autopus-adk

# macOS / Linux
curl -sSfL https://raw.githubusercontent.com/Insajin/autopus-adk/main/install.sh | sh

# Windows (CMD or PowerShell)
powershell -c "irm https://raw.githubusercontent.com/Insajin/autopus-adk/main/install.ps1 | iex"

🎬 See It In Action

# Brainstorm with 3 AI models debating each other
/auto idea "Add OAuth2 with Google and GitHub providers" --multi --ultrathink

# One command does the rest — plan, build with 16 agents, ship with docs
/auto dev "Add OAuth2 with Google and GitHub providers"

Or if you prefer step-by-step control:

/auto plan "Add OAuth2 with Google and GitHub providers" --auto --multi --ultrathink
/auto go SPEC-AUTH-001 --auto --loop --team
/auto sync SPEC-AUTH-001

🐙 Pipeline ─────────────────────────────────────────────
  ✓ Phase 1:   Planning         planner decomposed 5 tasks
  ✓ Phase 1.5: Test Scaffold    12 failing tests created (RED)
  ✓ Phase 2:   Implementation   3 executors in parallel worktrees
  ✓ Phase 2.5: Annotation       @AX tags applied to 8 files
  ✓ Phase 3:   Testing          coverage: 62% → 91%
  ✓ Phase 4:   Review           TRUST 5: APPROVE | Security: PASS
  ───────────────────────────────────────────────────────
  ✅ 5/5 tasks │ 91% coverage │ 0 security issues │ 4m 32s

💡 One command. Production-ready code with tests, security audit, documentation, and decision history.

⭐ Star History

😤 The Problem

You're using AI coding tools. They're powerful. But...

• 🔄 Platform lock-in — Switch from Claude to Codex? Rewrite all your rules and prompts from scratch.
• 🎲 Hope-driven development — "Add auth" → AI writes code, skips tests, ignores security, forgets docs. Maybe it works.
• 🧠 Amnesia — Next session, the AI forgets every decision. "Why did we use this pattern?" → silence.
• 👤 Solo agent — One model, one context, one shot. Multi-file refactoring? Good luck.

🧠 The Philosophy: AX — Agent Experience

AX is not "AI Transformation." AX is Agent Experience — how AI agents perceive, navigate, and operate within your codebase. Just as UX designs for users and DX designs for developers, AX designs for agents.

flowchart LR
    UX["🧑 UX\nUser Experience"]
    DX["👩‍💻 DX\nDeveloper Experience"]
    AX["🤖 AX\nAgent Experience"]

    UX -->|"designs for"| U["Users"]
    DX -->|"designs for"| D["Developers"]
    AX -->|"designs for"| A["AI Agents"]

    style AX fill:#ff6b6b,stroke:#c92a2a,color:#fff

Most AI coding tools are designed around a simple model: you prompt, it responds.

Autopus starts from a different question: What if the agent is the primary audience of your project's documentation?

Think about onboarding a new engineer. You wouldn't hand them a blank editor and say "build the auth system." You'd give them: - An architecture overview so they understand the system - Coding conventions so their code fits in - Decision history so they don't repeat past mistakes - A review process so mistakes get caught before shipping

AI agents need the same things. The difference is that every session is their first day.

Autopus is a harness — a structured environment that gives agents the context, constraints, and workflows they need to produce code that a senior engineer would approve. Not through hope. Through design.

Of the agents. By the agents. For the agents.

flowchart TB
    subgraph OF ["🧬 Of the Agents"]
        direction TB
        O1["16 specialized agents\nform a software team"]
        O2["Planner · Executor · Tester\nReviewer · Architect · ..."]
    end

    subgraph BY ["⚡ By the Agents"]
        direction TB
        B1["Agents run the pipeline\nautonomously"]
        B2["Self-healing gates\nParallel worktrees\nMulti-model debate"]
    end

    subgraph FOR ["🎯 For the Agents"]
        direction TB
        F1["Every file, rule, and doc\nis designed for agents to parse"]
        F2["300-line limit · @AX tags\nStructured Lore · SPEC format"]
    end

    OF --> BY --> FOR

    style OF fill:#4c6ef5,stroke:#364fc7,color:#fff
    style BY fill:#7950f2,stroke:#5f3dc4,color:#fff
    style FOR fill:#f06595,stroke:#c2255c,color:#fff

🐙 Autopus doesn't make agents smarter. It makes them informed. That's AX.

🔥 What Makes Autopus Different

📏 Code That Agents Can Actually Read

Most codebases aren't written for AI. A 1,200-line file overwhelms context windows. Tangled responsibilities confuse intent. Autopus enforces a hard 300-line limit on every source file — not for aesthetics, but because agents work better when each file has one job and fits in one read.

❌ Traditional:
   service.go (1,200 lines) → Agent loses context halfway through

✅ Autopus:
   service.go       (180 lines)  Handler logic
   service_auth.go  (120 lines)  Auth middleware
   service_repo.go  (150 lines)  Data access
   → Every file fits in one context window. Every file has one job.

This isn't just about file size. The entire harness is agent-readable by design:

🐙 Human-readable is a bonus. Agent-readable is the requirement.

🤖 AI Agents That Form a Team, Not a Chatbot

Autopus doesn't give you one AI assistant — it gives you a software engineering team of 16 specialized agents with defined roles, quality gates, and retry logic.

🧠 Planner        →  Decomposes requirements into tasks
⚡ Executor ×N    →  Implements code in parallel worktrees
🧪 Tester         →  Writes tests BEFORE code (TDD enforced)
✅ Validator       →  Checks build, lint, vet
🔍 Reviewer       →  TRUST 5 code review
🛡️ Security       →  OWASP Top 10 audit
📝 Annotator      →  Documents code with @AX tags
🏗️ Architect      →  System design decisions
🔬 Deep Worker    →  Long-running autonomous exploration + implementation
... and 7 more

⚔️ AI Models That Debate Each Other (--multi)

One model has blind spots. Three models catch each other's mistakes.

Every AI model has its own strengths and biases — Claude is thorough but verbose, Codex is fast but sometimes shallow, Gemini brings a different perspective entirely. When you use --multi, they don't just work in parallel — they review, challenge, and build on each other's ideas.

# Add --multi to any command for multi-model intelligence
/auto idea "new feature" --multi          # 3 models brainstorm → cross-pollinate → ICE score
/auto plan "new feature" --multi          # 3 models review your SPEC independently
/auto go SPEC-ID --multi                  # 3 models debate your code review

flowchart TB
    C["🔍 Claude\nIndependent Analysis"] --> D["⚔️ Cross-Pollination\nEach model sees others' ideas"]
    X["🔍 Codex\nIndependent Analysis"] --> D
    G["🔍 Gemini\nIndependent Analysis"] --> D
    D --> R["🔄 Round 2\nAcknowledge · Integrate · Risk"]
    R --> J["🏛️ Blind Judge\nAnonymized scoring"]

Why this matters: - A bug that Claude misses, Codex catches. An edge case Codex ignores, Gemini flags. - Ideas that one model would never generate emerge from cross-pollination. - The blind judge scores anonymized results — no model favoritism. - Research shows multi-agent debate produces higher-quality outputs than any single model alone.

💡 /auto dev enables --multi by default. Every plan gets multi-model review. Every code review gets cross-checked. You don't have to think about it.

4 strategies: Consensus (merge agreements) · Debate (adversarial review + judge) · Pipeline (chain outputs) · Fastest (first wins)

🔁 Self-Healing Pipeline (RALF Loop)

Quality gates don't just fail — they fix themselves and retry.

flowchart LR
    R["🔴 RED\nRun Phase"] --> G["🟢 GREEN\nGate Check"]
    G -->|PASS| Done["✅ Next Phase"]
    G -->|FAIL| F["🔧 REFACTOR\nFix Issues"]
    F --> L["🔁 LOOP\nRetry"]
    L --> R
    L -.->|"3× no progress"| CB["⛔ Circuit Break"]

    style R fill:#ff6b6b,stroke:#c92a2a,color:#fff
    style G fill:#51cf66,stroke:#2b8a3e,color:#fff
    style F fill:#ffd43b,stroke:#f08c00,color:#000
    style L fill:#748ffc,stroke:#4263eb,color:#fff
    style CB fill:#868e96,stroke:#495057,color:#fff

/auto go SPEC-AUTH-001 --auto --loop

🐙 RALF [Gate 2] ──────────────────
  Iteration: 1/5 │ Issues: 3
  → spawning executor to fix golangci-lint warnings...

🐙 RALF [Gate 2] ──────────────────
  Iteration: 2/5 │ Issues: 3 → 0
  Status: PASS ✅

RALF = RED → GREEN → REFACTOR → LOOP — TDD principles applied to the pipeline itself. Built-in circuit breaker prevents infinite loops.

🌳 Parallel Agents in Isolated Worktrees

Multiple executors work simultaneously — each in its own git worktree. No conflicts. No corruption.

