Skip to content

feat: add stetkeep protocol agents (brain-router, craft-specialist, perf-specialist)#232

Open
chanjoongx wants to merge 2 commits intoVoltAgent:mainfrom
chanjoongx:feat/add-stetkeep-protocol-agents
Open

feat: add stetkeep protocol agents (brain-router, craft-specialist, perf-specialist)#232
chanjoongx wants to merge 2 commits intoVoltAgent:mainfrom
chanjoongx:feat/add-stetkeep-protocol-agents

Conversation

@chanjoongx
Copy link
Copy Markdown

Summary

Adds three protocol-driven agents from stetkeep, an XML protocol framework + false-positive catalog for Claude Code.

Agents added

09 — Meta & Orchestration

  • brain-router: top-level orchestrator that classifies MD files and routes compound/ambiguous code-quality requests to the right specialist. Triggers on "refactor", "optimize", "full checkup", etc. Uses the Agent tool to actually spawn craft-specialist or perf-specialist — it does not just advise, it delegates.

04 — Quality & Security

  • craft-specialist: structural refactor agent restricted to Read/Grep/Glob/Edit/Bash — no Write tool, so it cannot create new files. Follows CRAFT.md's 20 anti-patterns (A1–A20) and an 8-entry false-positive catalog.
  • perf-specialist: measurement-first performance agent with a hard gate — refuses any optimization without baseline Lighthouse/Profiler numbers. Same tool restriction as craft-specialist. Includes 8-entry false-positive catalog for patterns V8/React already handle internally.

What makes these different from existing agents

  • vs code-reviewer: craft-specialist is narrower — behavior-preserving refactor only, with a false-positive catalog that explicitly lists patterns Claude must NOT flag (profiled hot paths, generated files, framework idioms, compliance constants).
  • vs performance-engineer: perf-specialist has a hard measurement gate. No numbers = no action. It will refuse and ask the user to run Lighthouse/Profiler first.
  • False-positive catalog: 16 total entries (8 CRAFT + 8 PERF) encode patterns that look like anti-patterns but aren't — this prevents "helpful" regressions that other agents don't guard against.

Note on agent file length

These agents are intentionally concise (~50 lines vs the typical 200+ line format here). The reason: they reference external XML protocol files (CRAFT.md / PERF.md / BRAIN.md) that ship with stetkeep and contain the full behavioral specification. The agent files act as controllers; the protocol files carry the detail. Happy to expand inline if the convention here requires it.

Note on model: inherit

These agents use model: inherit rather than sonnet/opus to respect whatever model the user's session is running. Functionally equivalent for sonnet users; easily overridden by editing the frontmatter.

Note on Agent tool in brain-router

brain-router uses the Agent tool because it actually spawns craft-specialist or perf-specialist as subagents — it does not just recommend delegation. This is standard Claude Code 2026 subagent spawning behavior.

Installation

Full framework (recommended): npx stetkeep install — copies protocol files (CRAFT.md, PERF.md, BRAIN.md) + these agents into the user's project.
Standalone: copy the agent files directly; add CRAFT.md / PERF.md to the project root from the stetkeep repo.

…erf-specialist)

- brain-router: XML-protocol orchestrator for code-quality routing (Meta & Orchestration)
- craft-specialist: structural refactor agent with 20 anti-patterns + 8-entry false-positive catalog (Quality & Security)
- perf-specialist: measurement-first performance agent with hard baseline gate + 8-entry false-positive catalog (Quality & Security)

All three reference stetkeep protocol files (BRAIN.md / CRAFT.md / PERF.md).
Install via: npx stetkeep install
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant