AgentSys

A modular runtime and orchestration system for AI agents.

Renamed from awesome-slash — The awesome- prefix implies a curated list of links, but this project is a functional software suite and runtime. Please update your installs: npm install -g agentsys

14 plugins · 43 agents · 30 skills (across all repos) · 26k lines of lib code · 3,357 tests · 3 platforms
Plugins distributed as standalone repos under agent-sh org — agentsys is the marketplace & installer

Commands · Installation · Website · Discussions

Built for Claude Code · Codex CLI · OpenCode

New skills, agents, and integrations ship constantly. Follow for real-time updates:

---

AI models can write code. That's not the hard part anymore. The hard part is everything around it — task selection, branch management, code review, artifact cleanup, CI, PR comments, deployment. AgentSys is the runtime that orchestrates agents to handle all of it — structured pipelines, gated phases, specialized agents, and persistent state that survives session boundaries.

---

Building custom skills, agents, hooks, or MCP tools? agnix is the CLI + LSP linter that catches config errors before they fail silently - real-time IDE validation, auto suggestions, auto-fix, and 155 rules for Cursor, Claude Code, Cline, Copilot, Codex, Windsurf, and more.

What This Is

An agent orchestration system — 14 plugins, 43 agents, and 30 skills that compose into structured pipelines for software development. Each plugin lives in its own standalone repo under the agent-sh org. agentsys is the marketplace and installer that ties them together.

Each agent has a single responsibility, a specific model assignment, and defined inputs/outputs. Pipelines enforce phase gates so agents can't skip steps. State persists across sessions so work survives interruptions.

The system runs on Claude Code, OpenCode, and Codex CLI. Install via the marketplace or the npm installer, and the plugins are fetched automatically from their repos.

---

The Approach

Code does code work. AI does AI work.

• Detection: regex, AST analysis, static analysis—fast, deterministic, no tokens wasted
• Judgment: LLM calls for synthesis, planning, review—where reasoning matters
• Result: 77% fewer tokens for /drift-detect vs multi-agent approaches, certainty-graded findings throughout

Certainty levels exist because not all findings are equal:

Level	Meaning	Action
HIGH	Definitely a problem	Safe to auto-fix
MEDIUM	Probably a problem	Needs context
LOW	Might be a problem	Needs human judgment

This came from testing on 1,000+ repositories.

---

Commands

Command	What it does

Each command works standalone. Together, they compose into end-to-end pipelines.

---

Skills

0 skills included across the plugins:

Category	Skills

Skills are the reusable implementation units. Agents invoke skills; commands orchestrate agents. When you install a plugin, its skills become available to all agents in that session.

---

Quick Navigation

Section	What's there
The Approach	Why it's built this way
Commands	All 13 commands overview
Skills	30 skills across plugins
Command Details	Deep dive into each command
How Commands Work Together	Standalone vs integrated
Design Philosophy	The thinking behind the architecture
Installation	Get started
Research & Testing	What went into building this
Documentation	Links to detailed docs

---

Command Details

/next-task

Purpose: Complete task-to-production automation.

What happens when you run it:

1. Policy Selection - Choose task source (GitHub Issues, GitHub Projects, GitLab, local file), priority filter, stopping point
2. Task Discovery - Shows top 5 prioritized tasks, you pick one
3. Worktree Setup - Creates isolated branch and working directory
4. Exploration - Deep codebase analysis to understand context
5. Planning - Designs implementation approach
6. User Approval - You review and approve the plan (last human interaction)
7. Implementation - Executes the plan
8. Pre-Review - Runs deslop-agent and test-coverage-checker
9. Review Loop - Multi-agent review iterates until clean
10. Delivery Validation - Verifies tests pass, build passes, requirements met
11. Docs Update - Updates CHANGELOG and related documentation
12. Ship - Creates PR, monitors CI, addresses comments, merges

Phase 9 uses the orchestrate-review skill to spawn parallel reviewers (code quality, security, performance, test coverage) plus conditional specialists.

