Compose Team

Purpose / When to Activate

Activate after evaluate-story returns the tier, before the first sub-agent is spawned.

The Orchestrator must give each sub-agent exactly the context it needs — no more, no less. Context pollution wastes tokens and introduces drift. Context starvation causes failures.

This skill determines what goes into each sub-agent's context bundle.

Process

Step 0 — Resolve memory file paths (always first)

Read contexts/memory/index.md. For each bundle below that references a memory category, resolve the actual file path from the index before including it. Never hardcode memory file paths — the index is the source of truth.

Key categories to resolve:

• project category → project context file (stack, architecture, constraints)
• decisions category → decisions log
• patterns category → conventions and code patterns

If a category is absent from the index, omit it from the bundle silently — do not fail.

For Tier 1 (MicroDelivery)

MicroDelivery bundle (minimal):

- Story artefact
- patterns category file (resolved from index.md)
- Directly affected file(s) — identified from acceptance criteria
- orchestration.rules.md (relevant sections only)

For Tier 2 / Tier 3 (Core Team)

Planning Sub-Agent bundle:

- Story artefact
- orchestration.rules.md + artefacts.rules.md
- project category file (resolved from index.md)
- decisions category file (filtered: relevant decisions only — resolved from index.md)
- patterns category file (resolved from index.md)
- codebase-scan artefact (if exists)

If risk_analysis_required: true from evaluate-story, unconditionally include the decisions category from memory in the Planning Sub-Agent bundle, regardless of other relevance filtering.

Implementation Sub-Agent bundle:

- Story artefact
- {id}.execution-plan.md (from Planning Sub-Agent)
- patterns category file (resolved from index.md)
- project category file (resolved from index.md)
- Codebase files identified in execution plan (file list, not full content)

QA Sub-Agent bundle:

- Story artefact (acceptance criteria is the test spec)
- {id}.execution-plan.md (test checkpoints)
- {id}.impl-report.md (from Implementation Sub-Agent)
- orchestration.rules.md + artefacts.rules.md

On remediation pass (QA re-spawn):

QA bundle + {id}.qa-report.md (prior failure record)
Implementation bundle + {id}.qa-report.md (failure diagnosis)

For Tier 3 Specialists

For each specialist in specialists_triggered:

• Read specialist entry from agents/specialists.registry.yaml
• Add context_bundle files from registry entry to Implementation Sub-Agent bundle
• Record which specialists are activated (for impl-report)

Output

Returns (inline, to the Orchestrator) the file list for each sub-agent's context bundle. The Orchestrator uses this list when spawning each sub-agent.

Not written to file — this is the Orchestrator's coordination state, not a durable artefact.

Non-Goals

This skill must NOT:

• Enrich bundles beyond what is listed for each tier — more context is not better context
• Load full memory archives into any bundle
• Make strategic decisions about which specialists to include (the Orchestrator decides based on evaluate-story output)
• Invoke any other skill (only agents orchestrate)

Quality Checks

• Every sub-agent receives its required inputs (Story, rules, relevant memory)
• No sub-agent receives another sub-agent's full context (isolation is structural)
• Specialist bundles include only the registry-defined files, not the full memory set
• Remediation passes receive the prior failure artefact — no exceptions