Work Item DAG

How complex tasks decompose into a dependency graph of concurrent work items.

Overview

For complex tasks that involve multiple steps, Codiv decomposes the work into a directed acyclic graph (DAG) of Work Items. This enables concurrent execution — independent steps run in parallel while respecting dependencies.

Work Item Data Model

Each Work Item tracks:

Field	Description
`id`	Unique identifier
`title`	Human-readable description of the task
`status`	Current state (see state machine below)
`dependencies`	List of Work Item IDs that must complete first
`artifacts`	Output files or data produced
`token_budget`	Maximum tokens this item may consume
`cost`	Actual token/cost usage

State Machine

Work Items follow a deterministic state machine:

stateDiagram-v2
    [*] --> Pending
    Pending --> Running : dependencies met
    Running --> Completed
    Running --> Failed
    note right of Pending : Additional states:\nBlocked, Cancelled

DAG Construction

When the agent receives a complex task, it:

Plans — breaks the task into discrete work items
Identifies dependencies — determines which items depend on others
Builds the DAG — constructs the graph with edges representing “must complete before”
Validates — checks for cycles (which would be invalid)

For example, “refactor the auth module and update all tests” might produce:

graph LR
    A[Read auth module] --> B[Plan refactoring]
    B --> C[Edit auth code]
    C --> D[Run tests]
    B --> E[Update test files]
    E --> D

Concurrent Execution

The scheduler uses Tokio to run independent Work Items in parallel:

Items with no pending dependencies start immediately
As items complete, newly unblocked items begin
The DAG structure guarantees correct ordering without manual coordination

Budget Enforcement

Each Work Item has a token budget. The scheduler:

Tracks token usage per item
Pauses items approaching their budget limit
Reports cost breakdowns per item and for the overall task

Artifact Storage

Work Items can produce artifacts — files created, test results, analysis outputs. These are stored alongside the Work Item and available to downstream items that depend on them.