Agents involved:

Agent	Model	Role
task-discoverer	sonnet	Finds and ranks tasks from your source
worktree-manager	haiku	Creates git worktrees and branches
exploration-agent	opus	Deep codebase analysis before planning
planning-agent	opus	Designs step-by-step implementation plan
implementation-agent	opus	Writes the actual code
test-coverage-checker	sonnet	Validates tests exist and are meaningful
delivery-validator	sonnet	Final checks before shipping
ci-monitor	haiku	Watches CI status
ci-fixer	sonnet	Fixes CI failures and review comments
simple-fixer	haiku	Executes mechanical edits

Cross-plugin agent:

Agent	Plugin	Role
deslop-agent	deslop	Removes AI artifacts before review
sync-docs-agent	sync-docs	Updates documentation

Usage:

/next-task              # Start new workflow
/next-task --resume     # Resume interrupted workflow
/next-task --status     # Check current state
/next-task --abort      # Cancel and cleanup

Full workflow documentation →

---

/agnix

Purpose: Lint agent configurations before they break your workflow. The first dedicated linter for AI agent configs.

agnix is a standalone open-source project that provides the validation engine. This plugin integrates it into your workflow.

The problem it solves:

Agent configurations are code. They affect behavior, security, and reliability. But unlike application code, they have no linting. You find out your SKILL.md is malformed when the agent fails. You discover your hooks have security issues when they're exploited. You realize your CLAUDE.md has conflicting rules when the AI behaves unexpectedly.

agnix catches these issues before they cause problems.

What it validates:

Category	What It Checks
Structure	Required fields, valid YAML/JSON, proper frontmatter
Security	Prompt injection vectors, overpermissive tools, exposed secrets
Consistency	Conflicting rules, duplicate definitions, broken references
Best Practices	Tool restrictions, model selection, trigger phrase quality
Cross-Platform	Compatibility across Claude Code, Cursor, Copilot, Codex, OpenCode, Gemini CLI, Cline, and more

155 validation rules (57 auto-fixable) derived from:

• Official tool specifications (Claude Code, Cursor, GitHub Copilot, Codex CLI, OpenCode, Gemini CLI, and more)
• Research papers on agent reliability and prompt injection
• Real-world testing across 500+ repositories
• Community-reported issues and edge cases

Supported files:

File Type	Examples
Skills	SKILL.md, */SKILL.md
Memory	CLAUDE.md, AGENTS.md, .github/CLAUDE.md
Hooks	.claude/settings.json, hooks configuration
MCP	*.mcp.json, MCP server configs
Cursor	.cursor/rules/*.mdc, .cursorrules
Copilot	.github/copilot-instructions.md

CI/CD Integration:

agnix outputs SARIF format for GitHub Code Scanning. Add it to your workflow:

- name: Lint agent configs
  run: agnix --format sarif > results.sarif
- uses: github/codeql-action/upload-sarif@v3
  with:
    sarif_file: results.sarif

Usage:

/agnix                       # Validate current project
/agnix --fix                 # Auto-fix fixable issues
/agnix --strict              # Treat warnings as errors
/agnix --target claude-code  # Only Claude Code rules
/agnix --format sarif        # Output for GitHub Code Scanning

Agent: agnix-agent (sonnet model)

External tool: Requires agnix CLI

npm install -g agnix         # Install via npm
# or
cargo install agnix-cli      # Install via Cargo
# or
brew install agnix           # Install via Homebrew (macOS)

Why use agnix:

• Catch config errors before they cause agent failures
• Enforce security best practices across your team
• Maintain consistency as your agent configs grow
• Integrate validation into CI/CD pipelines
• Support multiple AI tools from one linter

---

/ship

Purpose: Takes your current branch from "ready to commit" to "merged PR."

What happens when you run it:

1. Pre-flight - Detects CI platform, deployment platform, branch strategy
2. Commit - Stages and commits with generated message (if uncommitted changes)
3. Push & PR - Pushes branch, creates pull request
4. CI Monitor - Waits for CI, retries on transient failures
5. Review Wait - Waits 3 minutes for auto-reviewers (Copilot, Claude, Gemini, Codex)
6. Address Comments - Handles every comment from every reviewer
7. Merge - Merges when all comments resolved and CI passes
8. Deploy - Deploys and validates (if multi-branch workflow)
9. Cleanup - Removes worktree, closes issue, deletes branch

Platform Detection:

Type	Detected
CI	GitHub Actions, GitLab CI, CircleCI, Jenkins, Travis
Deploy	Railway, Vercel, Netlify, Fly.io, Render
Project	Node.js, Python, Rust, Go, Java

Review Comment Handling:

Every comment gets addressed. No exceptions. The workflow categorizes comments and handles each:

• Code fixes get implemented
• Style suggestions get applied
• Questions get answered
• False positives get explained

If something can't be fixed, the workflow replies explaining why and resolves the thread.

