zo-dispatcher

分类: 工具与效率 | 上传者: JackALaing | 下载: 0 | 版本: v1.0（最新）
Create, edit, debug, and manage zo-dispatcher agents. Covers dispatch patterns, agent file format, prompt authoring, webhook setup, channel configuration, scheduling, and debugging. Triggers on: 'create an agent', 'edit an agent', 'schedule a task', 'add a webhook', 'cron job', 'recurring task', 'debug agent', 'agent not running', 'scheduled task', 'zo-dispatcher'.
更新日志: Source: GitHub https://github.com/JackALaing/zo-dispatcher
目录结构

当前层级: skill/
📁 references/
- 📄 cli.md 3.4 KB
- 📄 editing-and-debugging.md 3.8 KB
- 📄 scheduled-agents.md 1.6 KB
- 📄 transforms.md 4.8 KB
- 📄 webhook-agents.md 8.5 KB
📄 SKILL.md 15.4 KB
SKILL.md

---
name: zo-dispatcher
description: "Create, edit, debug, and manage zo-dispatcher agents. Covers dispatch patterns, agent file format, prompt authoring, webhook setup, channel configuration, scheduling, and debugging. Triggers on: 'create an agent', 'edit an agent', 'schedule a task', 'add a webhook', 'cron job', 'recurring task', 'debug agent', 'agent not running', 'scheduled task', 'zo-dispatcher'."
compatibility: Created for Zo Computer
metadata:
  author: JackALaing
---

# zo-dispatcher

zo-dispatcher is a custom agent dispatcher that replaces Zo's built-in agents. Do NOT use `create_agent`, `edit_agent`, or `delete_agent` when this skill is installed.

Key paths (relative to the zo-dispatcher install directory):

- Agent definitions: configured via `agents_dir` in `config/config.json`
- Config: `config/config.json`
- State: SQLite database at path configured via `db_path`
- Logs: `/dev/shm/zo-dispatcher.log` (indexed by Loki)
- Transforms: configured via `transforms_dir`
- CLI: `dispatcher-cli` (available on PATH after `pip install .`)

The dispatcher reads agent files, fires `/zo/ask` calls on schedule or webhook triggers, and routes output to notification channels.

**Reference docs** (load only when needed):

- `references/scheduled-agents.md` — MANDATORY before creating any Cron agent
- `references/webhook-agents.md` — MANDATORY before creating any webhook, Sentinel, or Inbox agent
- `references/editing-and-debugging.md` — MANDATORY before debugging or modifying agent lifecycle
- `references/transforms.md` — MANDATORY before registering a webhook source (includes registration commands and copy-paste-ready transform scripts for common sources)
- `references/cli.md` — CLI command reference

---

## Dispatch Patterns

Every agent maps to one of four dispatch patterns. Each pattern corresponds to a distinct configuration shape.

| Pattern | Config | One-Liner |
|---------|--------|-----------|
| **Cron** | `trigger: schedule` | Runs on a clock, no event input |
| **Trigger** | `trigger: webhook` | Fires immediately per event |
| **Sentinel** | `trigger: both`, `defer_to_cron: false` | Fires immediately per event AND runs on schedule independently |
| **Inbox** | `trigger: both`, `defer_to_cron: skip_if_empty` or `always_run` | Queues events, drains on schedule |

For Cron details, see `references/scheduled-agents.md`. For Trigger, Sentinel, and Inbox details, see `references/webhook-agents.md`.

### Choosing a Pattern

```
Does the agent need external events at all?
├── No → Cron (trigger: schedule)
└── Yes →
    Is the event urgent (needs response in minutes)?
    ├── Yes → Trigger (trigger: webhook)
    │         Also need periodic independent work? → Sentinel (trigger: both, defer_to_cron: false)
    └── No → Inbox (trigger: both)
              Does the agent have work beyond processing events?
              ├── No → defer_to_cron: skip_if_empty
              └── Yes → defer_to_cron: always_run
```

### Cost Mental Model

| Pattern | Cost Driver | Typical Range |
|---------|-------------|---------------|
| Cron | 1 LLM call per scheduled tick | Predictable, based on frequency |
| Trigger | 1 LLM call per event, capped by `max_runs` | Depends on event volume |
| Sentinel | 1 LLM call per event + 1 per scheduled tick | Highest (reactive + proactive) |
| Inbox `skip_if_empty` (heartbeat, 30-min) | ~48 LLM calls/day max, only when events exist | Low-Medium |
| Inbox `skip_if_empty` (daily) | 1 LLM call/day, only when events exist | Very Low |
| Inbox `always_run` (daily) | 1 LLM call/day, always | Low (predictable like Cron, but also drains events) |

Inbox is dramatically cheaper than Trigger or Sentinel for high-volume sources because it amortizes the baseline token overhead (system prompt, tool loading, context bootstrapping) across all events in a batch.

