feat(agents): execute approved tool calls even when others are rejected

[pawel-twardziak]

humanInTheLoopMiddleware currently uses an all-or-nothing approach for mixed-decision batches: if any
tool call in a batch is rejected by the human operator, execution returns to the model and the
approved/edited calls are not executed in that round. This diverges from the Python SDK behavior where
rejected actions produce a ToolMessage with status: "error" while approved actions proceed to execution
immediately.

This PR adds an opt-in executeApprovedOnReject option that aligns JS behavior with Python.

Changes

libs/langchain/src/agents/middleware/hitl.ts

• Adds executeApprovedOnReject: boolean (default false) to HumanInTheLoopMiddlewareConfig
• Adds "tools" to canJumpTo so the middleware can jump directly to tool execution when applicable
• When executeApprovedOnReject: true and the batch contains a mix of approved/rejected decisions:
  • Rejected tool calls produce an artificial ToolMessage with status: "error" and the rejection message as content
  • Approved/edited tool calls remain in the AI message's tool_calls and jumpTo is set to "tools" — ToolNode automatically skips calls that already have a ToolMessage, so only the approved calls execute
  • If all calls in the batch are rejected, jumpTo falls back to "model" (same as legacy behavior)
• executeApprovedOnReject: false (default) preserves the existing behavior exactly — no breaking changes

Behavior comparison

Scenario	executeApprovedOnReject: false (default)	executeApprovedOnReject: true
All approved	Tools execute	Tools execute
All rejected	Returns to model	Returns to model
Mixed (approve + reject)	Returns to model; approved not executed yet	Approved tools execute; rejected → ToolMessage(status="error")

Tests

Two new test cases added to libs/langchain/src/agents/middleware/tests/hitl.test.ts:

1. Mixed batch — one approved, one rejected with executeApprovedOnReject: true: asserts the approved tool function is called once, the rejected function is never called, and the resulting ToolMessage for the rejected call has status: "error" with the rejection message as content
2. All-rejected batch with executeApprovedOnReject: true: asserts neither tool is called and the graph returns to the model (same as legacy)

---

Related to #10176

[changeset-bot]

🦋 Changeset detected

Latest commit: aaf0902

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
langchain	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[pawel-twardziak]

Christian Bromann (@christian-bromann) and Hunter Lovell (@hntrl) here is the python PR for this feature langchain-ai/langchain#32996