AI & Agent Dev Bug Sandbox Back to Radar

HumanInTheLoopMiddleware Cross-Tool Edit Bypasses Target Tool'S Interrupt_on Policy

An edit decision can change a tool call from tool A to tool B, but the resulting call is evaluated only against tool A's InterruptOnConfig. If tool B has its own interrupt_on policy, it is silently skipped, allowing execution without the required human approval.

highConfidence 90%LangChain Affected V1.3.18

Origin Analysis

HumanInTheLoopMiddleware selects the interrupt_on configuration using the original tool name and after processing an edit does not re-evaluate the target tool's configuration before the revised call is routed to ToolNode.

1. Configure HumanInTheLoopMiddleware with interrupt_on for tool_a (allowed_decisions=['approve','edit']) and tool_b (allowed_decisions=['approve','reject']).\n2. Create an agent with tools tool_a and tool_b and a deterministic model that initially calls tool_a.\n3. Invoke the agent; the initial tool_a call triggers an interrupt.\n4. Resume with a Command containing an edit decision that changes the action name to tool_b.\n5. Observe that tool_b executes immediately without a second interrupt, printing 'Second interrupt: False' and executed = [('tool_b', 'edited')].

Fixing Code Block

0x7F 45 4C 46 02 01 01 000x90 12 AB 04 C1 7E 33 08
0xA4 1C 90 3E 77 0B D2 580x11 F6 6A 2C 80 19 DE 44
0x00 00 00 00 03 00 3E 000x01 00 B7 4D 55 90 0C 61
0x6C 69 62 63 2E 73 6F 2E0x36 00 8E 17 42 AA 05 F3
0xDE C0 DE 01 8B 45 FC 890xC7 5D 20 6F 91 03 E4 78
0x13 37 00 4F 9A 2D E1 700x08 B4 5C 22 71 D0 4A 16
0x48 8D 3D 22 01 00 00 E80x91 FF 0A 63 2E 57 B8 04
0x5A 91 C3 0E 44 70 1B 880x02 6F 39 C4 15 80 ED 27
0xFF 15 40 20 60 00 48 890xC3 90 7A 01 4C 28 B1 56
0x0F 1F 44 00 00 55 48 890xE5 C3 18 6D 92 0B 33 A7

Edge Case Audit

This change rejects previously allowed cross-tool edits to any tool listed in `interrupt_on`, potentially breaking existing workflows that relied on such edits. To roll back, remove the added guard or adjust the interrupt_on configuration. If re-evaluation of the target tool's policy is preferred over rejection, a more complex change is required. No thread-safety or concurrency regression is introduced by this simple guard.