Team-Based Subsystem Design

Orchestrate parallel Designer agents for complex designs with multiple subsystems. The interactive session (command) acts as team lead; decomposition and synthesis run as spawned sub-agents for context isolation.

This workflow is strictly opt-in. The command assesses complexity and suggests team mode when appropriate. The user always decides whether to proceed.

Assessment (Lightweight, In-Command)

Assessment happens inside the /forge:design command before spawning any agents. It is a quick scan, not a heavy operation.

Path A: design-approach.md Exists

Read {slice_path}/design-approach.md and scan for:

• Component count in "Affected Components" table
• Cross-boundary interactions in "Key Technical Decisions"
• External system integrations mentioned

If 3+ components with cross-boundary interactions: Suggest team mode with AskUserQuestion:

"The design approach identifies {N} components with cross-boundary interactions ({list}). Team-based design produces deeper edge-case coverage for complex multi-subsystem designs.

Note: Team mode uses significantly more tokens (~Nx for N subsystems plus synthesis overhead).

1. Team mode — Architect decomposition, then parallel designers per subsystem, then synthesis
2. Standard mode — Single designer handles the entire design"

Path B: No design-approach.md

Read {slice_path}/requirements.md and quickly scan for signals of multi-subsystem complexity:

• 3+ distinct components or services mentioned
• Multiple external system boundaries
• Complex data flows across component boundaries

If complexity signals are present: Suggest running design-approach first:

"The requirements mention {N} components with external integrations. Running the design approach first would help identify whether team-based design is appropriate.

1. Run design approach first (Recommended)
2. Proceed with standard full design
3. Proceed with team mode directly (higher token cost, less signal)"

If no complexity signals: Proceed to standard mode without mentioning team design.

Assessment Criteria

Team mode is likely beneficial when:

• Design approach identifies 3+ distinct components
• Components exchange data across boundaries or share state
• Multiple external system integrations exist
• Complex state machines span components
• Shared infrastructure requires upfront coordination

Team mode is NOT appropriate when:

• Fewer than 3 subsystems
• Single-component changes
• Simple changes following well-established patterns

Structural Parallel with Research Lenses

This follows the same "decompose, parallel explore, synthesize" pattern as Research Lenses (Skill: parallel-research), but applied differently:

• Lenses = different perspectives on the same scope (external viewpoints)
• Team design = same perspective on different scopes (internal decomposition)

They remain separate skills because the decomposition unit differs: lenses slice by concern, team design slices by subsystem.

Three Stages (After User Opts In)

Stage 1: Architect Decomposition

Who: The Architect Decomposer agent. Runs as a sub-agent to isolate code discovery and decomposition work from the interactive context.

Spawning the decomposer:

Use Task tool with:
- subagent_type: "Architect Decomposer"
- prompt: |
    ## Inputs
    - slice_path: {slice_path}
    - slice_name: {slice_name}

Stage 2: Parallel Subsystem Design (Team)

Who: N Designer agents, one per subsystem, spawned as teammates or parallel sub-agents.

Spawning each designer:

Use Task tool with:
- subagent_type: "Designer"
- prompt: |
    ## Inputs
    - mode: "full"
    - slice_path: {slice_path}
    - slice_name: {slice_name}
    - subsystem_name: {subsystem_name}
    - subsystem_number: {N}

    ## Subsystem Scope
    {paste the subsystem's section from
     architect-decomposition.md}

    ## Interface Contracts
    {paste the shared interface contracts section}

    ## Output Location
    Write output to:
    {slice_path}/artifacts/subsystem-{N}-{name}.md
    (NOT to design.md -- synthesizer creates that)

Execution modes:

• If agent teams available and user selects team mode: create a team with one designer per subsystem.
• Otherwise: spawn parallel sub-agents, one per subsystem.

Stage 3: Synthesis

Who: The Design Synthesizer agent. Runs as a sub-agent to isolate synthesis work (reading N subsystem designs) from the interactive context.

Spawning the synthesizer:

Use Task tool with:
- subagent_type: "Design Synthesizer"
- prompt: |
    ## Inputs
    - slice_path: {slice_path}
    - slice_name: {slice_name}
    - subsystem_count: {N}
    - subsystem_names: [{name1}, {name2}, ...]

Post-Synthesis Evolution

After synthesis produces design.md:

• Most changes: Edit design.md directly (PR discussions, reviewer feedback, refinements)
• Re-synthesis: Reserve only for fundamental architectural changes where second-order effects are hard to trace manually

Cost and Benefit

Dimension	Impact
Token cost	~Nx for N subsystems + synthesis
Wall clock	Parallel execution offsets overhead
Edge cases	Deeper — each designer focuses fully
Lifecycles	Complete — no cut corners
Interface quality	Pre-agreed contracts prevent drift
Conflict detection	Synthesis catches gaps

Artifact Summary

Artifact	Location	Producer
Architect decomposition	artifacts/architect-decomposition.md	Architect Decomposer
Subsystem designs	artifacts/subsystem-{N}-{name}.md	Designer agents
Synthesis notes	artifacts/synthesis-notes.md	Design Synthesizer
Unified design	design.md	Design Synthesizer