay.
Why It Matters
OpenClaw nodes are supposed to extend where work can run, not create a different personality or memory model for the same agent. If a local session has the full bootstrap and a node session silently falls back to a generic prompt, the same request can behave differently depending on placement.
That hurts reliability in several ways:
- Node sessions may omit memory-recall guidance.
- Agent persona and standing instructions can disappear.
- Runtime context can fail to survive replay.
- Tool prompt surfaces can drift from local sessions.
- Follow-up turns can lose the shared system context users expect.
This PR makes node placement feel less like a special mode and more like the same OpenClaw session running on a different executor.
Verification
The PR body includes live provider-boundary proof comparing the landed prerequisite with the final candidate. In that controlled capture, all 24 turns completed with exact expected responses. The final candidate's local and node tool arrays matched byte-for-byte, both included selected bootstrap markers and Memory Recall guidance, and each session retained stable system, tool, and cache-key bytes across six turns.
The remaining differences were the ones OpenClaw should preserve: paths, host facts, operating system facts, shell facts, session identity, and project location. In other words, the node says where it is, but it no longer loses the agent's selected context.
The PR also reports final CI and openclaw/ci-gate passing, plus ClawSweeper review with replay findings resolved.
Bottom Line
This is one of those fixes that should make distributed OpenClaw feel calmer. Users should not need to know whether a turn ran locally or on a paired node to trust that the agent received the same instructions and memory policy.
For teams relying on node execution, PR #164647 is a high-signal reliability improvement.