Farmer Manager

First Step

Check what already exists before asking the user anything:

1. Scan existing farmers: Glob for .claude/agents/*-farmer.md
2. Scan scheduled triggers: Use /schedule (the schedule skill) with action list to see what's already running — or use RemoteTrigger directly if available
3. Present status: Show the user what farmers exist and which are scheduled
4. Ask what they need: Build a new farmer, or schedule an existing one?

Build a New Farmer

1. Source & MCP

Ask what source to farm. Then check if an MCP server is already connected — use ToolSearch to look for tools matching that service name. If nothing found, search for an official remote-hosted MCP server and help them connect it.

2. Probe the connector

Use ToolSearch to discover all tools for the service. Test 1-2 read-only tools to verify connectivity and understand data shape.

3. Gather details

Ask what to monitor (channels, senders, calendars, keywords, URLs) and any custom classification rules beyond vault defaults.

4. Generate the subagent

Create .claude/agents/<name>-farmer.md. Read an existing farmer (e.g., slack-farmer.md) as a structural reference.

Frontmatter:

---
name: <name>-farmer
description: Farms context from <source> into the vault
model: sonnet
permissionMode: acceptEdits
---

Rules:

• No tools field — MCP tools and ToolSearch inherited from parent
• No mcpServers — doesn't work for remote MCPs; ToolSearch handles discovery
• Always permissionMode: acceptEdits
• Name must end in -farmer — matches SubagentStop hook for auto-push
• Body is the entire system prompt — no CLAUDE.md, no parent context

Body must include:

1. Process — Discover tools via ToolSearch (search by service name), use generic tool references (never hardcoded prefixed names). Then: read config, check state, read source (24hr max), classify, deduplicate, git pull, write, update state. If no new data, skip to state update.
2. Classification Rules — Source data → vault entity types
3. Useful MCP Tools — | Tool | Purpose | table with bare names. Read/search tools only.
4. Guidelines — Signal-to-noise, naming, attribution

All farmer-created files must include source: farmer/<name> and farmed: <timestamp> in frontmatter.

5. Allow read-only MCP tools

Add the farmer's read-only MCP tools to .claude/settings.json under permissions.allow. Only add read/search/list tools — never write/send/delete.

6. Update watchlists

Add a section to context/watchlists.md for the new source's filters.

7. Offer to schedule

After building, ask if they want to schedule it now. If yes, follow the scheduling flow below.

Schedule a Farmer

Use /schedule to create a cloud scheduled trigger. The prompt should be /farm <name>.

Key details to confirm:

• Frequency: Daily (recommended for most), weekdays, hourly, etc.
• Time: Convert user's local time to UTC for cron
• MCP connector: Must be attached to the trigger
• Environment: full-access unless they specify otherwise

Farmers must run as subagents via /farm <name>, not inline prompts — this gives isolated context and triggers the SubagentStop hook for auto-push.

Other scheduling options

• Desktop scheduled: Desktop app > Schedule > New local task. Needs machine on.
• Session-scoped: /loop 30m /farm <name> — active session only.
• Manual: /farm <name> — on demand.