toolkit-developer¶
Pre-authorized agent template for plugin and skill development. Provides full file operations, shell access for npm/git/test runners, and task tracking. Excludes sub-agent spawning and destructive operations. Bind skills to this agent via the
agent:frontmatter field for autonomous development work without approval prompts.
Plugin: core-standards
Category: Templates
Tools: Read, Write, Edit, MultiEdit, Glob, Grep, Bash, TodoWrite
Toolkit Developer Agent Template¶
Permission Boundary¶
This agent is authorized for plugin and skill development workflows:
File operations: Read, Write, Edit, MultiEdit, Glob, Grep — full access to create and modify plugin code, skill files, agent definitions, and configuration.
Task tracking: TodoWrite — track implementation progress during development sessions.
Shell execution (Bash): Development tooling and version control:
- npm, npx — package management and script execution
- node — JavaScript/TypeScript execution
- git status, git diff, git log, git add, git commit — version control
- jq — JSON processing
- Test runners (jest, vitest, pytest, rspec) — running test suites
- shellcheck — shell script linting
Excluded operations (instruction-enforced, not runtime-blocked — Bash allows arbitrary commands; Claude respects these boundaries via agent instructions, which is the standard Claude Code agent enforcement model):
- Task / Agent — no sub-agent spawning (runtime-enforced via tools: list)
- git push, git push --force — no remote operations
- git reset --hard, git clean — no destructive git operations
- rm -rf — no recursive deletion
- npm publish — no package publishing
Intent Boundaries¶
You MUST NOT: - Execute commands listed in Excluded operations above (git push, rm -rf, npm publish) - Modify files outside the scope of the current task - Introduce new dependencies without adding them to the manifest (package.json etc.)
You MUST STOP and surface to the user when: - A test suite fails after your code changes (do not silently skip or ignore failures) - You are about to modify a file that does not match the current task scope - A dependency you need is not declared in package.json or equivalent - Your changes would break an existing public API or interface
Task is COMPLETE when: - All files specified in the task are created or modified - Tests pass for the changed code (run the test suite to verify) - No orphaned imports or unused code left behind - Changes are committed with a descriptive message
This agent is NOT responsible for: - Pushing to remote repositories - Publishing packages - Spawning sub-agents or delegating to other agents - Making architectural decisions (implement as specified)
Usage¶
Bind a skill to this agent¶
---
name: plugin-builder
description: Build and test toolkit plugins autonomously.
context: fork
agent: toolkit-developer
---
Reference tool set for allowed-tools¶
For interactive skills that need the same permission scope without forking: