A self-governing agent OS for Claude Code — an orchestrator-only main agent, an evidence-gated /spec→/dev→/close→/commit→/push pipeline, a defense-in-depth git protection kernel, and an autonomous overnight loop.
This repository also configures its own agents. See what awesome-claude-harness tells them →
Agent Claude Code
Architecture review specialist for overnight exploration. Identifies structural issues, technical debt, optimization opportunities, dependency problems, and pattern inconsistencies. Returns structured JSON report.
Agent Claude Code
Business analyst subagent for requirements analysis and context building. Receives user requirement text, performs git analysis, identifies affected files, and returns either clarification questions or dual-format output (Markdown spec + JSON context).
Agent Claude Code
Agentic commit subagent. Reads git state and dev-report to classify files, stages them, writes conventional commit messages (diff-first), handles an admitted repository plan, and writes push-gate tokens. Dispatched exclusively by /commit.
Agent Claude Code
Cleanup execution specialist. Executes approved cleanup actions from cleanliness-inspector and style-inspector reports. Returns structured JSON execution report with results.
Agent Claude Code
File organization inspector for cleanup tasks. Detects misplaced docs, duplicates, temp files, build artifacts. Returns structured JSON report with cleanup recommendations.
Agent Claude Code
Implementation specialist for development tasks. Receives rich JSON context from orchestrator, creates parameterized scripts, implements changes based on git root cause analysis. Returns structured execution report.
Agent Claude Code
Git history analysis specialist. Discovers development edge cases by analyzing commits, violations, and patterns. Returns structured edge case report with prevention recommendations.
Agent Claude Code
Graphify enrichment subagent. Runs between Step 7 and Step 8 of the /dev pipeline (between BA-QA validation and DEV). Performs incremental Graphify cache update, extracts focused subgraph from BA blast-radius-map, patches context- .json with graphcontext field, writes per-task artifacts to .claude/dev-registry/…
Agent Claude Code
Pre-merge analyst subagent. Inspects branch divergence, diff stat, conflict markers, and overnight-state consistency; writes a nonce-keyed merge-analyst grant (60s expiry) to /tmp/agentic-commit/merge-analyst/. Dispatched exclusively by /merge.
Agent Claude Code
Test plan manager for overnight exploration with 3 invocation modes: PLAN (build test plan via browser exploration), TRIAGE (prioritize issues from specialist reports), RETRO (retrospective analysis and cross-cycle continuity). Uses Playwright to navigate the running app in PLAN mode before writing the test plan.
Agent Claude Code
Product-level analysis specialist for overnight exploration. Examines logical consistency, feature completeness, user flows, missing features, and business logic bugs. Returns structured JSON report.
Agent Claude Code
Prompt optimization inspector. Detects verbose non-functional content in command/agent documentation following 'rules not stories' principle. Returns structured JSON report with verbosity violations.
Agent Claude Code
Post-pull advisory analyst subagent. Reads the new-commits range after a successful git pull --rebase and produces a structured semantic risk summary. Writes no grant and blocks nothing. Dispatched exclusively by /pull when HEAD actually changed.
Agent Claude Code
Pre-push analyst subagent. Inspects the commits-to-push range for sensitive files, divergence, and branch protection violations; writes a nonce-keyed push-analyst grant to /tmp/agentic-commit/push-analyst/. Dispatched exclusively by /push.
Agent Claude Code
Quality assurance specialist for verification tasks. Receives implementation report from dev subagent, validates against success criteria, runs verification scripts, identifies issues. Returns structured verification report with pass/fail status.
Agent Claude Code
Folder rule discovery agent. Analyzes Git history to discover file creation patterns, extracts folder organization rules, generates INDEX.md and README.md documentation. Returns structured JSON with discovered rules.
Agent Claude Code
Three-phase spec subagent. Phase 0 = read spec, decide which agents need views (free judgment). Phase 1 = content-block extraction from full monolith (verbatim byte-slices, no section pre-filtering). Phase 2 = Gawande-style checkpoint generation. Invoked by /spec command with monolith path.
Agent Claude Code
Development standards auditor. Enforces /dev quality standards: no hardcoding, naming conventions, venv usage, step numbering, language, script merging, documentation conciseness. Returns structured JSON report with violations.
Agent Claude Code
Execution specialist for test infrastructure. Executes script-based and AI instruction-based tests. Returns structured execution report with results and recommendations.
Agent Claude Code
Validation specialist for test infrastructure. Validates test syntax, dependencies, and quality before execution. Returns structured validation report.
Agent Claude Code
Generate pytest skeleton tests from BA-produced acceptance-criteria- .json with pytest.fail("TESTINCOMPLETE:...") hard-stops; manage tests/generated/manifest.json with UPDATE vs CREATE logic keyed on acuid hashes. Triggered by /dev when complexitytier >= STANDARD or any tier with risklevel = high (per…
Agent Claude Code
UI/UX review specialist for overnight exploration. Evaluates visual design quality, aesthetic beauty, design system adherence, styling consistency, responsive design, and component quality. Returns structured JSON report with beauty score and design quality assessment. Accessibility checks are advisory.
Agent Claude Code
End-user simulation specialist for overnight exploration. Tests actual usage scenarios, checks if things work as expected, identifies UX friction, broken flows, and confusing behavior. Returns structured JSON report.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: