DTRACDTRAC Project distinguishes runtime agent roles from supporting pipeline capabilities. The current classification is defined in agentTaxonomy.mjs; prompt-file mappings are defined in agentPromptRegistry.mjs. A filename alone does not establish the existence of an additional agent.
Runtime agents
The canonical list currently contains the following four entries. Registration and actual invocation by a particular task are separate facts to verify.
| ID | Prompt file |
|---|---|
research-partner | research-partner.md |
protocol-agent | protocol-agent.md |
ui-agent | ui-agent.md |
data-agent | data-collection-specialist.md |
The registry owns these mappings. For example, the file for data-agent retains the older specialist filename.
Pipelines and compatibility roles
The taxonomy classifies flow-design, feasibility-review, screen-generation, and build-validation as pipeline capabilities. They should not be counted as four additional runtime agents.
The orchestrator and older specialist names also remain in the compatibility classification. The prompt registry separately maintains persona files and legacy stage files. Do not remove or rename those files before checking compatibility with stored data and existing callers.
How prompts are loaded
agentDocs.ts reads the Markdown files selected by the registry to construct role-specific instructions. Specialist prompts are read again when called. The versioning function hashes the role, contract version, and instruction text to produce a traceable prompt version.
This guide covers role classification and loading. Model selection, execution order, failure recovery, and human review conditions require detailed guides grounded in each execution path. Start with the builder guide for the relationship to the UI workspaces.
Architecture reference
Refer to DTRAC architecture — Assurance tab alongside the builder and agent structure. This link opens the external architecture reference with the Assurance tab selected.
The reference site currently requires ChatGPT sign-in.