Usage:

/ship                       # Full workflow
/ship --dry-run             # Preview without executing
/ship --strategy rebase     # Use rebase instead of squash

Full workflow documentation →

---

/deslop

Purpose: Finds AI slop—debug statements, placeholder text, verbose comments, TODOs—and removes it.

How detection works:

Three phases run in sequence:

1. Phase 1: Regex Patterns (HIGH certainty)

   • console.log, print(), dbg!(), println!()
   • // TODO, // FIXME, // HACK
   • Empty catch blocks, disabled linters
   • Hardcoded secrets (API keys, tokens)

2. Phase 2: Multi-Pass Analyzers (MEDIUM certainty)

   • Doc-to-code ratio (excessive comments)
   • Verbosity ratio (AI preambles)
   • Over-engineering patterns
   • Buzzword inflation
   • Dead code detection
   • Stub functions

3. Phase 3: CLI Tools (LOW certainty, optional)

   • jscpd, madge, escomplex (JS/TS)
   • pylint, radon (Python)
   • golangci-lint (Go)
   • clippy (Rust)

Languages supported: JavaScript/TypeScript, Python, Rust, Go, Java

Usage:

/deslop              # Report only (safe)
/deslop apply        # Fix HIGH certainty issues
/deslop apply src/ 10  # Fix 10 issues in src/

Thoroughness levels:

• quick - Phase 1 only (fastest)
• normal - Phase 1 + Phase 2 (default)
• deep - All phases if tools available

Pattern reference →

---

/perf

Purpose: Structured performance investigation with baselines, profiling, and evidence-backed decisions.

10-phase methodology (based on recorded real performance investigation sessions):

1. Setup - Confirm scenario, success criteria, benchmark command
2. Baseline - 60s minimum runs, PERF_METRICS markers required
3. Breaking Point - Binary search to find failure threshold
4. Constraints - CPU/memory limits, measure delta vs baseline
5. Hypotheses - Generate up to 5 hypotheses with evidence and confidence
6. Code Paths - Use repo-map to identify entrypoints and hot files
7. Profiling - Language-specific tools (--cpu-prof, JFR, cProfile, pprof)
8. Optimization - One change per experiment, 2+ validation passes
9. Decision - Continue or stop based on measurable improvement
10. Consolidation - Final baseline, evidence log, investigation complete

Agents and skills:

Component	Role
perf-orchestrator	Coordinates all phases
perf-theory-gatherer	Generates hypotheses from git history and code
perf-theory-tester	Validates hypotheses with controlled experiments
perf-analyzer	Synthesizes findings into recommendations
perf-code-paths	Maps entrypoints and likely hot paths
perf-investigation-logger	Structured evidence logging

Usage:

/perf                 # Start new investigation
/perf --resume        # Resume previous investigation

Phase flags (advanced):

/perf --phase baseline --command "npm run bench" --version v1.2.0
/perf --phase breaking-point --param-min 1 --param-max 500
/perf --phase constraints --cpu 1 --memory 1GB
/perf --phase hypotheses --hypotheses-file perf-hypotheses.json
/perf --phase optimization --change "reduce allocations"
/perf --phase decision --verdict stop --rationale "no measurable improvement"

---

/drift-detect

Purpose: Compares your documentation and plans to what's actually in the code.

The problem it solves:

Your roadmap says "user authentication: done." But is it actually implemented? Your GitHub issue says "add dark mode." Is it already in the codebase? Plans drift from reality. This command finds the drift.

How it works:

1. JavaScript collectors gather data (fast, token-efficient)

   • GitHub issues and their labels
   • Documentation files
   • Actual code exports and implementations

2. Single Opus call performs semantic analysis

   • Matches concepts, not strings ("user auth" matches auth/, login.js, session.ts)
   • Identifies implemented but not documented
   • Identifies documented but not implemented
   • Finds stale issues that should be closed

Why this approach:

Multi-agent collection wastes tokens on coordination. JavaScript collectors are fast and deterministic. One well-prompted LLM call does the actual analysis. Result: 77% token reduction vs multi-agent approaches.

Tested on 1,000+ repositories before release.

Usage:

/drift-detect              # Full analysis
/drift-detect --depth quick  # Quick scan

---

/audit-project

Purpose: Multi-agent code review that iterates until issues are resolved.

