Agents
Agents (brewers) are specialized roles defined in markdown files with frontmatter metadata. Each agent has its own model, toolset, and system prompt. Brewcode includes 13 agents, from developer to task coordinator.
Summary table
| # | Agent | Model | Write tools | Purpose |
|---|---|---|---|---|
| 1 | developer | opus | Read, Write, Edit, Bash | Feature implementation, bug fixes |
| 2 | tester | sonnet | Read, Write, Edit, Bash | Test execution, failure analysis |
| 3 | reviewer | opus | Read (read-only) | Code review, architecture, security |
| 4 | architect | opus | Read (read-only) | Architecture analysis, patterns |
| 5 | skill-creator | opus | Read, Write, Edit | Skill creation and improvement |
| 6 | agent-creator | opus | Read, Write, Edit | Agent creation and improvement |
| 7 | hook-creator | opus | Read, Write, Edit, Bash | Hook creation and debugging |
| 8 | text-optimizer | sonnet | Read, Write, Edit | Text optimization for LLMs |
| 9 | bash-expert | opus | Read, Write, Edit, Bash | Bash/sh script creation |
| 10 | bc-coordinator | haiku | Read, Write, Edit, Bash | Phase and knowledge coordination |
| 11 | bc-knowledge-manager | haiku | Read, Write | KNOWLEDGE.jsonl compaction |
| 12 | bc-grepai-configurator | opus | Read, Write, Edit, Bash | grepai config.yaml generation |
| 13 | bc-rules-organizer | sonnet | Read, Write, Edit, Bash | .claude/rules/ organization |
developer
opusFull-stack developer: implements features, writes code, fixes bugs.
| Parameter | Value |
|---|---|
| Tools | Read, Write, Edit, Glob, Grep, Bash, Task, NotebookEdit, WebFetch, WebSearch |
| Write access | Full |
| Triggers | Implement features, fix bugs, write code |
Before starting work:
- Reads rules:
.claude/rules/*-best-practice.md,.claude/rules/*-avoid.md - Checks
CLAUDE.mdfor stack, patterns, commands - Detects stack from build files
Verification: Build (no tests) -> Lint/Format -> Unit tests -> report.
tester
sonnetSDET/QA — runs tests, analyzes results, debugs flaky tests.
| Parameter | Value |
|---|---|
| Tools | Read, Write, Edit, Glob, Grep, Bash, Task |
| Responsibilities | Test execution, failure analysis, flaky debugging, run configuration |
| Out of scope | Fixing production code (-> developer), major test rewrites |
Supports: Jest, pytest, JUnit 5, Go testing, and other frameworks. Automatically detects the test stack from configuration files.
architect
opusSystem architect — analysis, patterns, trade-offs, scaling.
| Parameter | Value |
|---|---|
| Tools | Read, Glob, Grep, Bash, WebFetch, WebSearch |
| Access | Read-only (Write and Edit are prohibited) |
| Triggers | Review architecture, design service, performance limits, scaling strategy |
Expertise: Layered, Clean/Hexagonal, Microservices, Event-Driven, CQRS.
Searches for existing solutions via grepai_search before analysis.
reviewer
opusArchitect-reviewer — architecture, quality, security, performance.
| Parameter | Value |
|---|---|
| Tools | Read, Glob, Grep, Bash, Task |
| Access | Read-only (Write and Edit prohibited via disallowedTools) |
| Delegation | Code -> developer, Tests -> tester |
Principle: Reuse over reinvention — apply existing patterns, reject duplicated functionality. New abstractions require justification.
Review areas:
- Architecture: Layered, Clean/Hexagonal, Microservices, Event-Driven, CQRS
- Quality: Correctness, Performance, Security, Maintainability, Testability, Scalability
Always searches for analogues via grepai_search before review.
skill-creator
opusCreates and improves Claude Code skills following official Anthropic best practices.
| Parameter | Value |
|---|---|
| Tools | Read, Write, Edit, Glob, Grep, Task, Skill, AskUserQuestion |
| Triggers | ”Create a skill”, “improve skill”, invocation issues |
Creation process: 6-Step Creation Process with a 1500-2000 word budget, Reference Splitting Strategy, description optimization for 84% activation rate.
agent-creator
opusCreates and improves Claude Code agents following Anthropic best practices.
