SuperTeam: Multi-Agent Delivery Framework and Quality Gating Practice for Claude Code

This article analyzes how the SuperTeam framework addresses the core contradiction in AI-assisted development: "relying on model self-discipline to ensure quality is unreliable". Through a seven-stage delivery workflow and Python hook mechanism, it transforms software quality constraints from "agent self-discipline" to "physical enforcement", enabling reliable delivery of AI-driven development. The framework is suitable for Claude Code team development, and its core concepts also have reference value for traditional development processes.

SuperTeam V4.5.0 and earlier versions relied on quality rules in .md files and self-supervised execution by the orchestrator, but post-mortem diagnosis revealed a fatal flaw: the orchestrator could rationalize skipping review/verification steps (like a fox guarding the henhouse). V4.6.0 moves the execution mechanism to Python hooks (registered in ~/.claude/settings.json), which run outside the Claude inference chain, achieving physical mandatory constraints that cannot be rationalized and bypassed.

SuperTeam defines a seven-stage delivery process (G1-G7), with clear entry/exit criteria for each stage:

• G1: Requirements clarification and planning, producing plan.md with classification IDs (F/UI/API/MIG) to ensure traceable requirements;
• G2: Architecture and technical solution, pre-identifying technical debt and formulating mitigation strategies;
• G3: Implementation and coding, enforcing TDD red/green state machine via hooks (record failed tests before writing production code);
• G4-G7: Verification (verification.md must PASS), review (issues must be confirmed in finish.md), deployment (git operations blocked unless verification passes), retrospective (retrospective.md must include improvement actions). Moreover, G4-G7 are executed automatically in a chain without user confirmation.

The core innovation of V4.6.0 is hook-level mandatory constraints, including:

1. TDD red/green state machine: Tracks test status; production code can only be edited after resolving failed tests;
2. Plan MUST accounting mechanism: Each MUST item in plan.md requires corresponding evidence in execution.md to ensure no omissions;
3. Orchestrator decision log: A decision chapter must be recorded before generating executors/reviewers; unit IDs must exist and be incomplete;
4. Entry log reconciliation: Downstream agents must restate the MUST items in the plan verbatim; blocking occurs if there is a mismatch;
5. Commit gating: Git commits are blocked unless verification.md has a PASS verdict, finish.md confirms review issues, and retrospective.md includes improvement actions.

SuperTeam introduces plan-progress.json to record the status of MUST items (COMPLETE/PENDING/BLOCKED), supporting interruption recovery, progress visualization, and blocking identification. Additionally, there are two feature modules:

• Frontend Aesthetics Pipeline: Defines UI/UX aesthetic standard checklists (color consistency, typesetting norms, etc.);
• Team Behavior Audit: Generates reports through logs and retrospective documents to reveal systemic collaboration issues (e.g., requirement deviations).

The design philosophy of SuperTeam brings four insights: 1. From self-discipline to heteronomy, using technical mechanisms to enforce quality standards; 2. Physical constraints are better than process documents; 3. Traceability is the foundation of quality; 4. Automated closed loops improve consistency and efficiency. Conclusion: SuperTeam represents the evolution direction of AI-assisted development from "intelligent assistant" to "quality infrastructure", providing a practical framework for Claude Code team development, and its core concepts are also applicable to traditional development.