input-after-effect¶
Severity: Error. Reported by ouncemcp check and fails it.
What it finds¶
A write or destructive handler calls ctx.require_input(...) after ctx.effect_applied(), and the tool is not registered with replay_safe=True.
Why it matters¶
In MCP 2026-07-28 the client answers an input_required result by sending the whole tools/call again, so the handler runs from the start. Whatever it did before asking happens twice: two charges, two emails, two deleted rows (spec SEC-7). OunceMCP also refuses this at run time with an internal error.
How to fix it¶
Ask for input before the effect:
@server.tool(name="ticket.close", effect=Effect.WRITE)
async def close(ticket_id: str, ctx: ToolContext) -> Closed:
if not ctx.input_responses:
ctx.require_input(CONFIRM, state={"ticket": ticket_id})
await api.close(ticket_id)
ctx.effect_applied()
return Closed(ticket_id=ticket_id)
If running the effect twice is harmless (it is idempotent, or keyed by an idempotency key), register the tool with replay_safe=True.