Go Multi-Agent Code Review — Deployment Guide¶

This directory contains ready-to-deploy agent definition files for the 7-Agent Go code review system described in bestpractice/Architecture.md §17–18.

Platform constraint: Claude Code does not allow subagents to spawn other subagents ("Subagents cannot spawn other subagents"). The go-review-lead orchestrator therefore runs as a Skill in the main conversation, not as an agent definition in .claude/agents/. The 7 Worker Agents in agents/ are dispatched by the main conversation, not by a Lead Agent.

What's in This Directory¶

agents/                            # 7 vertical Worker Agents only
├── go-security-reviewer.md        # Worker: SQL injection, XSS, secrets, auth flaws
├── go-concurrency-reviewer.md     # Worker: race conditions, goroutine leaks, deadlocks
├── go-performance-reviewer.md     # Worker: allocations, N+1, slice pre-allocation
├── go-error-reviewer.md           # Worker: error wrapping, resource lifecycle, panics
├── go-quality-reviewer.md         # Worker: naming, structure, modern Go idioms
├── go-test-reviewer.md            # Worker: table-driven tests, coverage, assertions
└── go-logic-reviewer.md           # Worker: boundary conditions, state machines, nil safety

Each file is a Claude Code agent definition (YAML frontmatter + system prompt). The agents are lightweight by design — they load their domain knowledge at runtime via the Skill tool, so the definition files stay short and maintainable.

The go-review-lead orchestrator is not in this directory. It lives in skills/go-review-lead/SKILL.md and is loaded by the main conversation.

Prerequisites¶

Each Worker Agent loads a corresponding vertical skill at runtime. These skills must be installed before the agents will work:

Agent file	Required skill
`go-concurrency-reviewer.md`	`go-concurrency-review` skill
`go-performance-reviewer.md`	`go-performance-review` skill
`go-error-reviewer.md`	`go-error-review` skill
`go-security-reviewer.md`	`go-security-review` skill
`go-quality-reviewer.md`	`go-quality-review` skill
`go-test-reviewer.md`	`go-test-review` skill
`go-logic-reviewer.md`	`go-logic-review` skill

The go-review-lead skill is loaded by the main conversation — it is the orchestrator, not a worker.

The source files for all 8 skills are in the skills/ directory of this repository (e.g. skills/go-concurrency-review/SKILL.md).

Installation¶

Step 1 — Install the skills¶

Copy the skill directories to your Claude Code user-level skills location. The default path is ~/.claude/skills/; adjust if your setup differs.

# Run from the repository root
for skill in go-review-lead \
             go-concurrency-review \
             go-performance-review \
             go-error-review \
             go-security-review \
             go-quality-review \
             go-test-review \
             go-logic-review; do
  cp -r "skills/$skill" ~/.claude/skills/
done

Step 2 — Install the Worker Agent definitions¶

Copy only the 7 Worker Agent definition files to ~/.claude/agents/. Do not copy go-review-lead.md — it does not belong in agents/.

mkdir -p ~/.claude/agents
# Copy all worker agents (go-review-lead is a Skill, not here)
for agent in go-security-reviewer \
             go-concurrency-reviewer \
             go-performance-reviewer \
             go-error-reviewer \
             go-quality-reviewer \
             go-test-reviewer \
             go-logic-reviewer; do
  cp "outputexample/go-review-lead/agents/${agent}.md" ~/.claude/agents/
done

Claude Code discovers agents in ~/.claude/agents/ automatically — no further configuration is needed.

Verify¶

List installed agents to confirm:

ls ~/.claude/agents/ | grep go-
# Expected output (7 workers, no go-review-lead):
# go-concurrency-reviewer.md
# go-error-reviewer.md
# go-logic-reviewer.md
# go-performance-reviewer.md
# go-quality-reviewer.md
# go-security-reviewer.md
# go-test-reviewer.md

Usage¶

The go-review-lead Skill runs in the main conversation. Invoke it by asking Claude to use the skill:

Use the go-review-lead skill to review this PR

or simply:

Review the Go changes in src/ using go-review-lead

The main conversation (running the go-review-lead Skill) will:

Run git diff -- '*.go' to identify changed Go files
Run a compile pre-check (go build)
Select review depth (Lite / Standard / Strict) based on file count and risk signals
Triage the diff through 4 phases to decide which Worker Agents to dispatch
Launch selected workers in parallel — each loads its vertical skill and runs independently
Consolidate all findings into a unified report with REV-NNN IDs, deduplication, and severity ordering

You can also invoke any Worker Agent directly for a focused review:

@go-concurrency-reviewer review internal/worker/pool.go

Model Configuration¶

All Worker Agents default to sonnet (see the model: field in each .md file). This is intentional: each Worker Agent focuses on a single dimension, so a mid-tier model with full attention on one domain outperforms a top-tier model splitting attention across all dimensions simultaneously. See Architecture.md §17.3.3 for the rationale.

To change the model for a specific agent, edit the model: field before copying:

# Example: use haiku for quality/logic workers to reduce cost further
sed -i '' 's/model: sonnet/model: haiku/' ~/.claude/agents/go-quality-reviewer.md
sed -i '' 's/model: sonnet/model: haiku/' ~/.claude/agents/go-logic-reviewer.md

Architecture Overview¶

For the complete design rationale — including the attention dilution problem, triage logic, Grep-Gated execution protocol, three-round validation results, and the model cost trade-off analysis — see:

English: bestpractice/Architecture.md §17–18
Chinese: bestpractice/架构篇.md §17–18