What happens when you run it:

Up to 10 specialized role-based agents run based on your project:

Agent	When Active	Focus Area
code-quality-reviewer	Always	Code quality, error handling
security-expert	Always	Vulnerabilities, auth, secrets
performance-engineer	Always	N+1 queries, memory, blocking ops
test-quality-guardian	Always	Coverage, edge cases, mocking
architecture-reviewer	If 50+ files	Modularity, patterns, SOLID
database-specialist	If DB detected	Queries, indexes, transactions
api-designer	If API detected	REST, errors, pagination
frontend-specialist	If frontend detected	Components, state, UX
backend-specialist	If backend detected	Services, domain logic
devops-reviewer	If CI/CD detected	Pipelines, configs, secrets

Findings are collected and categorized by severity (critical/high/medium/low). All non-false-positive issues get fixed automatically. The loop repeats until no open issues remain.

Usage:

/audit-project                   # Full review
/audit-project --quick           # Single pass
/audit-project --resume          # Resume from queue file
/audit-project --domain security # Security focus only
/audit-project --recent          # Only recent changes

Agent reference →

---

/enhance

Purpose: Analyzes your prompts, plugins, agents, docs, hooks, and skills for improvement opportunities.

Seven analyzers run in parallel:

Analyzer	What it checks
plugin-enhancer	Plugin structure, MCP tool definitions, security patterns
agent-enhancer	Agent frontmatter, prompt quality
claudemd-enhancer	CLAUDE.md/AGENTS.md structure, token efficiency
docs-enhancer	Documentation readability, RAG optimization
prompt-enhancer	Prompt engineering patterns, clarity, examples
hooks-enhancer	Hook frontmatter, structure, safety
skills-enhancer	SKILL.md structure, trigger phrases

Each finding includes:

• Certainty level (HIGH/MEDIUM/LOW)
• Specific location (file:line)
• What's wrong
• How to fix it
• Whether it can be auto-fixed

Auto-learning: Detects obvious false positives (pattern docs, workflow gates) and saves them for future runs. Reduces noise over time without manual suppression files.

Usage:

/enhance                    # Run all analyzers
/enhance --focus=agent      # Just agent prompts
/enhance --apply            # Apply HIGH certainty fixes
/enhance --show-suppressed  # Show what's being filtered
/enhance --no-learn         # Analyze but don't save false positives

---

/repo-map

Purpose: Builds an AST-based map of symbols and imports for fast repo analysis.

What it generates:

• Cached file→symbols map (exports, functions, classes)
• Import graph for dependency hints

Output is cached at {state-dir}/repo-map.json and exposed via the MCP repo_map tool.

Why it matters:

Tools like /drift-detect and planners can use the map instead of re-scanning the repo every time.

Usage:

/repo-map init        # First-time map generation
/repo-map update      # Incremental update
/repo-map status      # Check freshness

Required: ast-grep (sg) must be installed.

---

/sync-docs

Purpose: Sync documentation with actual code changes—find outdated refs, update CHANGELOG, flag stale examples.

The problem it solves:

You refactor auth.js into auth/index.js. Your README still says import from './auth'. You rename a function. Three docs still reference the old name. You ship a feature. CHANGELOG doesn't mention it. Documentation drifts from code. This command finds the drift.

What it detects:

Category	Examples
Broken references	Imports to moved/renamed files, deleted exports
Version mismatches	Doc says v2.0, package.json says v2.1
Stale code examples	Import paths that no longer exist
Missing CHANGELOG	feat: and fix: commits without entries

Auto-fixable vs flagged:

Auto-fixable (apply mode)	Flagged for review
Version number updates	Removed exports referenced in docs
CHANGELOG entries for commits	Code examples needing context
	Function renames

Usage:

/sync-docs              # Check what docs need updates (safe)
/sync-docs apply        # Apply safe fixes
/sync-docs report src/  # Check docs related to src/
/sync-docs --all        # Full codebase scan

---

/learn

Purpose: Research any topic online and create a comprehensive learning guide with RAG-optimized indexes.

