Imported from kirankumar-ritz/FRIDAY (workspace/AGENTS.md). Install upstream with npx skills add kirankumar-ritz/FRIDAY --skill workspace. Copyright stays with the author.

AGENTS.md — Agent Routing & Configuration

==========================================

Defines which agent handles which type of request.

The AgentRouter reads this file on every message.

Hot-reloads within 5 seconds — no restart needed.

Agent Roster

main

• Role: Default personal assistant. Handles all inbound messages.
• Persona: Reads SOUL.md for personality. Full tool access.
• Channel: All channels unless overridden below
• Memory: Full persistent memory access
• Tools: All 16 action modules + browser + shell + skills

proactive

• Role: Background agent. Runs scheduled jobs from HEARTBEAT.md.
• Trigger: Cron schedule — wakes, runs task, sleeps. No user message needed.
• Memory: Read-only memory access (doesn't store mid-task)
• Tools: All tools — but no TTS unless job explicitly requests it
• Output: Delivers results via the active channel (Telegram/voice/chat)

delegate ← Layer 3

• Role: Multi-task orchestrator. Decomposes complex requests into parallel subtasks.
• Trigger: Routed when request contains 2+ independent tasks (see patterns below)
• Memory: Full persistent memory access (reads before decompose, writes after synthesis)
• Tools: Spawns SubtaskRunners — each has access to all executor tools
• Output: One synthesised reply combining all subtask results
• Fallback: Single-task requests are handed to MainAgent with zero overhead

Routing Rules

Format: [trigger_pattern] → [agent_name]

Patterns are matched against the incoming message text (case-insensitive)

First match wins. Falls through to 'main' if no match.

── Delegate patterns — multi-task or compound requests ──────────────────────

Trigger delegate when the user asks for multiple independent things at once,

or uses compound connectors that imply parallel work.

[briefing|morning briefing|daily briefing] → delegate [and also|as well as|at the same time|simultaneously] → delegate [weather and news|news and weather|weather, news] → delegate [summarise.*and|summarize.*and|check.*and.*send|search.*and.*book] → delegate [compare.*and|research.*and|find.*and.*email|look up.*and] → delegate [give me.*and.*also|tell me.*and.*also] → delegate

── Proactive patterns — scheduled / reminder tasks ──────────────────────────

[remind|schedule|every|daily|morning|briefing] → proactive

── Main agent catch-all (default) ───────────────────────────────────────────

[code|script|python|debug|error|function|class|refactor] → main [default] → main

Delegate Routing Notes

The delegate agent is smart about overhead:

• If TaskDecomposer returns strategy="single" → hands off to MainAgent (no extra LLM call)
• If strategy="parallel" → runs subtasks in parallel threads (up to 6 workers)
• If strategy="sequential" → runs subtasks in DAG order (respects depends_on)
• Synthesis always produces ONE reply — user never sees "sub-agent" language

Agent Defaults

defaults:
  model: groq/llama-3.1-8b-instant     # override per-agent below
  max_tokens: 512
  temperature: 0.3
  context_turns: 6
  memory_recall_limit: 10

agents:
  main:
    model: groq/llama-3.1-8b-instant
    always_on: true

  proactive:
    model: groq/llama-3.1-8b-instant
    heartbeat_file: workspace/HEARTBEAT.md
    speak_results: false              # true = TTS output after each job

  delegate:
    model: groq/llama-3.1-8b-instant
    max_workers: 6                    # parallel subtask thread pool size
    subtask_timeout: 20               # seconds per subtask before timeout
    decomposer_max_tokens: 512        # tokens for decomposition LLM call
    synthesis_max_tokens: 600         # tokens for final synthesis call