🔥 Embr DSL: A Shell-Inspired Chat Abstraction for Language Programs

🧠 Overview

Embr is a domain-specific language (DSL) designed for interactive, composable natural language pipelines, inspired by the syntax and philosophy of Unix shell scripting. 🐚

It treats a conversational state (chat log) as a mutable object that can be transformed, extended, and queried using intuitive operators like <<, >>, |, and |=.

✨ Core Design Principles

• 🧱 Declarative + Composable: Like piping commands in a shell.
• 🧑‍🎤 Role-aware: Messages are typed with @ (e.g., user @ "text").
• 🛠️ Reducers as Functions: Language models and tools act like reducers.
• 🔁 Stateful by Default: All operators mutate a shared conversation state (Embr).

---

🧮 Operator Table

🔣 Operator	📝 Description	💡 Example Usage
@	Assign role to a message	user @ "message"
<<	Append message to chat (in-place)	chat << "text"
>>	Also appends in-place.	updated_chat = chat >> gpt
>	Send chat to model, get response	chat > gpt
|	Apply model and append output	chat = chat << "hello" | gpt
| =	Register reducer into chat's pipeline	chat |= gpt | claude

🔄 Chat Composition and Transfer Patterns

Operator Precedence

In Embr, operators follow standard Python precedence rules. For our operators, from highest to lowest:

Precedence	Operators	Description
Highest	@	Role annotation
High	>	Process through model
Medium	>>	Chat transfer
Low	|	Pipe and modify
Low	<<	Append message

This affects how expressions are evaluated when chained together.

Chat Transfer Pattern

The chat transfer pattern allows you to process a chat through a model and transfer the results to another chat:

# Process task_chat and transfer results to global_chat
(task_chat > gpt) >> global_chat

This pattern creates a clean separation between task-specific conversations and your main conversation context.

┌────────────┐    ┌─────┐    ┌─────────────┐    ┌────────────┐
│ task_chat  │ >  │ gpt │ -> │ response    │ >> │ global_chat│
└────────────┘    └─────┘    │ (Spark)     │    └────────────┘
                             └─────────────┘

Parenthesis-Free Alternatives

For cleaner code without parentheses, several alternatives exist:

python
# Manual two-step approach
response = task_chat > gpt
global_chat << response

# Using helper methods (proposed)
task_chat.process(gpt).append_to(global_chat)

# Using a new syntax pattern (proposed)
task_chat > gpt > global_chat

The last approach would require updating the implementation of the > operator to handle different right-hand types.

---

🚀 Example Use Cases

1. 🛠️ Manual append + model roundtrip

python
chat = Embr()
chat << user @ "hello"
response = chat > gpt
chat << response

2. 🧪 Pipe-style chaining

python
chat = Embr()
chat = chat << "what is the weather?" | gpt

3. 🔗 Register reducers to the chat

python
chat |= gpt | claude | other_ai
chat.apply_reducers()

4. ❓ Conditional reducers

python
def gpt_if_question(chat):
if "?" in chat.last.content:
return gpt(chat)


chat |= gpt_if_question
chat << "this is a statement"
chat << "what is this?"
chat.apply_reducers()

5. 🤖 Multi-agent routing

python
chat |= on_call @ gpt | on_call @ images

---

🧩 Functional Patterns

Currying and Composition

Embr uses a functional programming pattern similar to currying, where functions like gpt are designed to:

1. Accept an Embr object: Each reducer (like gpt, claude) takes a chat as input
2. Return a result: Either a Spark (message) or a modified Embr object
3. Enable operator-based application: The same function works with >, >>, and | operators

This pattern enables multiple styles of interaction:

# Direct application returning a response
response = chat > gpt

# Application with returned response
response = chat >> gpt

# Pipeline style with in-place modification
chat = chat | gpt

Embr brings the expressiveness of Unix pipelines 🧵 to AI-assisted conversation flows 💬, making it easy to test, iterate, and layer intelligence like shell commands.