What it does:

1. Progressive Discovery - Uses funnel approach (broad → specific → deep) to find quality sources
2. Quality Scoring - Scores sources by authority, recency, depth, examples, uniqueness
3. Just-In-Time Extraction - Fetches only high-scoring sources to save tokens
4. Synthesis - Creates structured learning guide with examples and best practices
5. RAG Index - Updates CLAUDE.md/AGENTS.md master index for future lookups
6. Enhancement - Runs enhance:enhance-docs and enhance:enhance-prompts

Depth levels:

Depth	Sources	Use Case
brief	10	Quick overview
medium	20	Default, balanced
deep	40	Comprehensive

Output structure:

agent-knowledge/
  CLAUDE.md                    # Master index (updated each run)
  AGENTS.md                    # Index for OpenCode/Codex
  recursion.md                 # Topic-specific guide
  resources/
    recursion-sources.json     # Source metadata with quality scores

Usage:

/learn recursion                    # Default (20 sources)
/learn react hooks --depth=deep     # Comprehensive (40 sources)
/learn kubernetes --depth=brief     # Quick overview (10 sources)
/learn python async --no-enhance    # Skip enhancement pass

Agent: learn-agent (opus model for research quality)

---

/consult

Purpose: Get a second opinion from another AI CLI tool without leaving your current session.

What it does:

1. Tool Detection - Detects which AI CLI tools are installed (cross-platform)
2. Interactive Picker - If no tool specified, shows only installed tools to choose from
3. Effort Mapping - Maps effort levels to per-provider models and reasoning flags
4. Execution - Runs the consultation with safe-mode defaults and 120s timeout
5. Session Continuity - Saves session state for Claude and Gemini (supports --continue)

Supported tools:

Tool	Default Model (high)	Reasoning Control
Claude	claude-opus-4-6	max-turns
Gemini	gemini-3.1-pro-preview	built-in
Codex	gpt-5.3-codex	model_reasoning_effort
OpenCode	(user-selected or default)	--variant
Copilot	(default)	none

Usage:

/consult "Is this the right approach?" --tool=gemini --effort=high
/consult "Review for performance issues" --tool=codex
/consult "Suggest alternatives" --tool=claude --effort=max
/consult "Continue from where we left off" --continue
/consult "Explain this error" --context=diff --tool=gemini

Agent: consult-agent (sonnet model for orchestration)

---

/debate

Purpose: Stress-test ideas through structured multi-round debate between two AI CLI tools.

What it does:

1. Tool Detection - Detects which AI CLI tools are installed (cross-platform)
2. Interactive Picker - If no tools specified, prompts for proposer, challenger, effort, rounds, and context in a single batch question
3. Proposer/Challenger Format - First tool argues for the topic; second tool challenges with evidence
4. Multi-Round Exchange - Each round the proposer defends and the challenger responds (1–5 rounds)
5. Verdict - Orchestrator delivers a final synthesis picking a winner with reasoning

Usage:

# Natural language
/debate codex vs gemini about microservices vs monolith
/debate with claude and codex about our auth implementation
/debate thoroughly gemini vs codex about database schema design
/debate codex vs gemini 3 rounds about event sourcing

# Explicit flags
/debate "Should we use event sourcing?" --tools=claude,gemini --rounds=3 --effort=high
/debate "Valkey vs PostgreSQL for caching" --tools=codex,opencode

# With codebase context
/debate "Is our current approach correct?" --tools=gemini,codex --context=diff

Options:

Flag	Description
--tools=TOOL1,TOOL2	Proposer and challenger (comma-separated)
--rounds=N	Number of debate rounds, 1–5 (default: 2)
--effort=low|medium|high|max	Reasoning depth per tool call
--context=diff|file=PATH|none	Codebase context passed to both tools

Agent: debate-orchestrator (opus model for orchestration)

/web-ctl

Purpose: Browser automation for AI agents - navigate, authenticate, and interact with web pages.

How it works:

Each invocation is a single Node.js process using Playwright. No daemon, no MCP server. Session state persists via Chrome's userDataDir with AES-256-GCM encrypted storage.

