7. Human approval

Finding the best index set is automatic. Creating those indexes on a production table is not free — it takes space and slows writes. So before applying them, the optimizer pauses for a human to approve.

Runnable file: examples/tutorial/queryopt/ch07_human.py.

Pause with ctx.interrupt

async def approve(state, ctx) -> dict:
    inc = state["search"]["incumbent"]
    answer = await ctx.interrupt({
        "question": "apply these indexes?",
        "indexes": list(inc["indexes"]),
        "value": inc["value"],
    })
    return {"applied": str(answer).lower() in ("y", "yes", "apply")}

On a fresh run ctx.interrupt suspends and checkpoints; a later resume feeds the answer back and the call returns it. Interrupting needs a checkpointer, because the paused state must live somewhere.

await app.invoke(init, thread="run-1")       # runs until the pause
state = await app.get_state("run-1")          # state.interrupted is True
req = state.interrupt_payload                 # the indexes awaiting approval
out = await app.resume("run-1", value="yes")  # apply them

optimizing until the approval pause:
  PAUSED: apply 2 indexes (value -8.0)?
applied=True

The same mechanism the agent uses

In chapter 5 the Codex agent ran with AllowAll(). Switch its policy to Interactive and every index the agent wants to create becomes a ctx.interrupt a human approves or denies — the identical pause-and-resume machinery, now driven by the agent's permission policy instead of a node you wrote. Approval is one mechanism whether you write the interrupt or an agent's policy writes it for you.

What you learned

• ctx.interrupt(payload) pauses the run for a human; resume(thread, value=...) continues it with the answer.
• Interrupting requires a checkpointer — the paused state has to persist.
• An agent backend's Interactive policy turns tool/index requests into the same interrupt automatically.

In chapter 8 the search survives a restart, and a token budget guards an LLM-backed run.