Phase 2: Implementation
  ├── ⚡ Executor 1 (worktree/T1) → pkg/auth/provider.go     ✓
  ├── ⚡ Executor 2 (worktree/T2) → pkg/auth/handler.go      ✓
  └── ⚡ Executor 3 (worktree/T3) → pkg/auth/middleware.go    ✓

Phase 2.1: Merge (task-ID order)
  ✓ T1 merged → T2 merged → T3 merged → working branch

File ownership prevents conflicts. GC suppression prevents corruption. Up to 5 concurrent worktrees.

📜 Lore: Your Codebase Never Forgets

Every commit captures the why, not just the what. Queryable forever.

feat(auth): add OAuth2 provider abstraction

Why: Need Google + GitHub support, extensible for future providers
Decision: Interface-based abstraction over direct SDK usage
Alternatives: Direct SDK calls (rejected: too coupled)
Ref: SPEC-AUTH-001

🐙 Autopus <[email protected]>

9 structured trailers. Query with auto lore query "why interface?". Stale decisions auto-detected after 90 days.

🧪 Autonomous Experiment Loop

Let AI iterate autonomously — measure, keep or discard, repeat.

/auto experiment --metric "go test -bench=BenchmarkProcess" --direction lower --max-iter 5

🐙 Experiment ───────────────────────
  Iter 1: baseline  │ 1200 ns/op
  Iter 2: optimize  │  850 ns/op  ✓ keep (29% improvement)
  Iter 3: refactor  │  900 ns/op  ✗ discard (regression)
  Iter 4: cache     │  620 ns/op  ✓ keep (27% improvement)
  ─────────────────────────────────────
  Result: 1200 → 620 ns/op (48% improvement)

Built-in circuit breaker prevents runaway iterations. Simplicity scoring penalizes over-complex solutions. Each iteration is a git commit — easy to review or revert.

⚠️ Status: Experimental — CLI commands (auto experiment) are available but skill-level integration is in progress. Core iteration loop works; full pipeline integration is coming.

🧠 Pipeline That Learns From Failures

Autopus pipelines don't just fail — they remember why and prevent the same mistake next time.

Gate 2 FAIL: golangci-lint — unused variable in pkg/auth/
→ Auto-recorded to .autopus/learnings/pipeline.jsonl
→ Next /auto go: learning injected into executor prompt
→ Same mistake never repeated

Every pipeline failure is captured as a structured learning entry. On the next run, relevant learnings are automatically injected into agent prompts — giving your pipeline institutional memory across sessions.

🏥 Post-Deploy Health Check

Deploy first, verify immediately. canary runs build verification, E2E tests, and browser health checks against your live deployment.

/auto canary                          # Build + E2E + browser auto-verification
/auto canary --url https://myapp.com  # Target a specific deployment URL
/auto canary --watch 5m               # Repeat every 5 minutes
/auto canary --compare                # Compare against previous canary report

Generates canary.md with full diagnostics — build status, test results, accessibility scores, and screenshot diffs.

🔀 Smart Model Routing

Not every task needs Opus. Autopus analyzes message complexity and routes to the right model automatically.

Simple query     → Haiku  (fast, cheap)
Code review      → Sonnet (balanced)
Architecture     → Opus   (deep reasoning)

No configuration needed — the router evaluates token count, code complexity, and domain signals to pick the optimal model. Override anytime with --quality ultra.

🔌 Provider Connection Wizard

Setting up AI providers shouldn't require reading docs. auto connect walks you through a 3-step guided setup.

auto connect         # Interactive wizard: server auth → workspace → OpenAI OAuth
auto connect status  # Deterministic local verify/readiness summary

The current release authenticates with the Autopus server, saves the selected workspace, and completes the OpenAI OAuth handoff. Use auto connect status or auto desktop status --json to verify the saved local state.

Desktop runtime ownership note: - The packaged autopus-desktop-runtime source/build/release provenance now lives in autopus-desktop/runtime-helper/. - ADK keeps auto connect, auto desktop ..., and auto worker ... as harness or compatibility surfaces, but normal desktop runtime shipping no longer depends on an autopus-adk checkout.

🤖 ADK Worker — Local Agent Execution

ADK Worker runs A2A + MCP hybrid tasks locally with browser login, JWT refresh, and direct platform connectivity. No separate bridge daemon or worker API key exchange is required for the default production path.

