QCSD Development Swarm v1.0

Shift-left quality engineering swarm for in-sprint code quality assurance.

---

Overview

The Development Swarm takes refined stories (that passed Refinement) and validates code quality during sprint execution. Where the Ideation Swarm asks "Should we build this?" and the Refinement Swarm asks "How should we test this?", the Development Swarm asks "Is the code quality sufficient to ship?"

QCSD Phase Positioning

Phase	Swarm	Decision	When
Ideation	qcsd-ideation-swarm	GO / CONDITIONAL / NO-GO	PI/Sprint Planning
Refinement	qcsd-refinement-swarm	READY / CONDITIONAL / NOT-READY	Sprint Refinement
Development	qcsd-development-swarm	SHIP / CONDITIONAL / HOLD	During Sprint
Verification	qcsd-cicd-swarm	RELEASE / REMEDIATE / BLOCK	Pre-Release / CI-CD
Production	qcsd-production-swarm	HEALTHY / DEGRADED / CRITICAL	Post-Release

Parameters

• SOURCE_PATH: Source code directory to analyze (required, e.g., src/auth/)
• TEST_PATH: Test directory for coverage analysis (optional, e.g., tests/auth/)
• OUTPUT_FOLDER: Where to save reports (default: ${PROJECT_ROOT}/Agentic QCSD/development/)

---

ENFORCEMENT RULES - READ FIRST

Rule	Enforcement
E1	You MUST spawn ALL THREE core agents (qe-tdd-specialist, qe-code-complexity, qe-coverage-specialist) in Step 2. No exceptions.
E2	You MUST put all parallel Task calls in a SINGLE message.
E3	You MUST STOP and WAIT after each batch. No proceeding early.
E4	You MUST spawn conditional agents if flags are TRUE. No skipping.
E5	You MUST apply SHIP/CONDITIONAL/HOLD logic exactly as specified in Step 5.
E6	You MUST generate the full report structure. No abbreviated versions.
E7	Each agent MUST read its reference files before analysis.
E8	You MUST apply qe-defect-predictor analysis on ALL code changes in Step 8. Always.
E9	You MUST execute Step 7 learning persistence. No skipping.

PROHIBITED BEHAVIORS:

• Summarizing instead of spawning agents
• Skipping agents "for brevity"
• Proceeding before background tasks complete
• Providing your own analysis instead of spawning specialists
• Omitting report sections or using placeholder text

---

Step Execution Protocol

This skill uses a micro-file step architecture. Each step is a self-contained file loaded one at a time to avoid "lost in the middle" context degradation.

Execute steps sequentially by reading each step file with the Read tool.

Steps

1. Flag Detection -- steps/01-flag-detection.md -- Scan source code and tests, detect all 6 flags
2. Core Agents -- steps/02-core-agents.md -- Spawn qe-tdd-specialist, qe-code-complexity, qe-coverage-specialist in parallel
3. Batch 1 Results -- steps/03-batch1-results.md -- Wait for core agents, extract all metrics
4. Conditional Agents -- steps/04-conditional-agents.md -- Spawn flagged conditional agents in parallel
5. Decision Synthesis -- steps/05-decision-synthesis.md -- Apply SHIP/CONDITIONAL/HOLD logic
6. Report Generation -- steps/06-report-generation.md -- Generate executive summary and full report
7. Learning Persistence -- steps/07-learning-persistence.md -- Store findings to memory, save persistence record
8. Defect Predictor -- steps/08-defect-predictor.md -- Run qe-defect-predictor analysis on all code changes
9. Final Output -- steps/09-final-output.md -- Display completion summary with all scores

Execution Instructions

1. Use the Read tool to load the current step file (e.g., Read({ file_path: ".claude/skills/qcsd-development-swarm/steps/01-flag-detection.md" }))
2. Execute the step's instructions completely
3. Verify all success criteria are met before proceeding
4. Pass the step's output as context to the next step
5. If a step fails, halt and report the failure point -- do not skip ahead

Resume Support

To resume from a specific step: specify --from-step N and the orchestrator will skip to step N. Ensure you have the required prerequisite data from prior steps.

