Refuse a blank question put to a person, the way a blank task handed to a Bot is refused

[kevin9327]

A Bot calls ask_person with {"question": ""}. It is told:

That was put to the person in this conversation. Ask it in your own words now, plainly, and stop
there: do not answer it yourself and do not hand it to another Bot.

so it stops, and its turn ends. On the Audit screen there is now an agent.escalated row — the row
an administrator counts escalations by — whose question is empty. It says a person was asked
something that was never said.

On a deployment whose EscalationRoute is a duty desk or an on-call rota rather than the person
already in the conversation, which is the seam the module is explicitly built around, that is a page
to somebody with no question on it.

Why

server/src/agents/escalation.ts already has the sentence for this case:

const parsed = parameters.safeParse(args);
if (!parsed.success) {
  return "That was not put to anybody: say what you need a person to answer.";
}

and server/tests/agent-escalation.test.ts already asserts it, under
a call with nothing in it is refused as a sentence, by calling execute({}).

But question is z.string(), which accepts "" and a run of spaces. So the refusal fired only
when the field was missing. Present and empty is the same call spelled the other way, and it went
straight through.

The tool this sits beside and competes with for the same decision does check. handoff.ts:

const task = envelope.task?.trim() ?? "";
if (!task) {
  return refuse(from, target, "no_task",
    "Nothing was sent: a handoff has to say what the other Bot is being asked to do.");
}

A blank task handed to a Bot is refused, recorded as no_task, and answered with a sentence. A blank
question put to a person was not.

The fix

Trim the question, and refuse it with the sentence that was already written for this. The trimmed
text is then what travels to the route and onto the row — for the same reason a handoff sends the
trimmed task: what is recorded should be what was asked, not the spacing around it.

Deliberately not changed: why stays optional and is left as the model wrote it, and every other
path through the tool is untouched.

Measured

bun test server/tests/agent-escalation.test.ts

• Against origin/main with only the new tests applied: 8 pass, 4 fail.
  • an empty question is refused as a sentence — returns the "That was put to…" sentence.
  • a question of spaces is refused as a sentence — same.
  • a blank question reaches nobody and leaves no row saying it did — the route is called and an
    agent.escalated row is written.
  • a question with room around it is recorded as the question — the row carries
    " which account? ".
• With the change: 12 pass, 0 fail.

The pins are the eight cases already in that file, which pass before and after: a real question still
reaches the route and comes back with the PUT_TO marker the transcript matches on, the row still
carries the question and the why, a routine's escalation is still filed under its initiator, a
route that reaches nobody is still recorded as agent.escalation_failed, and execute({}) is still
refused.

Also run, all green:

• bun test server/tests/agent-handoff-tool.test.ts server/tests/agent-handoff.test.ts server/tests/control-refusal-end-to-end.test.ts server/tests/human-input-end-to-end.test.ts — 41 pass, 0 fail
• bun run --filter server typecheck — exit 0
• bunx biome check on both changed files — clean

Note on the CHANGELOG

A turn that used to end now does not, so there is an entry. It goes at the top of ## Unreleased,
the one line every entry goes at, so it will conflict with any other PR open against that anchor.
Happy to rebase whenever it suits you.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[davidmckayv]

Deep-reviewed against live code (correctness, governance, no vendor/secret/scale issues). Composed build+tests green. CHANGELOG/format rebase on CI-validated substance.