Block Format

llmist uses a simple block format that works with any text model. This format uses marker prefixes to delimit gadget calls and their parameters—no native tool calling or structured outputs required.

Basic Structure

A gadget call consists of a start marker, parameters, and an end marker:

!!!GADGET_START:GadgetName
!!!ARG:parameter_name
parameter_value
!!!ARG:another_param
another_value
!!!GADGET_END

Markers

Marker	Purpose
`!!!GADGET_START:Name`	Begins a gadget call with the gadget name
`!!!ARG:pointer`	Declares a parameter (value on following line(s))
`!!!GADGET_END`	Ends the gadget call

Parameter Syntax

Parameters use !!!ARG: followed by a JSON Pointer path (without leading /) to specify where to place the value.

Simple Parameters

!!!ARG:filename
calculator.ts
!!!ARG:language
typescript

Result: { filename: "calculator.ts", language: "typescript" }

Nested Objects

Use / to create nested structures:

!!!ARG:config/timeout
30
!!!ARG:config/retries
3

Result: { config: { timeout: 30, retries: 3 } }

Arrays

Use numeric indices (0-based):

!!!ARG:items/0
first
!!!ARG:items/1
second
!!!ARG:items/2
third

Result: { items: ["first", "second", "third"] }

Arrays of Objects

Combine array indices with nested paths:

!!!ARG:users/0/name
Alice
!!!ARG:users/0/age
25
!!!ARG:users/1/name
Bob
!!!ARG:users/1/age
30

Result: { users: [{ name: "Alice", age: 25 }, { name: "Bob", age: 30 }] }

Automatic Type Coercion

Single-line values are automatically coerced to appropriate types:

Value	Coerced Type	Result
`true`	boolean	`true`
`false`	boolean	`false`
`42`	number	`42`
`3.14`	number	`3.14`
`-17`	number	`-17`
`hello`	string	`"hello"`
(multiline)	string	preserved as-is

Multiline Values

Multiline content is preserved exactly as written:

!!!ARG:code
function hello() {
  console.log("Hello, World!");
  return { key: "value" };
}
!!!ARG:filename
example.ts

Internal newlines are preserved
A single trailing newline is stripped
Great for code, documentation, or any multi-line content

Dependencies

Gadgets can specify dependencies on other gadgets using invocation IDs. This enables DAG (Directed Acyclic Graph) execution: independent gadgets run in parallel, dependent gadgets wait.

Syntax

!!!GADGET_START:GadgetName:invocation_id:dep1,dep2

Component	Required	Description
`GadgetName`	Yes	Name of the gadget to call
`invocation_id`	Optional	Unique ID for this call (auto-generated if omitted)
`dep1,dep2`	Optional	Comma-separated IDs this gadget depends on

Example

!!!GADGET_START:FetchData:fetch_users
!!!ARG:url
/api/users
!!!GADGET_END

!!!GADGET_START:FetchData:fetch_orders
!!!ARG:url
/api/orders
!!!GADGET_END

!!!GADGET_START:MergeData:merge_all:fetch_users,fetch_orders
!!!ARG:format
json
!!!GADGET_END

In this example:

fetch_users and fetch_orders execute in parallel (no dependencies)
merge_all waits for both to complete before executing
If fetch_users fails, merge_all is automatically skipped

Customizing Markers

You can customize the marker prefixes if needed:

await LLMist.createAgent()
  .withGadgetStartPrefix('<<<START:')
  .withGadgetEndPrefix('<<<END:')
  .withGadgetArgPrefix('@param:')
  .ask('...');

This would expect:

<<<START:FloppyDisk
@param:filename
DOOM.ZIP
@param:megabytes
50
<<<END:

Implicit Termination

The end marker (!!!GADGET_END) is optional. A gadget block terminates when:

An explicit !!!GADGET_END is encountered
A new !!!GADGET_START: begins (implicit termination)
The stream ends (finalization)

This allows streaming parsers to handle incomplete or truncated responses gracefully.

Error Handling

The parser reports errors for:

Duplicate pointers: Same path specified twice in one gadget
Array index gaps: Non-sequential array indices (e.g., items/0 then items/5)
Invalid array indices: Negative or non-numeric indices

Parse errors are captured in the parseError field while parametersRaw preserves the original text for debugging.

EBNF Grammar Reference

For implementers, here’s the formal grammar:

(* Top-level structure *)
gadget_block     = start_marker , header_line , newline , parameters , [ end_marker ] ;

(* Markers - these are configurable *)
start_marker     = START_PREFIX ;   (* default: "!!!GADGET_START:" *)
end_marker       = END_PREFIX ;     (* default: "!!!GADGET_END" *)
arg_prefix       = ARG_PREFIX ;     (* default: "!!!ARG:" *)

(* Header line: gadget name with optional ID and dependencies *)
header_line      = gadget_name , [ ":" , invocation_id , [ ":" , dependencies ] ] ;
gadget_name      = identifier ;
invocation_id    = identifier ;
dependencies     = identifier , { "," , identifier } ;

(* Parameters section: zero or more argument definitions *)
parameters       = { parameter } ;
parameter        = arg_prefix , pointer , newline , value ;

(* JSON Pointer path (without leading /) for nested structures *)
pointer          = segment , { "/" , segment } ;
segment          = identifier | array_index ;
array_index      = digit , { digit } ;

(* Value: all text until next arg_prefix or end_marker *)
value            = { any_character } ;

(* Basic tokens *)
identifier       = letter_or_underscore , { letter_or_underscore | digit } ;
letter_or_underscore = letter | "_" ;
letter           = "A" | "B" | ... | "Z" | "a" | "b" | ... | "z" ;
digit            = "0" | "1" | "2" | "3" | "4" | "5" | "6" | "7" | "8" | "9" ;
newline          = LF ;
any_character    = (* any Unicode character *) ;