Agent Teams Lite Documentation

Instead of rewriting entire specification files, use Delta Specs to describe only what has changed. This allows for incremental updates that are eventually merged into the main specifications.

There are two primary types of deltas:

1. ADDED Requirements: Used to introduce new functionality. Use the ## ADDED Requirements heading followed by the new requirement and its associated scenarios.
2. MODIFIED Requirements: Used to update existing functionality. Use the ## MODIFIED Requirements heading, state the new requirement, and include a note about the previous state (e.g., (Previously: ...)).

Once a change is archived, these deltas are automatically merged into the main specs.

## ADDED Requirements

### Requirement: CSV Export
The system SHALL support exporting data to CSV format.

#### Scenario: Export all observations
- GIVEN the user has observations stored
- WHEN the user requests CSV export
- THEN a CSV file is generated with all observations
- AND column headers match the observation fields

## MODIFIED Requirements

### Requirement: Data Export
The system SHALL support multiple export formats.
(Previously: The system SHALL support JSON export.)

Agent Teams Lite uses a delegate-only Orchestrator to manage a sequence of specialized sub-agents. The Orchestrator does not perform real work; instead, it coordinates tasks, tracks state, and asks for user approval between phases.

Key Components

• Orchestrator: The coordinator. It launches sub-agents via a Task tool, shows summaries to the user, and manages the dependency graph. It uses minimal context (only state and summaries).
• Sub-agents: Specialized workers (e.g., EXPLORE, PROPOSE, SPEC, DESIGN, TASKS, APPLY, VERIFY, ARCHIVE) that receive pre-resolved compact rules from the Orchestrator to execute specific phases.
• Skill Registry: An auto-discovered catalog of coding skills and project conventions (defined in agents.md) that sub-agents use to follow project-specific patterns.
• Pluggable Store: A mechanism to persist artifacts using modes like engram, openspec, hybrid, or none.

The sdd-orchestrator is a COORDINATOR, not an executor. Its primary responsibility is to maintain a thin conversation thread, delegate all substantive work to specialized sub-agents, and synthesize their results.

Delegation Rules

To prevent context inflation, follow these guidelines for deciding when to act inline versus when to delegate:

Anti-patterns to avoid:

• Reading 4+ files to "understand" the codebase inline (delegate an exploration instead).
• Writing a feature across multiple files inline (delegate the task).
• Running tests or builds inline (delegate the execution).
• Reading files as preparation for edits, then editing (delegate the whole process together).

The Skill Resolver is a universal protocol used by any agent that delegates work to sub-agents (such as the ATL orchestrator, judgment-day, or pr-review). It ensures sub-agents receive relevant project standards and technical rules without needing to manually search for SKILL.md files.

The Workflow

1. Obtain Registry: The orchestrator retrieves a skill registry containing a Compact Rules section (pre-digested rules of 5-15 lines per skill).
2. Match Skills: Skills are matched based on Code Context (file patterns like .tsx or app/**) and Task Context (actions like "Create a PR" or "Run tests").
3. Inject Rules: The matching Compact Rules blocks are injected into the sub-agent's prompt under a ## Project Standards (auto-resolved) header before the task instructions.
4. Include Conventions: If available, project-specific file paths and notes from the Project Conventions section are also included.

Token Management

• Aim for 50-150 tokens per skill.
• If more than 5 skill blocks match, only include the 5 most relevant, prioritizing Code Context matches over Task Context matches.