What it is for: - Connecting a local workspace to the Autopus platform worker loop - Receiving platform-dispatched tasks and executing them with local tools - Reusing the same security, budget, and audit rails as the main harness

What to do today: - If you're here for auto init, Codex @auto ..., or OpenCode /auto ..., you can ignore Worker for now - auto worker ... is an optional advanced surface that is still being rolled out and documented

💰 Iteration Budget Management

Workers don't run forever. Each executor gets a tool-call budget — preventing runaway agents while ensuring enough room to complete complex tasks.

📦 Context Compression

As pipelines progress through phases, earlier context gets compacted automatically into a fixed schema: Goal, Constraints, Progress, Decisions, Relevant Files, Next Steps, and Critical Context. Tool calls and results are pruned as pairs, unsafe provider payload bodies are omitted, and every applied compaction emits metadata with summary ids, source refs, reason codes, and budget/blocker state.

🔄 Pipeline That Never Dies

Crash mid-pipeline? Resume exactly where you left off.

/auto go SPEC-AUTH-001 --continue    # Resume from last checkpoint

YAML-based checkpoints save pipeline state after every phase. Stale detection prevents resuming outdated sessions. Combined with --auto --loop, you get a fully resilient autonomous pipeline.

🧪 E2E Scenarios from Your Code

Auto-generate and execute E2E test scenarios — no manual test writing needed.

auto test run                    # Run all scenarios
auto test run -s init --verbose  # Run a specific scenario

Autopus analyzes your codebase (Cobra commands, API routes, frontend pages) and generates typed scenarios with verification primitives (exit_code, stdout_contains, status_code, json_path, etc.). Incremental sync keeps scenarios up-to-date as code evolves.

🌐 Browser Automation — AI Agents That See and Click

AI agents can directly interact with web pages — open URLs, read accessibility trees, click elements, fill forms, and capture screenshots.

/auto browse --url https://example.com/settings

- @e1 heading "AI Settings"
- @e2 button "Provider Mode"
- @e3 switch "Auto Fallback" [checked]
- @e7 button "Save"

Terminal-aware: automatically selects cmux browser (in cmux) or agent-browser (fallback). Snapshot → Act → Verify loop — agents see the page as an accessibility tree and interact by reference.

📺 Live Agent Dashboard

In --team mode, each team member gets its own terminal pane with real-time log streaming.

┌─ lead ──────────┬─ builder-1 ───────┐
│ Phase 1: Plan   │ T1: auth.go       │
│ 5 tasks created │ implementing...   │
├─ tester ────────┼─ guardian ────────┤
│ scaffold: 12    │ waiting...        │
│ RED state ✓     │                   │
└─────────────────┴───────────────────┘

Works in cmux and tmux. Plain terminals degrade gracefully to log-only output.

📚 Auto-Documentation with Context7

Before implementation, Autopus fetches latest library docs automatically — so agents never work with stale API knowledge.

Phase 1.8: Doc Fetch
  → Detected: cobra v1.9, testify v1.11
  → Fetched: 2 libraries (6000 tokens)
  → Injected into executor + tester prompts

Context7 MCP → WebSearch fallback → skip (never blocks pipeline). Adaptive token budget: 1 lib → 5000 tokens, 5 libs → 2000 tokens each.

🔌 Hook-Based Result Collection

Instead of scraping terminal output, Autopus uses each provider's native hook system to collect structured JSON results.

Fallback: providers without hooks use ReadScreen + idle detection (SPEC-ORCH-006).

🔧 More Power Tools

🌐 One Config, Five Platforms

auto init   # auto-detects supported installed AI coding CLIs

One autopus.yaml generates native configuration for every detected supported platform.

Codex note: - Use $auto plan ..., $auto go ..., $auto idea ... immediately after auto init or auto update - Install the generated local plugin from the marketplace entry in .agents/plugins/marketplace.json (.autopus/plugins/auto) to unlock the friendlier @auto ... syntax - The local plugin provides the @auto ... router surface; detailed workflow instructions stay in repo skills and .codex/prompts/ so Codex does not see duplicate auto* skill entries - With skills.compiler.mode: split, long-tail Codex skills are emitted under .autopus/plugins/auto/skills/ while repo-visible helper skills stay under .codex/skills/ - .codex/hooks.json is still generated by default. The project-local .codex/config.toml uses [features].hooks; legacy [features].codex_hooks should not be emitted.