Agent calls skill -> node scripts/web-ctl.js <args> -> Playwright API -> JSON result

Session lifecycle:

1. session start <name> - Create session (encrypted profile directory)
2. session auth <name> --url <login-url> - Opens headed Chrome for human login (2FA, CAPTCHAs). Polls for success URL/selector, encrypts cookies on completion
3. run <name> <action> - Headless actions using persisted cookies
4. session end <name> - Cleanup

Actions:

Action	Description	Key flag
goto <url>	Navigate to URL
snapshot	Get accessibility tree (primary page inspection)
click <sel>	Click element	--wait-stable
click-wait <sel>	Click and wait for DOM + network stability	--timeout <ms>
type <sel> <text>	Type with human-like delays
read <sel>	Read element text content
fill <sel> <value>	Clear field and set value
wait <sel>	Wait for element to appear	--timeout <ms>
evaluate <js>	Execute JS in page context	--allow-evaluate
screenshot	Full-page screenshot	--path <file>
network	Capture network requests	--filter <pattern>
checkpoint	Open headed browser for user (CAPTCHAs)	--timeout <sec>

click-wait waits for network idle + no DOM mutations for 500ms before returning. Cuts SPA interactions from multiple agent turns to one.

Error handling:

All errors return classified codes with actionable recovery suggestions:

Code	Recovery suggestion
element_not_found	Snapshot included in response for selector discovery
timeout	Increase --timeout
browser_closed	session start <name>
network_error	Check URL; verify cookies with session status
no_display	Use --vnc flag
session_expired	Re-authenticate

Security: Output sanitization (cookies/tokens redacted), prompt injection defense ([PAGE_CONTENT: ...] delimiters), AES-256-GCM encryption at rest, anti-bot measures (webdriver=false, random delays), read-only agent (no Write/Edit tools).

Selector syntax: role=button[name='Submit'], css=div.class, text=Click here, #id

Usage:

/web-ctl goto https://example.com
/web-ctl auth twitter --url https://x.com/i/flow/login
/web-ctl   # describe what you want to do, agent orchestrates it

Install:

agentsys install web-ctl
npm install playwright
npx playwright install chromium

Agent: web-session (sonnet model)

Skills: web-auth (human-in-the-loop auth), web-browse (headless actions)

---

How Commands Work Together

Standalone use:

/deslop apply          # Just clean up your code
/sync-docs             # Just check if docs need updates
/ship                  # Just ship this branch
/audit-project         # Just review the codebase

Integrated workflow:

When you run /next-task, it orchestrates everything:

/next-task picks task → explores codebase → plans implementation
    ↓
implementation-agent writes code
    ↓
deslop-agent cleans AI artifacts
    ↓
Phase 9 review loop iterates until approved
    ↓
delivery-validator checks requirements
    ↓
sync-docs-agent syncs documentation
    ↓
[/ship](#ship) creates PR → monitors CI → merges

The workflow tracks state so you can resume from any point.

---

Design Philosophy

Architecture decisions and trade-offs (click to expand)

The Actual Problem

Frontier models write good code. That's solved. What's not solved:

• Context management - Models forget what they're doing mid-session
• Compaction amnesia - Long sessions get summarized, losing critical state
• Task drift - Without structure, agents wander from the actual goal
• Skipped steps - Agents skip reviews, tests, or cleanup when not enforced
• Token waste - Using LLM calls for work that static analysis can do faster
• Babysitting - Manually orchestrating each phase of development
• Repetitive requests - Asking for the same workflow every single session

How This Addresses It

1. One agent, one job, done extremely well

Same principle as good code: single responsibility. The exploration-agent explores. The implementation-agent implements. Phase 9 spawns multiple focused reviewers. No agent tries to do everything. Specialized agents, each with narrow scope and clear success criteria.

2. Pipeline with gates, not a monolith

Same principle as DevOps. Each step must pass before the next begins. Can't push before review. Can't merge before CI passes. Hooks enforce this—agents literally cannot skip phases.

3. Tools do tool work, agents do agent work

If static analysis, regex, or a shell command can do it, don't ask an LLM. Pattern detection uses pre-indexed regex. File discovery uses glob. Platform detection uses file existence checks. The LLM only handles what requires judgment.

4. Agents don't need to know how tools work

The slop detector returns findings with certainty levels. The agent doesn't need to understand the three-phase pipeline, the regex patterns, or the analyzer heuristics. Good tool design means the consumer doesn't need implementation details.

5. Build tools where tools don't exist

Many tasks lack existing tools. JavaScript collectors for drift-detect. Multi-pass analyzers for slop detection. The result: agents receive structured data, not raw problems to figure out.

