OpenClaw merged PR #96005, a large Copilot harness update that brings its behavior closer to the Codex and PI agent lanes.
The headline change is parity. Copilot was missing several harness capabilities that other OpenClaw-backed agents already had: interactive ask_user, compact tool-search and code-mode routing, and shared user-input formatting logic. This PR moves those behaviors into reusable harness helpers instead of leaving each provider lane to reinvent them.
Shared Harness Helpers
The patch adds shared helpers for two important pieces of agent runtime behavior.
First, it standardizes user-input prompt delivery and answer normalization. Copilot now wires SDK ask_user through OpenClaw active-run queueing, including channel-safe prompt escaping and abort cleanup. Codex request-user-input formatting is also refactored onto the shared helper while keeping Codex app-server protocol parsing local.
Second, Copilot now routes tool-search and code-mode behavior through a shared compact catalog helper. That includes hidden-catalog allowlist filtering, cleanup, docs updates, and support for source-qualified tool entries from the Copilot SDK.
Why This Matters
Harness parity is not just internal polish. OpenClaw is increasingly a multi-provider runtime, and users expect core interaction patterns to behave consistently across agent backends.
If one lane can pause and ask a user a structured question while another cannot, workflows become provider-specific. If one lane sees a compact, policy-aware tool surface while another takes a different path, testing and debugging get harder. Shared helpers reduce that drift.
This also matters for future maintenance. By moving common user-input and compact-tool behavior into generic harness helpers, OpenClaw makes it easier to bring the next provider lane up to the same standard without copying provider-specific glue.
Proof And Scope
The PR cites Copilot SDK contracts for availableTools and built-in ask_user, along with Codex protocol and session references for request-user-input behavior. That dependency check is useful because this work sits at the boundary between OpenClaw's harness layer and provider-specific SDK behavior.
Validation included focused Vitest runs for the plugin SDK harness runtime, Codex user-input bridge, Copilot user-input bridge, Copilot attempts, and Copilot tool bridge. A live Copilot agent test also passed against local GitHub/Copilot auth, and the final changed-lane Testbox run passed.
For OpenClaw users, the practical result is that Copilot-backed agent runs should now feel less like a special case. They can participate in the same interactive prompt and compact tool-routing patterns that have become part of the broader OpenClaw runtime.