---

Agent Inventory

Agent	Type	Domain	Batch
qe-tdd-specialist	Core (always)	test-generation	1
qe-code-complexity	Core (always)	code-intelligence	1
qe-coverage-specialist	Core (always)	coverage-analysis	1
qe-security-scanner	Conditional (HAS_SECURITY_CODE)	security-compliance	2
qe-performance-tester	Conditional (HAS_PERFORMANCE_CODE)	chaos-resilience	2
qe-mutation-tester	Conditional (HAS_CRITICAL_CODE)	test-generation	2
qe-message-broker-tester	Conditional (HAS_MIDDLEWARE)	enterprise-integration	2
qe-sap-idoc-tester	Conditional (HAS_SAP_INTEGRATION)	enterprise-integration	2
qe-sod-analyzer	Conditional (HAS_AUTHORIZATION)	enterprise-integration	2
qe-defect-predictor	Analysis (always)	defect-intelligence	3

Total: 10 agents (3 core + 6 conditional + 1 analysis)

---

Quality Gate Thresholds

Metric	SHIP	CONDITIONAL	HOLD
TDD Adherence	>= 80%	60 - 79%	< 60%
Code Complexity	Avg <= 10	Avg 11-15	Avg > 15
Test Coverage	>= 80%	60 - 79%	< 60%
Mutation Score	>= 70%	50 - 69%	< 50%
Security Issues	No HIGH/CRITICAL	MEDIUM only	HIGH/CRITICAL found

---

Report Filename Mapping

Agent	Report Filename	Step
qe-tdd-specialist	02-tdd-analysis.md	2
qe-code-complexity	03-complexity-analysis.md	2
qe-coverage-specialist	04-coverage-analysis.md	2
qe-security-scanner	05-security-scan.md	4
qe-performance-tester	06-performance-analysis.md	4
qe-mutation-tester	07-mutation-testing.md	4
qe-message-broker-tester	08-middleware-health.md	4
qe-sap-idoc-tester	09-sap-integration.md	4
qe-sod-analyzer	10-sod-compliance.md	4
Learning Persistence	11-learning-persistence.json	7
qe-defect-predictor	12-defect-prediction.md	8
Synthesis	01-executive-summary.md	6

---

Execution Model Options

Model	When to Use	Agent Spawn
Workflow (PRIMARY, ADR-102)	Harness with the Workflow tool	Workflow({ name: "qcsd-development-review", args: { sourcePath, testPath } })
Task Tool (fallback)	Claude Code sessions without Workflow support	Task({ subagent_type, run_in_background: true })
MCP Tools	MCP server available	fleet_init({}) / task_submit({})
CLI	Terminal/scripts	swarm init / agent spawn

Workflow execution (ADR-102)

.claude/workflows/qcsd-development-review.js runs the review as a deterministic pipeline: one finder per quality dimension (TDD adherence, complexity, coverage gaps — args.dimensions selects a subset) → 3 blind adversarial refuters per finding (Loki-mode, ADR-074: refuters see only the bare claim + evidence, never the finder's confidence or each other; uncertainty defaults to refuted) → deterministic synthesis into finding-verdict@1 envelopes (ADR-103, schemas/finding-verdict.schema.json). A finding survives only if fewer than ⌈N/2⌉ refuters kill it. The final report contains ONLY confirmed findings; killed findings are retained under killed with their refutations for audit.

Args: sourcePath (required), testPath, dimensions (subset of tdd-adherence|complexity|coverage-gaps), maxFindings per dimension (default 5).

When the Workflow tool is unavailable, fall back to the Task-tool protocol below — the report format and gates are identical, minus the adversarial verification stage (note this in the report header as verification: none).

---

Key Principle

Code quality is measured by evidence, not intentions. This swarm provides in-sprint quality assessment to ensure code meets engineering standards before entering the CI/CD pipeline.