OpenClaw Stops No-Op Tool Loops

OpenClaw merged PR #97044, "fix(tools): treat no-op writes and edits as terminal tool-loop failures", a P2 agent-runtime reliability fix for file editing tools.

The bug was easy to understand and annoying in practice. If an agent called the write tool with content identical to the existing file, or called edit with a replacement that produced unchanged content, the loop could treat the no-op as progress and try again.

That turns a harmless non-change into wasted work. In the worst case, the agent keeps attempting the same mutation even though the file is already in the requested state.

What Changed

The patch makes no-op writes and edits terminal tool results. In the PR author's proof, the result now includes terminate: true, which tells the tool loop to stop treating the call as a reason to continue retrying.

The implementation covers several paths:

• write now returns a terminal no-op result when the precheck detects identical content.
• edit filters out entries where oldText and newText are identical.
• All-no-op edit batches now return a terminal result.
• Mixed edit batches still execute the valid sibling replacements.
• No changes made errors from the edit-diff path are converted into terminal no-op results.

That mixed-batch behavior is the useful nuance. OpenClaw is not rejecting an entire edit request because one replacement is redundant. It removes the redundant part and preserves the real edit.

Why It Matters

No-op tool calls are common in agentic coding. A model may ask to write a file after it has already made the intended change. It may propose an edit whose replacement text matches the current file after another tool call, a formatter pass, or a nearby correction.

The right behavior is not to spin. The right behavior is to tell the loop that no change is needed and let the agent move on.

PR #97044 gives OpenClaw that boundary. A no-op is now a terminal answer from the tool, not a vague failure and not fake progress.

Validation

The PR adds regression coverage for both tool families: two tests for write and three tests for edit, including all-no-op, mixed-batch, and real-change cases.

The focused test command reported 14 passing tests across the write and edit tool suites. The PR body also includes real behavior proof from Linux on OpenClaw 2026.6.10 with Node 24.13.1, covering identical writes, different writes, all-no-op edits, mixed edit batches, and real edits.

The one caveat is explicit: a full end-to-end LLM loop was not tested because it requires real model inference. Still, the tool contract is now clearer. When OpenClaw can prove a requested file mutation changes nothing, it can stop the loop cleanly instead of letting the same request echo.