code-quality-evaluation

Code Quality Evaluation

Orchestrate 8 specialized teammate agents for comprehensive code quality evaluation. Each agent is a plugin subagent — spawn them as teammates and aggregate their findings.

Agents

All agents are read-only — they analyze code and report findings via SendMessage.

• the-crucible:namer — Naming: misleading, vague, type-encoded, scope-mismatched identifiers
• the-crucible:complexity-reviewer — Complexity: nesting, flag arguments, duplication, premature abstraction
• the-crucible:comment-reviewer — Comments: noise, staleness, refactoring signals, commented-out code
• the-crucible:test-reviewer — Tests: false confidence, implementation coupling, flakiness, coverage gaps
• the-crucible:error-handling-reviewer — Errors: silent swallowing, context loss, resource leaks, async error loss
• the-crucible:security-reviewer — Security: injection, access control, secrets, crypto, data exposure
• the-crucible:observability-reviewer — Observability: logging, metrics, tracing, cardinality, context propagation
• the-crucible:docs-auditor — Documentation: missing API docs, stale docs, contract gaps

Workflow

Phase 0: Create Team

Create a review team and 8 tasks — one per agent:

TeamCreate(name="code-review")

For each agent in the table above:
    TaskCreate(team_name="code-review", title="{agent_name}", description="Review {target}")

Phase 1: Spawn Reviewers

Spawn all 8 agents in parallel as teammates using their subagent_type. The agent files define each reviewer's expertise, patterns, and constraints — the prompt parameter only needs to specify the target and team context:

For each agent:
    Agent(
        subagent_type="{subagent_type}",
        prompt="Review all code in {target}. Send your findings to the leader via SendMessage.",
        team_name="code-review",
        task_id="{task_id}"
    )

Phase 2: Aggregate and Present Findings

As teammates send messages, aggregate findings:

1. Collect all messages from the 8 reviewers
2. Deduplicate overlapping findings across agents
3. Present to user organized by severity (Critical > Issues > Recommendations)
4. Ask user which issues to address

Phase 3: Apply Fixes

Address findings based on user direction:

1. Fix critical issues first, then issues, then recommendations
2. User may skip categories or individual findings
3. Continue until user says to stop or all addressed

Usage

Run code quality evaluation on src/auth/

Example interaction flow:

1. User: "Run code quality evaluation on src/auth/"
2. Create team "code-review" with 8 tasks
3. Spawn 8 teammate agents by subagent_type
4. Teammates send findings via messages
5. Aggregate and present: "Found 2 critical, 5 issues, 3 recommendations. Which to address?"
6. User picks categories or specific findings to fix
7. Apply fixes iteratively

Constraints

- Create the team and all tasks BEFORE spawning any agents - Use `subagent_type` to spawn agents — their `.md` files define review behavior - The `prompt` parameter provides only the target path and team instructions - All agents are READ-ONLY — they report, the leader fixes - Agents communicate via SendMessage — no file-based reports - Deduplicate findings that overlap across agents before presenting - Present findings organized by severity, not by agent