AGENTS.md git:20260618.ad9fd21A

Immutable. This exact content is served forever at /api/v1/blob/7900db6f6abf4613.

# Agent Instructions

This repository is a public reference repository for portable AI-agent phone-call workflows.

## Language

All repository-facing content must be written in English.

Do not add Chinese text or other non-English prose to repository files unless it is explicitly part of a user-provided example and the maintainer requests it.

## Scope

Keep the repository focused on AI-agent phone-call workflows:

- Agent Skills
- apps
- workflow plugins
- provider adapters
- scheduler recipes
- automation patterns
- safety patterns
- reference implementations

Do not turn this repository into a general voice-agent list, telephony vendor directory, or call-center software list unless the resource directly helps AI agents package phone-call workflows.

## Directory rules

- Put installable Agent Skills in `skills/`.
- Put runnable apps and integration demos in `apps/`.
- Put no-code and low-code workflow plugins in `plugins/`.
- Put long-form guidance in `docs/`.
- Keep provider-specific or host-specific details out of generic skill frontmatter.

## Skill design rules

Use the Agent Skills folder pattern:

```text
skill-name/
├── SKILL.md
├── references/
├── scripts/
└── assets/
```

Do not put `README.md` files inside installable skill directories. Keep skill folders limited to
runtime-facing skill resources. Put human-facing or long-form guidance under `docs/`, preferably
`docs/<skill-name>/README.md`, and keep reusable operational details in `references/` so agents can
load them through progressive disclosure.

Every `SKILL.md` must include YAML frontmatter with at least:

```yaml
---
name: skill-name
description: Clear description of what the skill does and when to use it.
---
```

For portability:

- keep `name` lowercase with hyphens
- keep `description` useful for discovery
- avoid host-specific frontmatter in generic skills
- put host-specific details in `references/`
- use progressive disclosure: keep `SKILL.md` focused and move long details to references

## App design rules

Apps must remain directly tied to AI-agent phone-call workflows.

- Prefer local fake servers, dry runs, or preview modes for tests and demos.
- Do not require live credentials or real outbound calls for default tests.
- Do not depend on unpublished private packages.
- Document setup, side effects, credentials, and cancellation or rollback behavior when applicable.
- Keep demo apps focused. If a demo becomes a reusable tool, keep it under `apps/` and document the supported workflow boundary.

## Plugin design rules

Plugins must remain directly tied to AI-agent phone-call workflows.

- Put no-code and low-code workflow-platform nodes, actions, connectors, templates, and recipes under `plugins/`.
- Document supported triggers, actions, required inputs, expected outputs, credentials, side effects, and rollback or disable behavior.
- Prefer preview, dry-run, or confirmation paths before a plugin can place calls or create recurring jobs.
- Keep platform-specific implementation details inside the plugin directory.

## Phone-call safety rules

Phone calls are real-world side effects.

Every skill, app, or adapter that can place a call must include rules for:

- explicit user intent
- E.164 phone numbers
- masking phone numbers in summaries
- no credential exposure
- no hidden recurring schedules
- no duplicate jobs
- clear cancellation behavior
- boundaries for medical, legal, financial, or emergency content

## Architecture rule

For recurring reminders, use this default architecture:

```text
Host scheduler handles recurrence.
Phone-call provider handles exactly one call per scheduled run.
```

Do not make provider-side recurrence mandatory. Many providers only support one-time calls or one-time scheduled calls.

## Validation

After editing, run:

```bash
python3 scripts/validate_repository.py
```