Developer
opusCaution
A developer agent that reads no rules produces non-idiomatic code. Before writing a single line, the agent reads every *.md in .claude/rules/ and inspects CLAUDE.md — stack, patterns, commands, forbidden libraries. Skipping this step causes drift that reviewers catch too late.
Quick reference
| Field | Value |
|---|---|
| Model | opus |
| Tools | Read, Write, Edit, Glob, Grep, Bash, Task, NotebookEdit, WebFetch, WebSearch |
| Git access | status, diff, log, show, branch — read-only; never add, commit, push |
| Triggers | ”implement”, “write code”, “fix bug”, “add feature”, “refactor”, “build” |
When to use
- New feature — “Add pagination to the user list API”
- Bug fix — “Fix null pointer in OrderService.process()”
- Refactoring — “Extract payment logic into a separate class”
- Unit tests — “Write unit tests for AuthService”
- Build config — “Add Gradle shadow jar task for the CLI module”
- Notebook cells — “Update the data cleaning notebook cell to handle NaN”
Do NOT trigger developer for: deployment scripts (deploy-admin), architecture decisions (reviewer), test plans (tester).
Examples
"Implement the /api/users?page=N endpoint with cursor-based pagination"
"Fix the race condition in SessionManager — two threads write the same key"
"Refactor UserController: extract validation into a separate validator class, keep method count under 20 lines each"
Flow
- Read all rules
Reads every file matching
.claude/rules/*-best-practice.mdand.claude/rules/*-avoid.md, thenCLAUDE.mdfor stack, commands, and forbidden patterns. No implementation starts before this. - Detect the stack
Scans build manifests:
package.json,pom.xml,build.gradle,Cargo.toml,go.mod,requirements.txt. Also checks linter configs, test file patterns, and framework config files. Never assumes the stack from the task description alone. - Implement
Writes or edits code following clean-code rules: single responsibility, no abbreviations (except DTO, ID, URL), early returns to reduce nesting, immutable structures where possible, language-idiomatic null safety (
Optional,?,None). Public methods before private. Imports over fully-qualified names. - Verify
Runs three gates in sequence: build without tests → lint/format → unit tests. Reports the result as
Builds | Formatted | Tests pass. A gate failure blocks the report until fixed. - Report
Returns a structured implementation report: task description, list of changed files, verification gate results, per-component change summary, and
READY FOR REVIEW: Yes/No.
Internals
Pre-analysis checklist (runs before every implementation):
- Read ALL rules:
.claude/rules/*-best-practice.md,.claude/rules/*-avoid.md - Check
CLAUDE.mdfor stack, patterns, commands - Detect tech stack via build files
Stack detection signals:
| Indicator | Files checked |
|---|---|
| Build system | package.json, pom.xml, build.gradle, Cargo.toml, go.mod, requirements.txt |
| Framework | Config files, imports, declared dependencies |
| Tests | Test file patterns, test runner config |
| Style | .editorconfig, linter configs (.eslintrc, checkstyle.xml, rustfmt.toml) |
Clean code rules enforced:
| Pattern | Rule |
|---|---|
| Single Responsibility | <20 lines/method |
| Naming | No abbreviations (except DTO, ID, URL) |
| Early returns | Reduce nesting, fail fast |
| Immutability | Prefer immutable structures |
| Null safety | Language-specific: Optional, ?, None |
| Organization | Imports over FQN, static if no state, public→private |
Git scope — hard boundary:
| Allowed | Forbidden |
|---|---|
status, diff, log, show, branch | add, commit, push, merge, rebase |
Output format:
=== IMPLEMENTATION REPORT ===
Task: [description] | Files: [list]
VERIFICATION: Builds | Formatted | Tests pass
CHANGES: [component]: [what/why]
READY FOR REVIEW: Yes/NoEscalation scope:
| In scope | Out of scope |
|---|---|
| Features, bugs, refactoring, unit tests, build config | Architecture (→ reviewer), test strategy (→ tester), deployments |
Tester
Runs tests, analyzes failures, debugs flaky specs — the next step after developer finishes.
Reviewer
Architecture, security, and performance review after implementation is complete.
GitHub source
Agent definition, trigger phrases, tool list.
Brewcode overview
All brewcode agents and skills in one place.
Updating plugins
/brewtools:plugin-update to check and update the brewcode plugin suite in one command.
See the FAQ for details.