| Parameter | Value |
|---|---|
| Tools | Read, Write, Edit, Glob, Grep, Task, Skill, WebFetch, WebSearch, AskUserQuestion |
| Triggers | ”Create agent”, “improve agent”, agent triggering issues |
Process: Agent Architect Process — parallel codebase analysis, System Prompt Patterns, Reference-Aware Skills.
hook-creator
opusCreates and debugs Claude Code hooks (bash/JS lifecycle event handlers).
| Parameter | Value |
|---|---|
| Tools | Read, Write, Edit, Glob, Grep, Bash, WebFetch, WebSearch |
| Triggers | ”Create PreToolUse hook”, “hook doesn’t work”, hook debugging |
Expertise: 10 Hook Patterns, Advanced Techniques, Multi-Stage hooks. Works with the hooks.json JSON schema and message routing.
text-optimizer
sonnetOptimizes text and documentation for efficient LLM token consumption.
| Parameter | Value |
|---|---|
| Tools | Read, Write, Edit, Glob, Grep, WebFetch |
| Skills | text-optimize |
| Triggers | ”Optimize prompt”, “reduce tokens”, “compress text”, “too verbose” |
Applies 30+ optimization rules across 5 categories: Claude behavior, Token efficiency, Structure, Reference integrity, Perception.
bash-expert
opusCreates production-quality bash/sh scripts for macOS/Linux.
| Parameter | Value |
|---|---|
| Tools | Read, Write, Edit, Glob, Grep, Bash, WebFetch |
| Triggers | ”Create script”, “bash script”, “shell script”, “install script” |
Patterns: strict mode (set -euo pipefail), error handling, argument parsing,
output formatting. Expert in Claude Code plugin scripts.
bc-coordinator
haikuInternal task coordinator — a lightweight agent for phase management.
| Parameter | Value |
|---|---|
| Tools | Read, Write, Edit, Bash |
| Mode | acceptEdits |
Responsibilities:
- Task initialization (lock, status)
- Phase Status table updates in PLAN.md
- Reading agent reports from disk
- Knowledge extraction -> KNOWLEDGE.jsonl (3-10 entries per phase)
- Auto-compaction when approaching the limit
- FINAL.md generation upon completion
2-step protocol
After each agent, the manager first WRITES the report to artifacts/, then CALLS bc-coordinator for knowledge extraction. This order is mandatory.
bc-knowledge-manager
haikuKnowledge manager — maintains quality and size of KNOWLEDGE.jsonl.
| Parameter | Value |
|---|---|
| Tools | Read, Write |
| Mode | acceptEdits |
Operations:
| Action | When |
|---|---|
| Deduplication | Coordinator reports duplicates |
| Merging similar entries | Before handoff |
| Prioritization | When limit is exceeded |
| Trimming | When limit is exceeded |
| Validation | Any operation |
Priority: ❌ (avoid) > ✅ (best practice) > ℹ️ (info).
bc-grepai-configurator
opusgrepai configuration specialist through deep project analysis.
| Parameter | Value |
|---|---|
| Tools | Read, Write, Edit, Bash, WebFetch, Glob, Grep |
| Mode | acceptEdits |
Environment: Ollama (bge-m3:1024), GOB storage, Java/Kotlin/JS/TS.
Analyzes project structure and generates an optimal .grepai/config.yaml.
bc-rules-organizer
sonnetRules organizer — creates and optimizes .claude/rules/*.md.
| Parameter | Value |
|---|---|
| Tools | Read, Write, Edit, Glob, Grep, Bash, Skill |
| Skills | text-optimize |
| Mode | acceptEdits |
| Triggers | ”Organize rules”, “path-specific rules”, “extract rules” |
Capabilities:
- Path-specific rules via
paths:frontmatter - Rule extraction from CLAUDE.md, documentation, code
- Lazy documentation — references instead of inline content
- LLM optimization via text-optimize
Models and cost
| Model | Agents | When to use |
|---|---|---|
| opus | developer, reviewer, architect, skill-creator, agent-creator, hook-creator, bash-expert, bc-grepai-configurator | Complex tasks, deep analysis, code generation |
| sonnet | tester, text-optimizer, bc-rules-organizer | Balanced tasks, review, optimization |
| haiku | bc-coordinator, bc-knowledge-manager | Lightweight operations: statuses, knowledge extraction, compaction |
Automatic selection
Agents are selected automatically based on the Phase Registry in PLAN.md.
The /brewcode:start skill reads agent assignments from the phase table
and launches the corresponding agents via Task API.