First AI Mutation

DUUMBI’s AI mutation system lets you describe changes in natural language. The AI generates structured graph patches — no text parsing, no syntax errors.

Prerequisites

• A DUUMBI project (see quickstart)
• An LLM provider configured (see installation)

Adding a function

From your project directory:

duumbi add "create a multiply function that takes two i64 parameters and returns their product"

The AI will:

1. Analyze the current graph
2. Generate a GraphPatch with an add_function operation
3. Validate the patched graph (schema + type check)
4. Show you a diff summary for confirmation

Changes:
  + Function: multiply(a: i64, b: i64) -> i64
    + Block: entry
      + Const: Load(a)
      + Const: Load(b)
      + Mul
      + Return

Apply changes? [y/N]

Type y to apply. The change is saved to main.jsonld and a snapshot is created in .duumbi/history/.

Modifying existing code

duumbi add "modify the main function to call multiply(6, 7) and print the result"

The AI generates modify_op and add_op patches to update the existing main function.

Undo

Made a mistake? Undo the last change:

duumbi undo

This restores from the latest snapshot in .duumbi/history/. Undo is a LIFO stack — you can undo multiple times.

Multi-provider support

DUUMBI supports multiple LLM providers with automatic fallback:

config.toml

[[providers]]
name = "anthropic"
api_key_env = "ANTHROPIC_API_KEY"
model = "claude-sonnet-4-20250514"

[[providers]]
name = "openai"
api_key_env = "OPENAI_API_KEY"
model = "gpt-4o"

If the primary provider fails (rate limit, network error), DUUMBI automatically tries the next provider in the chain.

How it works internally

1. The current graph (main.jsonld) is sent to the LLM as context
2. The LLM responds with tool calls defining graph patches (6 patch types)
3. Patches are applied atomically (all-or-nothing)
4. The result is validated against the schema
5. On failure: automatic retry with error feedback (up to 3 attempts)

The 6 patch operations:

Operation	Description
add_function	Append a complete function to the graph
add_block	Add a block to an existing function
add_op	Append an operation to a block
modify_op	Change a field on any node
remove_node	Remove a node (op, block, or function)
set_edge	Set an @id reference between nodes

Next steps

• Build a multi-module project with separate modules
• Use the intent system for complex, multi-step programs