OpenCode note: - /auto ... and direct aliases like /auto-plan ... are generated under .opencode/commands/ - Native rule/agent/plugin files live under .opencode/, while reusable skills are published under .agents/skills/ - With skills.compiler.mode: split, shared/core skills stay under .agents/skills/ while OpenCode long-tail skills move to .opencode/skills/ - Helper workflows like /auto status, /auto map, /auto why, /auto verify, /auto secure, /auto test, /auto dev, and /auto doctor are generated as OpenCode-native command wrappers - opencode.json now registers the managed hook plugin automatically, so .opencode/plugins/autopus-hooks.js is live immediately after auto init or auto update

OMP role routing and context optimization (opt-in)

Operator commands:

auto platform omp models                 # installed model catalog
auto platform omp profile init --name omp-balanced --plan
auto platform omp profile apply omp-balanced
auto platform omp explain
auto status --platform omp

With the current upstream OMP 17.2.7 catalog, models returns the exact native selector, context window, reasoning flag, thinking levels, and input modes as a successful degraded/display-only result. OMP does not yet expose the family, Autopus capability, or authorization metadata required for safe automatic routing, so strict_routing_ready remains false and profile init/apply stays blocked with catalog_metadata_insufficient. Autopus does not guess those fields or inspect credentials.

OMP policies are provider-neutral and inactive until a named profile is selected. Start with the non-destructive overlay mode only after the strict catalog is ready, then replace the illustrative selectors below with exact, authenticated provider/model selectors.

role_model_policy:
  version: v1
  profile: omp-balanced
  profiles:
    omp-balanced:
      config_mode: overlay
      capabilities:
        deep_reasoning:
          required: true
          candidates:
            - selector: provider-a/reasoner
              family: family-a
              thinking: high
        coding_tool_use:
          required: true
          candidates:
            - selector: provider-a/coder
              family: family-a
              thinking: medium
        fast_validation:
          required: true
          candidates:
            - selector: provider-b/fast
              family: family-b
              thinking: medium
        vision_design:
          required: true
          candidates:
            - selector: provider-b/vision
              family: family-b
              thinking: high
        independent_dissent:
          required: true
          candidates:
            - selector: provider-b/reviewer
              family: family-b
              thinking: high
        deterministic_transform:
          required: true
          candidates:
            - selector: provider-a/transform
              family: family-a
              thinking: medium
      family_diversity:
        enabled: true
        roles: [advisor]

• Candidate order is the fallback order. selector, family, and thinking must match the probed catalog; the placeholders above are not model recommendations.
• safety is optional. Add explicit approval_mode or isolation_mode values only after the installed-version capability probe reports support; omission leaves those keys unclaimed.
• Use project-managed only when the project intentionally owns the target OMP keys. Every claimed key needs the observed prior_fingerprint and complete: true; an array claim such as retry.fallbackChains also needs full_array_ownership: true.

Context optimization is separately opt-in. The selected profile below keeps history in shadow and memory off; the active profile is defined but not selected.

omp_context_policy:
  profile: omp-safe-shadow
  profiles:
    omp-safe-shadow:
      history_mode: shadow
      memory_mode: off
      history_target_tokens: 1000
      fallback: canonical_full
      capability_policy: probe_required
      runtime_root_policy: isolated_task_owned
      mutation_scope: session_overlay
    omp-active-probed:
      history_mode: active
      memory_mode: off
      history_target_tokens: 1000
      fallback: canonical_full
      capability_policy: probe_required
      runtime_root_policy: isolated_task_owned
      mutation_scope: session_overlay

• Select omp-active-probed only when an exact capability probe succeeds, the runtime is task-owned (isolated_task_owned), and a fresh promotion attestation has been recomputed from at least 20 complete balanced AB/BA raw pairs and bound to the exact policy, session/binding, and canary digest. A missing, stale, mismatched, duplicate, or aggregate-only claim remains in shadow or is blocked. Use no_session only when no session runtime exists.
• The generated OMP bridge currently proves hash-only event forwarding, not a long-lived supervisor ACK, authoritative CanonicalSource, or real dispatch admission. It therefore cannot activate optimization by itself and remains fail-closed. canonical_full rollback may be claimed only after an authoritative new-session rebuild is verified and its delivery dispatch is acknowledged; absent that evidence, rollback remains unverified and blocked.
• Memory supports off or observation-only shadow. Shadow memory requires a memory_namespace; memory is never actively injected, and active receipts cannot claim memory injections or document omissions.
• Model resolution evidence is written to the gitignored runtime artifact .autopus/omp-model-resolution-v1.json. Task-scoped context evidence uses .autopus/runtime/omp-context/<task-id>/<session-id>/receipt.json.
• Run auto doctor (or auto doctor --json) to re-probe the installed CLI and configuration. Receipts must contain neither credentials, prompt bodies, nor absolute paths. Live provider transport checks can incur cost and run only with explicit auto doctor --provider-smoke opt-in; a stale receipt is not proof that a current canary passed.

