HACH developer: safe agent cancellation needs 3 layers

Cancellation is a problem of consistency rather than a simple stop button. While text generation can be cut at any moment by discarding unshown streaming output, tool calls that produce side effects — sending mail, updating a ticket — can only be interrupted at checkpoints before or after execution. A cancel signal that arrives mid-call is therefore left to finish, with the decision deferred to the next checkpoint.

A second concern is what the agent remembers. The design keeps a cancel-time record in both the conversation history and the user notice, distinguishing steps cancelled before execution from those already completed, and pointing to an undo path where one exists. Without that record, the next turn can misread the state and run the same action twice — the very duplicate-execution failure that cancelling was meant to prevent.

The design also draws a line between explicit and implicit cancellation. A new message that plainly says to stop triggers an immediate interrupt, but an unrelated request is queued so the current turn can finish, and ambiguous cases are allowed to complete because the inconsistency risk from interrupting outweighs a slightly verbose answer. The author notes this comes from building HACH, an agent that lives in Slack / Teams / Chatwork / LINE WORKS and collects chat, email, calendars and meeting notes for a daily morning report.