ZNAK CallOps

What it does

ZNAK CallOps is an experimental Agent Skill for a bounded service inquiry: ask about availability and price without booking or accepting a quote, then keep the source of every answer visible.

The working Python helper prepares the inquiry brief and checks locally normalized transcript evidence. It accepts only complete, exact provider-turn quotations. Unsupported or clipped quotes are rejected; different statements remain DISPUTED; missing answers remain UNKNOWN even when the call status says completed. QUOTED means the text is present in the supplied transcript, not that the statement is independently true.

CALL-E integration

The installable skill directs an authenticated agent host through the official MCP workflow: plan_call, then one authorized run_call, then get_call_run. Current tool schemas and original provider responses remain the source of truth. The host supplies the real transcript and proposes field associations for local checking. An optional executable Python adapter imports the official calle-ai 0.7.0 SDK and issues one calls.create request after explicit live opt-in. It defaults to a no-network preview, pins the official API origin, disables automatic retries, stores a private receipt, and reads status separately by saved call ID. A live CALL-E call has not been verified for this project.

Why it matters

A service inquiry can return a price with conditions, a later correction, or no answer at all. Flattening those into a confident summary hides the information a person needs before choosing or booking a service. CallOps keeps the quotation, its source, and the unresolved fields together.

Demonstration and tests

The public demonstration shows actual execution of the local Python helper on explicitly synthetic fixtures: inquiry preparation, complete quotes, an unknown price, conflicting price statements, and passing local tests. No phone call occurs in the video. The current suite has 16 passing tests with the optional SDK installed, including real SDK request serialization through an in-memory HTTP transport. These are local tests with synthetic HTTP responses; the video shows the 10 core helper tests recorded before the adapter tests were added. The repository contains runnable examples and setup instructions; the local path requires no credentials or paid service.

How AI and Codex are used

The live workflow is designed for an AI agent host and CALL-E's voice agent. The host proposes transcript quotations; deterministic code checks their identity. Codex helped implement the helper, tests, integration instructions, and demonstration. The bundled demonstration uses synthetic input, not an AI-generated record of a real conversation.

Limits

This is an experimental contribution. Quote identity does not verify speaker authenticity, field relevance, factual truth, transcript completeness, or whether a quoted time remains available. Distinct quotes require human review even when one may be a correction. The local helper cannot place or cancel a call, and the voice agent's compliance with the no-booking instructions has not been tested live.

Run it

Get the feat/znak-callops branch from the public repository. From the repository root run: python3 skills/znak-callops/scripts/callops.py demo python3 -m unittest discover -s skills/znak-callops/scripts -p 'test_*.py'

See skills/znak-callops/references/examples.md for expected results and the optional host integration path.

Built With

• agent-skills
• call-e
• mcp
• python