ZNAK CallOps

Turn a bounded service inquiry into a CALL-E plan and an inspectable result. This skill adds a small executable evidence checker: a completed call is not a verified answer, and a quote present in a transcript is not independent proof of its truth.

Setup and verification

Requires Python 3.10+ for the local helper, and an Agent Skills-compatible host with authenticated CALL-E MCP tools or Python 3.11+ with the optional official SDK adapter for live calls. Copy this whole directory to the host's skill directory. For Codex, use ~/.codex/skills/znak-callops/. Use references/call-e.md to connect the official provider. The helper uses only Python's standard library and never connects to a network.

From the repository root, run the synthetic no-call demonstration:

python3 skills/znak-callops/scripts/callops.py demo
python3 -m unittest discover -s skills/znak-callops/scripts -p 'test_*.py'

Read references/examples.md for inputs, expected results, and manual verification. These tests exercise local code, not the CALL-E service. The live host integration has not been verified by a real call in this contribution. For the optional executable SDK path, read references/sdk-dispatch.md. dispatch.py defaults to a no-network preview and imports the official SDK only after explicit live/read opt-in. No live CALL-E call has been verified through either path.

Workflow

1. Establish the user's exact question, service scope, preferred time window, known destination, language, and permission to call that recipient for this purpose. Read references/safety.md before dispatch.
2. Use prepare-goal with the resolved request to produce the bounded call brief. Show the masked destination, purpose, and inquiry only; do not book, accept a quote, pay, cancel, or disclose additional personal details. A request to prepare a plan alone does not authorize a phone call.
3. Invoke CALL-E plan_call with the user's latest message verbatim in user_input. Pass only known fields; use the prepared brief as goal. Follow the live tool schema. A plan does not place a call.
4. After the user clearly authorizes this call and the plan is ready, invoke run_call once with exactly the returned plan_id and confirm_token. Keep those opaque values out of public artifacts. Retain the returned run_id in the host's private task record. Follow the recovery rules in the provider reference; a timeout is not permission to redial.
5. Read get_call_run until a terminal outcome or a monitoring pause. If transcript turns are available, preserve their exact text, order, and speaker. Map them to the helper's documented input. The provider's response schema is discovered at runtime; this normalized packet is local, not a CALL-E wire format.
6. Propose full provider-turn quotations for availability, quoted_price, and follow_up. Never clip qualifying words. Run validate-result on that packet. If the provider exposes only a summary, leave the unsupported fields UNKNOWN.
7. Return the checked quotes and their source pointers. Review relevance and ambiguity: the checker verifies quote identity, not semantic interpretation, speaker authenticity, provider honesty, or that a time remains available. Leave booking and other commitments to a separate user decision.

Result meaning

• QUOTED: one complete provider turn is present in the supplied transcript. This is a reported statement, not a verified real-world fact.
• DISPUTED: distinct supported quotes remain for one field; show both for review.
• UNKNOWN: no acceptable source quote; show the rejected evidence reason.

Keep call status, source mode, and field status separate. COMPLETED with no transcript still yields UNKNOWN answers. source_mode: synthetic identifies the bundled demonstration; a caller-supplied live label is not independent attestation. The helper has no booking API and cannot enforce what a remote voice model says. Any observed commitment request or boundary breach must be surfaced to the user.