AgentSkillsCN

return-stack

你曾踏足之地,正是你未来可以重返的所在。

SKILL.md
--- frontmatter
name: return-stack
description: "Where you've been is where you can go back to."
license: MIT
tier: 1
allowed-tools:
  - read_file
  - write_file
related: [room, adventure, character, memory-palace, action-queue, card, session-log, prototype, debugging]
tags: [moollm, navigation, history, breadcrumbs, continuations, deoptimization, introspection]

Return Stack

"Where you've been is where you can go back to."


What Is It?

Return Stack treats navigation history as a first-class continuation — a stack of saved positions you can manipulate like browser history or a call stack.


The Metaphor

ProgrammingBrowserMOOLLM
Call stackHistoryReturn stack
Return addressBack buttonPrevious room
Stack frameTabRoom context
PushNavigateENTER
PopBackBACK

Commands

CommandEffect
ENTER roomPush current room, enter new one
BACKPop stack, return to previous room
FORWARDRedo after BACK (if available)
HISTORYShow the stack
BOOKMARKSave current position
GOTO bookmarkJump to saved position
STACKShow all open "tabs" (parallel stacks)
FORKCreate new tab from current position

Example Session

code
> ENTER workshop
[Stack: lobby]

> ENTER storage
[Stack: lobby → workshop]

> ENTER archive
[Stack: lobby → workshop → storage]

> BACK
Returning to storage...
[Stack: lobby → workshop]

> BACK
Returning to workshop...
[Stack: lobby]

> HISTORY
  1. lobby (start)
  2. workshop
  3. storage
  4. archive ← furthest
  
Current: workshop (position 2)

Bookmarks

Save positions for later:

code
> BOOKMARK "interesting-spot"
Bookmarked: workshop as "interesting-spot"

> ENTER research
> ENTER data-room
> ENTER sub-analysis
[Deep in the hierarchy]

> GOTO interesting-spot
Returning to workshop...
[Stack cleared, at bookmark]

Forking (Tabs)

Create parallel exploration paths:

code
> FORK
Created new tab from workshop.
Tab 1: lobby → workshop
Tab 2: workshop (active) ←

> ENTER experiment-A
[Tab 2: workshop → experiment-A]

> STACK
Tab 1: lobby → workshop
Tab 2: workshop → experiment-A ←

> TAB 1
Switching to Tab 1...
[Now at workshop via tab 1]

As Continuation

The return stack IS a continuation:

yaml
# Stored in character's pocket
return_stack:
  - path: "./lobby"
    context: {examining: "welcome-sign"}
  - path: "./workshop" 
    context: {crafting: "blueprint-v2"}
  - path: "./storage"
    context: {searching: "rare-materials"}
    
# Current position
current: "./archive"
context: {reading: "old-records"}

When you BACK, you don't just return to the room — you restore the context you had there.


Portable Journey

The stack travels with you:

code
> INVENT
Inventory:
  - notebook
  - pen
  - return_stack: [lobby → workshop → storage]

You can:

  • Save your journey to a file
  • Share it with others
  • Replay someone else's exploration
  • Branch from any point in their journey

HyperCard Heritage

HyperCard had:

  • Stacks of cards
  • "Go back" button
  • Breadcrumb trail
  • Bookmarks

MOOLLM extends this:

  • Rooms as cards
  • BACK command
  • Return stack as data
  • Bookmarks as saved positions
  • FORK for parallel exploration

Implementation

yaml
# character.yml
name: explorer
location: ./archive

navigation:
  return_stack:
    - room: ./lobby
      entered: "2024-01-15T10:00:00"
    - room: ./workshop
      entered: "2024-01-15T10:05:00"
      context:
        active_task: "crafting"
    - room: ./storage
      entered: "2024-01-15T10:15:00"
      
  bookmarks:
    interesting-spot:
      room: ./workshop
      context: {task: "blueprint-review"}
    start:
      room: ./lobby
      
  forward_stack: []  # After BACK, stores where you came from

Dynamic Deoptimization

The Self programming language (source of our prototype inheritance) pioneered dynamic deoptimization: aggressively inlining code for performance, then reconstructing the "logical" call stack on demand when debugging.

The LLM does this naturally for narrative:

SelfMOOLLM
Inlined bytecodeFlattened conversation
Deoptimized framesReconstructed causality
Breakpoint trigger"How did we get here?"
Stack traceCausal chain from evidence

The stack isn't explicitly maintained, but it's recoverable.

How It Works

When you ask for history, the LLM examines:

yaml
evidence_sources:
  - session_log: "Append-only narrative trail"
  - room_state: "Accumulated changes"
  - character_location: "Current position"
  - file_timestamps: "Order of modifications"
  - chat_context: "Recent decisions"

And synthesizes a virtual stack trace:

code
Deduced navigation:
1. lobby (start)
2. workshop (examining blueprints)
3. storage (found locked chest)
4. workshop (returned for key) ← BACK
5. storage (unlocked chest)
6. archive (following map) ← current

This is introspection without instrumentation — the same insight that made Self's debugger magical.


Dovetails With


Protocol Symbols

code
RETURN-STACK      — Navigation history as data
BACK / FORWARD    — Stack manipulation
BOOKMARK / GOTO   — Saved positions
FORK              — Parallel exploration
HYPERCARD-HIERARCHY — The room/card model

See: PROTOCOLS.yml