### Anti-Patterns

1. **Trigger on high-volume sources.** If a source fires 100+ events/day, use Inbox instead.
2. **Sentinel or Trigger without `max_runs`.** Any agent with webhook triggers can run away without a budget cap.
3. **Inbox with too-short drain interval.** Draining every 1-2 minutes defeats the purpose of batching. Use 15-30 min minimum.
4. **Single agent trying to do both urgent filtering and batch processing.** Use two agents: Trigger for urgency, Inbox for batching. They can listen to the same events.
5. **Forgetting transform scripts.** Transforms strip payloads to essentials and can drop events entirely by returning `None` — the cheapest possible filter.
6. **Inbox prompt without `{{ queue_file }}`.** The agent has no way to read queued events. The parser warns, but the agent will fire and produce a generic response with no event data.
7. **Prompts that assume memory across runs.** Each dispatch is a fresh `/zo/ask` call. The agent doesn't remember previous runs. If it needs continuity, the prompt must include explicit file reads for state.
8. **Agents writing to the same file without coordination.** Two agents that both write to the same output file on overlapping schedules will clobber each other. Use distinct output files or make one agent depend on the other's output.
9. **Transform scripts doing LLM-level reasoning.** Transforms are cheap deterministic filters. If your transform needs to "understand" the payload, that work belongs in the agent prompt. Keep transforms to field extraction, keyword matching, and sender filtering.
10. **Prompts without context bootstrapping.** The agent wakes up blank. If it needs workspace context (files, config, recent state), the prompt must explicitly instruct it to read those files.

---

## Agent File Format

Each agent is a markdown file with YAML frontmatter and a markdown body (the prompt).

### Location & Organization

Files live under `agents_dir`, optionally organized in subfolders. Any directory structure works — the example below is just one convention:

```
agents/
├── schedules/
│   ├── daily-summary.md
│   └── weekly-report.md
└── webhooks/
    ├── github-issue.md
    └── readwise-daily.md
```

Agent IDs are path-based: `schedules/daily-summary`, `webhooks/github-issue`.

### Frontmatter Fields

| Field             | Relevant Triggers  | Required | Default        | Description                                                                       |
| ----------------- | ------------------ | -------- | -------------- | --------------------------------------------------------------------------------- |
| `title`           | all                | yes      | —              | Display name                                                                      |
| `trigger`         | all                | yes      | —              | `schedule`, `webhook`, or `both`                                                  |
| `rrule`           | `schedule`, `both` | yes      | —              | RFC 5545 recurrence rule                                                          |
| `event`           | `webhook`, `both`  | yes      | —              | Webhook event filter: string or list of strings (dot-notation). List = match any. |
| `persona`         | all                | no       | —              | Persona ID (`per_XXX`).                                                           |
| `model`           | all                | no       | config default | Model ID                                                                          |
| `notify_channel`  | all                | no       | —              | `<channel>` or `<channel>/<sub_channel>`                                          |
| `notify`          | all                | no       | `errors`       | `always`, `errors`, `never`                                                       |
| `timeout`         | all                | no       | config default | Seconds for `/zo/ask` timeout                                                     |
| `retry_delays`    | all                | no       | config default | List of retry delay seconds                                                       |
| `max_runs`        | all                | no       | unlimited      | Max dispatches in `max_runs_window`                                               |
| `max_runs_window` | all                | no       | `3600`         | Budget window in seconds                                                          |
| `defer_to_cron`   | `both` only        | no       | `false`        | `false`, `skip_if_empty`, or `always_run`. Requires `trigger: both`.              |
| `active`          | all                | no       | `true`         | Set false to disable                                                              |

### Template Variables

| Variable | Available | Description |
|----------|-----------|-------------|
| `{{ payload }}` | webhook | JSON payload, pretty-printed |
| `{{ event_type }}` | webhook | Extracted event type |
| `{{ date }}` | all | Current date (e.g., "2026-03-04") |
| `{{ timestamp }}` | all | Current ISO 8601 timestamp |
| `{{ agent_id }}` | all | Namespaced ID (e.g., "schedules/daily-summary") |
| `{{ queue_file }}` | defer_to_cron | Snapshot JSONL path when events exist, or `"No events queued."` when empty (`always_run`) |

### Event Matching (Dot-Notation)

| Agent `event` | Incoming source + event_type | Match? |
|---------------|------------------------------|--------|
| `github` | github / any event | Yes (source-level) |
| `github.opened` | github / `opened` | Yes (exact) |
| `github.open` | github / `opened` | No (not a prefix) |
| `todoist.item:added` | todoist / `item:added` | Yes (exact) |

Multiple agents can listen to different events from the same source. The `event` field accepts a YAML list for multi-event matching (logical OR). Patterns can span multiple webhook sources for cross-source batching.

