action-queue（动作队列）

Action Queue

*"Queue up what to do next. The Sims showed us the way."*

What Is It?

Action Queue is The Sims-inspired task scheduling: instead of doing one thing at a time, you queue up a sequence of actions that execute in order.

While return-stack tracks *where you've been*, action-queue tracks *what you're going to do*.

The Sims Connection

In The Sims, you click multiple actions and they queue up:

Sim's Queue:
  1. Go to fridge     [executing]
  2. Get snack
  3. Eat snack
  4. Watch TV
  5. Go to bed

The Sim works through them in order. You can:

• Add more actions
• Cancel any action
• Reorder the queue
• Insert urgent actions at the front

Objects push actions too! When you try to eat, the fridge pushes "get food" onto the queue. The food pushes "prepare" and "cook". The stove pushes "serve". Each object dangles the next carrot.

MOOLLM does the same for agents and characters.

The Food Chain Pattern

Objects guide you through multi-step processes by pushing actions:

> DO eat

Fridge intercepts:
  "You need food first."
  → pushes GET-FOOD to front

Queue: [GET-FOOD, EAT]

Food intercepts:
  "This needs preparation."
  → pushes PREPARE to front

Queue: [PREPARE, GET-FOOD, EAT]

Stove intercepts:
  "Needs cooking."
  → pushes COOK to front

Queue: [COOK, PREPARE, GET-FOOD, EAT]

Agent executes in order:
  1. COOK (at stove)
  2. PREPARE (at counter)
  3. GET-FOOD (from fridge)
  4. EAT (finally!)

Dangling a carrot in front of a donkey. Each object advertises what it needs, pushing prerequisites onto the queue. The agent surfs the possibility space of advertisements, guided step by step.

Why This Works

• No hardcoded recipes — objects know their own prerequisites
• Emergent behavior — complex sequences from simple rules
• Interruptible — cancel anytime, queue adapts
• Discoverable — agent learns by trying
• Adapts to context — different appliances, different paths
• Handles routing — crowded kitchen? Find another way

Routing Around Problems

> DO cook

Stove advertises COOK... but:
  "Blocked! Alice is using the stove."
  → Checks alternatives
  → Microwave advertises COOK (score: 70)
  → pushes USE-MICROWAVE instead

Queue: [USE-MICROWAVE, PREPARE, GET-FOOD, EAT]

If the kitchen is crowded:

Kitchen has 4 Sims blocking workspaces.
Counter: "Blocked by Bob"
Stove: "Blocked by Carol"
Fridge: "Accessible!"

Agent routes around obstacles:
  → Wait for counter? (cost: time)
  → Use outdoor grill? (cost: distance)
  → Order takeout? (cost: money)

Highest-scored alternative wins.

Emergent traffic flow. No pathfinding algorithm needed — objects simply advertise availability, agents pick the best option. Crowded rooms naturally disperse as scores drop.

This is how The Sims creates rich behavior from simple object definitions.

Commands

Example Session

> DO examine-workbench
Added: examine-workbench

> DO craft-tool
Added: craft-tool

> DO go-north
Added: go-north

> QUEUE
Action Queue:
  1. examine-workbench
  2. craft-tool
  3. go-north

> NEXT
Executing: examine-workbench
You examine the workbench. It has blueprints and tools.

> URGENT check-inventory
Inserted at front: check-inventory

> QUEUE
Action Queue:
  1. check-inventory    ← urgent
  2. craft-tool
  3. go-north

Autonomous Execution

For autonomous agents, the queue runs automatically:

# agent.yml
mode: autonomous
action_queue:
  - examine: hypothesis.yml
  - analyze: data/
  - write: findings.md
  - notify: user

execution:
  auto_advance: true
  pause_on_error: true
  pause_on_user_input: true

Agent works through queue until done, paused, or interrupted.

Queue + Stack Together

Return stack and action queue complement each other:

Together they form a complete temporal model:

• Stack = memory of the past
• Queue = intentions for the future
• Current = the present moment

Compound Actions

Queue items can be compound:

> DO [go-to-library, find-book, read-chapter-1]
Added compound action (3 steps)

> QUEUE
Action Queue:
  1. [go-to-library, find-book, read-chapter-1]  (compound)
  2. write-summary

Compound actions expand when executed:

> NEXT
Expanding compound action...
Action Queue:
  1. go-to-library     [executing]
  2. find-book
  3. read-chapter-1
  4. write-summary

Conditional Actions

Actions can have conditions:

action_queue:
  - action: craft-tool
    if: has_materials

  - action: gather-materials
    if: not has_materials

  - action: test-tool
    after: craft-tool

The queue adapts based on state.

Integration with Advertisements

Objects advertise actions. The agent queues the best ones:

Agent enters workshop.

Objects advertise:
  workbench: CRAFT (90), EXAMINE (50)
  bookshelf: READ (70)
  door: EXIT (40)

Agent queues:
  1. CRAFT at workbench (highest score)
  2. READ at bookshelf (if time permits)

Advertisements suggest. Queue commits.

Implementation

# character.yml
name: researcher
location: ./lab

action_queue:
  - action: analyze-sample
    target: sample-A
    added: "2024-01-15T10:00:00"

  - action: write-notes
    target: notebook
    added: "2024-01-15T10:01:00"

  - action: report-findings
    target: user
    added: "2024-01-15T10:02:00"

queue_state:
  paused: false
  current_index: 0
  mode: manual  # or autonomous

Dovetails With

• Return Stack — Past complements future
• Advertisement — Objects suggest, queue commits
• Room — Where actions happen
• Speed of Light — Execute multiple actions per epoch
• Coherence Engine — Orchestrates queue execution

Protocol Symbol

ACTION-QUEUE

Invoke when: Scheduling sequences of actions, autonomous agent behavior.

See: PROTOCOLS.yml