Codex vs OpenCode

Split compiler note: - skills.compiler.mode: split is opt-in. Default full keeps the current backward-compatible surface layout. - In split mode, .agents/skills/ is reserved for shared/core skills, .opencode/skills/ carries OpenCode long-tail skills, and .autopus/plugins/auto/skills/ carries Codex plugin-scoped long-tail skills.

🚀 Quick Start Guide

Get from zero to your first AI-powered feature in under 5 minutes.

Step 1 · Install (one line)

Paste this command into your AI coding agent's chat (Claude Code, Codex, OpenCode, etc.) — the agent will run it for you. Or run it directly in your terminal.

# macOS / Linux — installs the binary and checks required tools
cd your-project    # go to your project folder (e.g., cd ~/my-app)
curl -sSfL https://raw.githubusercontent.com/Insajin/autopus-adk/main/install.sh | sh

# Windows (CMD or PowerShell)
cd your-project
powershell -c "irm https://raw.githubusercontent.com/Insajin/autopus-adk/main/install.ps1 | iex"

That's it. The installer installs the auto CLI plus an autopus alias, checks required tools, skips anything already present, and auto-installs missing essentials like git, GitHub CLI, and Antigravity CLI. It does not run auto init for you.

Platform command syntax: - Codex: install the generated local plugin, then use @auto ...; until then, use $auto ... - OpenCode: use /auto ... or /auto-<subcommand> ... - Claude Code / Antigravity CLI: use /auto ...

Note: If you run the Windows installer from Git Bash via powershell -c ..., restart Git Bash after install so it reloads the updated user PATH. The installer prints the exact install directory and a one-line export PATH=... fallback for that case.

# Homebrew (macOS) — canonical new install
brew install --cask Insajin/autopus/auto

# go install (requires Go 1.26+)
go install github.com/Insajin/autopus-adk/cmd/auto@latest

# Build from source
git clone https://github.com/Insajin/autopus-adk.git
cd autopus-adk && make build && make install

# After manual install, initialize:
cd your-project && auto init

brew uninstall --formula auto
brew install --cask Insajin/autopus/auto

After install, the script explains these commands:

• auto init: initialize the current project and generate autopus.yaml plus platform files
• auto update --self: update only the auto CLI binary
• auto update: refresh the current project's generated rules, skills, agents, and platform files

Step 2 · Initialize the Project

cd your-project
auto init

auto init scans your machine for supported installed AI coding CLIs (Claude Code, Codex, Antigravity CLI, OpenCode, Oh My Pi) and generates native configuration for each one — rules, skills, agents, commands, and platform-specific settings — all from a single autopus.yaml.

Claude Code statusline note: - If .claude/settings.json already has a statusLine.command, auto init / auto update now lets you choose keep, merge, or replace in interactive mode. - You can force the same behavior non-interactively with --statusline-mode keep|merge|replace.

✓ Detected: claude-code, codex, antigravity-cli, opencode, omp
✓ Generated: .claude/rules/, .claude/skills/, .claude/agents/, CLAUDE.md
✓ Generated: .codex/, AGENTS.md
✓ Generated: .gemini/, GEMINI.md
✓ Generated: .opencode/, .agents/skills/, AGENTS.md, opencode.json
✓ Generated: .omp/agents/, .omp/skills/, .omp/commands/, .omp/extensions/, .omp/config.yml
✓ Created: autopus.yaml

Step 3 · Set Up Project Context (/auto setup)

This is the most important step. AI agents lose all memory between sessions — every conversation is their first day on the job. /auto setup creates the "onboarding documents" that let agents understand your project instantly.

/auto setup     # Claude Code, Antigravity CLI, OpenCode, Oh My Pi
@auto setup     # Codex after local plugin install
$auto setup     # Codex fallback before plugin install

This analyzes your codebase and generates 5 context documents:

ARCHITECTURE.md                    # Domains, layers, dependency map
.autopus/project/product.md       # What this project does, core features
.autopus/project/structure.md     # Directory layout, package roles, entry points
.autopus/project/tech.md          # Tech stack, build system, testing strategy
.autopus/project/scenarios.md     # E2E test scenarios extracted from code

💡 Why this matters: Without these documents, an AI agent looking at your project is like a new hire with no onboarding — they'll guess at architecture, miss conventions, and reinvent patterns that already exist. With /auto setup, every agent session starts informed.