6. Research-backed prompt engineering

Documented techniques that measurably improve results:

• Progressive disclosure - Agents see only what's needed for the current step
• Structured output - JSON between delimiters, XML tags for sections
• Explicit constraints - What agents MUST NOT do matters as much as what they do
• Few-shot examples - Where patterns aren't obvious
• Tool calling over generation - Let the model use tools rather than generate tool-like output

7. Validate plan and results, not every step

Approve the plan. See the results. The middle is automated. One plan approval unlocks autonomous execution through implementation, review, cleanup, and shipping.

8. Right model for the task

Match model capability to task complexity:

• opus - Exploration, planning, implementation, review orchestration
• sonnet - Pattern matching, validation, discovery
• haiku - Git operations, file moves, CI polling

Quality compounds. Poor exploration → poor plan → poor implementation → review cycles. Early phases deserve the best model.

9. Persistent state survives sessions

Two JSON files track everything: what task, what phase. Sessions can die and resume. Multiple sessions run in parallel on different tasks using separate worktrees.

10. Delegate everything automatable

Agents don't just write code. They:

• Clean their own output (deslop-agent)
• Update documentation (sync-docs-agent)
• Fix CI failures (ci-fixer)
• Respond to review comments
• Check for plan drift (/drift-detect)
• Analyze their own prompts (/enhance)

If it can be specified, it can be delegated.

11. Orchestrator stays high-level

The main workflow orchestrator doesn't read files, search code, or write implementations. It launches specialized agents and receives their outputs. Keeps the orchestrator's context window available for coordination rather than filled with file contents.

12. Composable, not monolithic

Every command works standalone. /deslop cleans code without needing /next-task. /ship merges PRs without needing the full workflow. Pieces compose together, but each piece is useful on its own.

What This Gets You

• Run multiple sessions - Different tasks in different worktrees, no interference
• Fast iteration - Approve plan, check results, repeat
• Stay in the interesting parts - Policy decisions, architecture choices, edge cases
• Minimal review burden - Most issues caught and fixed before you see the output
• No repetitive requests - The workflow you want, without asking each time
• Scale horizontally - More sessions, more tasks, same oversight level

---

Installation

Claude Code (Recommended way)

/plugin marketplace add agent-sh/agentsys
/plugin install next-task@agentsys
/plugin install ship@agentsys

All Platforms (npm)

npm install -g agentsys && agentsys

Interactive installer for Claude Code, OpenCode, and Codex CLI.

# Non-interactive install
agentsys --tool claude              # Single tool
agentsys --tools "claude,opencode"  # Multiple tools
agentsys --development              # Dev mode (bypasses marketplace)

Full installation guide →

---

Requirements

Required:

• Git
• Node.js 18+

For GitHub workflows:

• GitHub CLI (gh) authenticated

For GitLab workflows:

• GitLab CLI (glab) authenticated

For /repo-map:

• ast-grep (sg) installed

For /agnix:

• agnix CLI installed (cargo install agnix-cli or brew install agnix)

Local diagnostics (optional):

npm run detect   # Platform detection (CI, deploy, project type)
npm run verify   # Tool availability + versions

---

Research & Testing

The system is built on research, not guesswork.

Knowledge base (agent-docs/): 8,000 lines of curated documentation from Anthropic, OpenAI, Google, and Microsoft covering:

• Agent architecture and design patterns
• Prompt engineering techniques
• Function calling and tool use
• Context efficiency and token optimization
• Multi-agent systems and orchestration
• Instruction following reliability

Testing:

• 1,818 tests passing
• Drift-detect validated on 1,000+ repositories
• E2E workflow testing across all commands
• Cross-platform validation (Claude Code, OpenCode, Codex CLI)

Methodology:

• /perf investigation phases based on recorded real performance investigation sessions
• Certainty levels derived from pattern analysis across repositories
• Token optimization measured and validated (77% reduction in drift-detect)

---

Documentation

Topic	Link
Installation	docs/INSTALLATION.md
Cross-Platform Setup	docs/CROSS_PLATFORM.md
Usage Examples	docs/USAGE.md
Architecture	docs/ARCHITECTURE.md

Workflow Deep-Dives

Workflow	Link
/next-task Flow	docs/workflows/NEXT-TASK.md
/ship Flow	docs/workflows/SHIP.md

Reference

Topic	Link
Slop Patterns	docs/reference/SLOP-PATTERNS.md
Agent Reference	docs/reference/AGENTS.md

---

Support

• Issues: github.com/agent-sh/agentsys/issues
• Discussions: github.com/agent-sh/agentsys/discussions

---

MIT License | Made by Avi Fenesh