Countercall

CounterCall

Pak Yanto took the morning off work, rode forty minutes to the Samsat counter, and was turned away because his photocopy was folio instead of A4. He lost a day's income, and he still did not know what else was missing for tomorrow.

Public service counters publish requirement lists that are incomplete, outdated, or quietly contradicted at the window. The one reliable source is the office's phone line, and that line is busy, IVR-gated, and answered on the fourth try. So people do not call. They travel, they queue, and they get turned away over one document.

This skill makes the call instead, and returns a checklist the person can screenshot and take with them.

It is a good fit for CALL-E's design: low-frequency, personal, high-stakes phone work against a number that a human would otherwise have to redial all morning.

Before you start

This skill places a real phone call to a real public office. Confirm with the user:

• which procedure they are asking about, in the office's own words
• which office, and the phone number in E.164 format, from the office's published page
• that they want a call placed now

Never infer a phone number. Not from a directory, not from a similar office, not by guessing a country code. A number enters the flow only when a human has read it off the office's own published page. Calling the wrong number means an automated caller reaches a stranger, which is the worst thing this skill can do.

Safety boundaries

This skill gathers counter requirements only.

• It does not submit an application, book an appointment, pay a fee, or commit the user to anything. It asks and reports.
• It does not give legal or immigration advice. If a clerk suggests an alternative route, report it verbatim as something the clerk said, never as a recommendation.
• A clerk's spoken answer is informational, not legally binding, and every rendered result says so. Requirements change and individual counters apply discretion.
• It states plainly at the start of the call that it is an automated assistant, and why it is calling. If the person declines to answer, it thanks them and ends the call. It does not persist, re-ask, or call back the same day.
• One call per office per procedure per day. These are public service lines staffed by people with a queue in front of them. The idempotency key enforces this; do not remove it.
• When the clerk is unsure, the result says unsure. It never fills a gap with what is typical, and it never presents a hedge as a fact. A person may travel across a city on the strength of this answer.
• Call offices during their published opening hours only. A public line ringing out at 22:00 is not a data point, it is a nuisance. The code does not check this: the seed file carries no hours, and call.mjs --live and bench.mjs --live will dial at any time of day. Checking the office's hours before either is the operator's job.

Read references/safety.md before adapting this skill to a new institution. Some callees are not appropriate targets for an automated caller at all.

What makes this different from a transcript

Most call skills return prose and let the reader interpret it. This one returns a pinned, validated object, and refuses to return anything else.

The pinned result_schema sets additionalProperties: false and constrains four fields to enumerations — payment_method, appointment_required, originals_or_copies and clerk_certainty, each of which includes a value for "the clerk did not know". If the call comes back shaped differently, the result is quarantined rather than rendered. A half-parsed checklist is worse than no checklist, because the user acts on it.

The full contract is in references/result-contract.md.

Workflow

1. Collect the request

Required: the procedure name, the office name, the office's phone number in E.164, and the published source URL the number came from.

Validate the number before anything else happens. E.164 is a plus, a non-zero country code digit, then 7 to 14 more digits: ^\+[1-9]\d{7,14}$. Reject anything else and ask the user. Never guess a country code. A local-format number reaching the dialler is how the wrong person gets called.

Run scripts/preflight.mjs to check the number and the contract without dialling.

2. Pick a transport, then make sure the contract is honest

There are two ways to reach CALL-E, and scripts/transport.mjs picks between them. Both place a real outbound call and both return the same validated shape, so nothing above the transport changes.

selectTransport() uses goals when COUNTERCALL_GOAL_ID is set and calls otherwise. COUNTERCALL_TRANSPORT=goals|calls forces one, and an unrecognised value throws rather than falling back — a typo must not silently dial on a path nobody chose.

Why calls is the default. Publishing a Goal is only possible inside CALL-E Chat: there is no POST /v1/goals, no MCP publish tool, and no action on the Goal detail page. That makes the Goals path dependent on a surface you may not be able to reach — during the 2026-09-02 CALL-E login suspension it was reachable by nobody. The Calls transport needs one credential and works regardless.

On calls, drift is impossible by construction. resultSchemaJSON() generates the request-scoped schema from the same CONTRACT object that validateResult checks the reply against, so the two cannot disagree. There is nothing to compare.

On goals, compare before every run. The published Goal owns the schema, so it can move underneath you:

import { diffContract, publishedRunSpec } from './scripts/_lib.mjs';

const goal = await client.goals.get(GOAL_ID);
const drift = diffContract(PINNED_CONTRACT, publishedRunSpec(goal));
if (drift.length) throw new ContractDrift(drift);   // refuse to dial

Use the publishedRunSpec(goal) helper, not goal.published_run_spec directly. The wire format and the docs spell it published_run_spec.result_schema; the TypeScript SDK camelCases it to publishedRunSpec.resultSchema. Reading only the documented name against the TS client returns undefined rather than throwing — the guard then concludes "no published spec" and refuses every single dial while looking like it is working. The helper reads both spellings. This cost us a day; it is written up in FEEDBACK.md at the repo root.

The CALL-E documentation recommends this comparison before deploying a variable change. Doing it on every run costs one API call and converts a silent failure into a loud refusal. Dialling a real person with a stale schema produces a result that looks fine and is quietly wrong.