Optional DESIGN.md for UI Work

Frontend verification and review can use a project-local DESIGN.md as lightweight design context. auto init creates a starter DESIGN.md next to autopus.yaml without overwriting an existing one, and auto update backfills the starter plus the design: config block for older harness installs. Keep it short and include the source of truth, palette roles, typography hierarchy, component guardrails, layout/responsive rules, and agent guidance. If a project has no DESIGN.md or configured design baseline, /auto verify, Phase 3.5, /auto review, and auto orchestra review continue normally and report Design context: skipped (not configured) as a non-error condition.

Design context is only injected for UI-related diffs such as .tsx, .jsx, CSS-family files, theme/token files, or design-system paths. UI findings check palette-role drift, typography hierarchy drift, component guardrail violations, layout/responsive regressions, and source-of-truth mismatch. Review surfaces remain read-only; they report issues and delegate fixes instead of editing files directly.

Generated platform surfaces are not canonical. Update autopus-adk content/templates and run auto update to refresh .claude/*, .codex/*, .gemini/*, .opencode/*, .agents/skills/*, and plugin surfaces in a target project.

External design references are untrusted until explicitly promoted. auto design import stores sanitized artifacts under .autopus/design/imports/<import-id>/; it must not replace a human-maintained DESIGN.md by default. URL imports are public-HTTPS-only and SSRF-guarded: they reject local/private/metadata targets and unsafe redirects, cap redirects, timeout, and response size, and persist only redacted diagnostics when rejected.

Step 4 · Build Your First Feature

Now you're ready. Describe what you want in plain language:

# 1. Plan — AI creates a full SPEC (requirements, tasks, acceptance criteria)
/auto plan "Add a health check endpoint at GET /healthz"

# 2. Build — 16 agents handle implementation, testing, and review
/auto go SPEC-HEALTH-001 --auto

# 3. Ship — Sync docs, update SPEC status, commit with decision history
/auto sync SPEC-HEALTH-001

╭────────────────────────────────────╮
│ 🐙 Pipeline Complete!              │
│ SPEC-HEALTH-001: Health Check      │
│ Tasks: 3/3 │ Coverage: 92%         │
│ Review: APPROVE                    │
╰────────────────────────────────────╯

That's it — production-ready code with tests, security audit, and full documentation.

Quick Reference

Keeping Autopus Up to Date

Autopus has two separate update steps. When adopting a new release, run them in order from the project you want to refresh.

1. Binary update — update the auto CLI itself:

auto update --self

Downloads the latest release from GitHub and replaces only the CLI binary. The updater included in v0.50.73 and later first authenticates checksums.txt with the ECDSA P-256 publisher envelope checksums.txt.signatures, then verifies the archive's SHA256 checksum before extraction. Missing, malformed, untrusted, or expired signing data fails closed. This command does not refresh generated project files. Check your current version with auto version. On macOS, the updater included in v0.50.72 and later also preserves the downloaded release bytes and Developer ID signature.

Darwin and Linux stage the new binary on the target filesystem and commit it with an atomic exchange. If the kernel or filesystem does not support atomic exchange, the update fails before changing the installed binary; use the package manager or reinstall the release manually instead.

Windows preserves the installed binary beside the target as <binary>.old while placing the new binary. A forced process or power interruption between those two moves can leave the target path temporarily absent. Restore it from PowerShell, using the actual installation path, and then rerun the update:

Move-Item -LiteralPath "C:\path\to\auto.exe.old" -Destination "C:\path\to\auto.exe"
auto update --self

If both auto.exe and auto.exe.old exist, the updater removes .old automatically only when its own <binary>.old.autopus-complete marker proves that the prior installation finished. An unmarked or invalid recovery file is never deleted automatically: confirm which binary you want to keep, then restore or remove .old manually.

One-time macOS migration from v0.50.71 or earlier: Do not use the one-line shortcut below for this migration. Run these commands separately and in order:

bash auto update --self auto update --self --force auto update

The first command is still executed by the legacy updater. It verifies the SHA256 checksum without authenticating the publisher envelope and may replace the downloaded Developer ID signature with an ad hoc signature. After that first hop installs v0.50.72 or later, the second command is executed by the fixed updater now on disk and reinstalls the exact release bytes, restoring the Developer ID signature and TeamIdentifier=GP2PFA2PUV. The final auto update refreshes the current project's generated files. If the installed CLI is already v0.50.72 or later, future binary self-updates need only one auto update --self; run auto update afterward when you also need to refresh a project. A fresh Cask install, including a migration from the legacy Formula to the Cask, installs the signed release artifact directly and does not require the two self-update commands.