---

## Notification Channels

### `notify_channel` Format

- `sms` — SMS to the user
- `email` — Email to the user
- `telegram` — Telegram to the user
- `discord/<channel-name>` — Discord, routed to the named channel in your server (e.g., `discord/general`, `discord/alerts`) - requires zo-discord
- Omit for silent agents (no notifications)

### Custom Channels

Configured in `config/config.json`:

```json
{
  "channels": {
    "discord": {
      "url": "http://localhost:8787/notify"
    }
  }
}
```

Custom channels receive a POST with `{"title": "...", "content": "...", "conversation_id": "..."}`. For Discord, the sub-channel is passed as `channel_name` when using `channel/sub_channel` notation.

### Builtin Channels

SMS, email, and Telegram need no config. Set `notify_channel: sms` (or `email`, `telegram`) in the agent frontmatter.

**Advisory:** SMS and Telegram are single-threaded — all agent output lands in one conversation. Use Discord for multi-agent notifications. SMS/Telegram work best for single-agent alerts.

### Notification Levels

- `notify: always` — notify on success and failure
- `notify: errors` — only on failure (default)
- `notify: never` — no notifications
- If `notify_channel` is not set, the agent runs silently regardless of `notify`

---

## Concurrency & Cost Control

Three-layer throttling model:

1. **Per-source serialization** — Events from the same webhook source dispatch sequentially. Cross-source events run in parallel.

2. **Global concurrency** (`max_concurrent_dispatches` in config) — Caps total simultaneous `/zo/ask` calls across **webhook agents only**. Scheduled agents bypass the semaphore. Default: 5.

3. **Per-agent dispatch budget** (`max_runs` / `max_runs_window`) — Caps how often a specific agent fires. Excess triggers are dropped with a system notification. **Strongly recommended for webhook and dual-trigger agents.** For `trigger: both`, the budget counts all runs regardless of trigger source.

---

## Writing Your First Agent

Create a file in your agents directory, e.g., `agents/webhooks/readwise-daily.md`:

```yaml
---
title: Readwise Daily Digest
trigger: both
event: readwise.reader.any_document.created
defer_to_cron: skip_if_empty
rrule: |-
  DTSTART;TZID=America/New_York:20260101T090000
  RRULE:FREQ=DAILY;BYHOUR=9;BYMINUTE=0
notify_channel: sms
notify: always
active: true
---

The following articles were saved to Readwise in the last 24 hours:

{{ queue_file }}

For each article:
1. Write a 2-sentence summary
2. Assign 1-3 tags based on topic
3. Note connections to previously saved articles on similar themes

Produce a brief daily reading digest with the summaries, tags, and any cross-references.
```

This is an **Inbox** agent: webhook events from Readwise are silently queued, then drained once daily at 9am. If nothing was saved that day, the agent doesn't run (`skip_if_empty`).

Test it manually:

```bash
dispatcher-cli agent run webhooks/readwise-daily
```

---

## Config Reference

`config/config.json`:

| Key                           | Default                   | Description                                                                           |
| ----------------------------- | ------------------------- | ------------------------------------------------------------------------------------- |
| `agents_dir`                  | `./agents`                | Path to agent markdown files                                                          |
| `db_path`                     | `./data/dispatcher.db`    | Path to SQLite database                                                               |
| `zo_api_url`                  | `https://api.zo.computer` | Zo API endpoint                                                                       |
| `default_model`               | —                         | Default model ID for agents without `model` (user must set)                           |
| `poll_interval_seconds`       | `60`                      | How often to check for due scheduled agents                                           |
| `transforms_dir`              | `./transforms`            | Path to transform scripts directory                                                   |
| `webhook_port`                | `8790`                    | HTTP server port                                                                      |
| `max_concurrent_dispatches`   | `5`                       | Global cap on simultaneous `/zo/ask` calls                                            |
| `zo_ask_timeout_seconds`      | `1800`                    | Default timeout for `/zo/ask` calls                                                   |
| `empty_response_retry_delays` | `[15, 30, 60]`            | Retry delays for empty/failed responses                                               |
| `session_pool_retry_delays`   | `[15, 30, 60, 120]`       | Retry delays for session pool errors                                                  |
| `notification_hours`          | `{"start": 9, "end": 21}` | Business hours window for all notifications. Set `{"start": 0, "end": 24}` to disable |
| `notification_timezone`       | `America/New_York`        | Timezone for business hours                                                           |
| `webhook_rate_limit`          | `60`                      | Max webhook requests per source per 60-second window                                  |
| `dedupe_hours`                | `24`                      | How long to remember webhook event IDs                                                |
| `system_notification_channel` | `sms`                     | Channel for system-level alerts                                                       |
| `channels`                    | `{}`                      | Custom channel configurations                                                         |
登录后下载/点赞/收藏 ❤ 2 | ★ 0
zo-dispatcher

目录结构

SKILL.md

举报内容

提示