2. Harness update — apply the installed CLI's templates to the current project:

auto update

Regenerates rules, skills, agents, and platform-specific files such as .claude/*, .codex/*, .gemini/*, .opencode/*, and .agents/skills/* from the templates in the installed CLI. With skills.compiler.mode: split, the update preview/apply flow also manages .opencode/skills/* and .autopus/plugins/auto/skills/*, including stale artifact pruning. Your custom edits outside AUTOPUS:BEGIN~AUTOPUS:END markers are preserved. Newly installed platforms are auto-detected.

If Claude Code already has a user-managed statusLine.command, the update flow defaults to preserving it, can merge it with the managed Autopus statusline, or replace it entirely via --statusline-mode keep|merge|replace.

Both at once (when the installed CLI is v0.50.72 or later):

auto update --self && auto update

When to update: auto update --self installs the new binary. The release is reflected in the current project's generated surfaces only after the following auto update succeeds.

Common Scenarios

/auto fix "500 error on login page"

# Small feature — SPEC only, skip PRD
/auto plan "Add GET /healthz health check endpoint" --skip-prd

# Large feature — full PRD + SPEC
/auto plan "OAuth2 Google + GitHub provider support"

# Exploring an idea first — multi-provider brainstorm
/auto idea "Should we migrate to microservices?" --multi

/auto review                    # TRUST 5 review of current changes
/auto secure                    # OWASP Top 10 security scan
/auto review --multi            # Multi-model cross-review (debate strategy)

/auto Add 2FA to the login page

🐙 Triage ────────────────────────────
  Request: "Add 2FA to the login page"
  Complexity: HIGH → /auto idea --multi (recommended)

🤖 The Pipeline

7-Phase Multi-Agent Pipeline

Every /auto go runs this:

sequenceDiagram
    participant S as SPEC
    participant P as 🧠 Planner
    participant T as 🧪 Tester
    participant E as ⚡ Executor ×N
    participant A as 📝 Annotator
    participant V as ✅ Validator
    participant R as 🔍 Reviewer + 🛡️

    S->>P: Phase 1: Task decomposition + agent assignment
    P->>T: Phase 1.5: Scaffold failing tests (RED)

    rect rgb(230, 245, 255)
        Note over E: Phase 2: TDD in parallel worktrees
        T->>E: T1, T2, T3 ... (parallel)
    end

    E->>A: Phase 2.5: @AX tag management
    A->>V: Gate 2: Build + lint + vet
    V->>T: Phase 3: Coverage → 85%+
    T->>R: Phase 4: TRUST 5 + OWASP audit
    R-->>S: ✅ APPROVE

16 Specialized Agents

Quality Modes

/auto go SPEC-ID --quality ultra      # Premium path for every role; Codex effort varies by role
/auto go SPEC-ID --quality balanced   # Adaptive: Opus/Sonnet/Haiku by task complexity

auto quality ultra --apply            # Persist Ultra and refresh this project's managed agents
auto quality balanced --apply         # Persist Balanced and refresh this project's managed agents
auto quality provider claude ultra --apply   # Claude only (claude-code is also accepted)
auto quality provider codex balanced --apply # Codex only
auto quality provider claude inherit --apply # Remove the Claude override
auto quality supervisor inherit --apply  # Use the user's Codex model for the primary session
auto quality show                     # Show the persisted mode and supervisor policy

Projects using both Claude Code and Codex can select their modes independently:

quality:
  default: balanced
  providers:
    claude: ultra
    codex: balanced

quality.providers accepts the canonical keys claude and codex. Custom preset names must be 1–64 ASCII characters, start with a letter or digit, and then contain only letters, digits, hyphens, or underscores. quality.default remains the fallback for providers without an override, so existing configuration files keep their behavior. A per-run --quality flag is an explicit global override and temporarily wins over both persisted provider values without rewriting YAML. Provider-specific --apply refreshes only that configured platform; the existing global auto quality <mode> --apply still refreshes every configured platform.

New projects default to supervisor_model_policy: inherit, so Autopus does not override the user's Codex model for the primary session. Quality mode still controls managed agents and quality-managed orchestra providers. Existing projects without this policy keep the legacy quality interpretation, while ambiguous markerless root assignments are preserved during migration. Run auto quality supervisor inherit --apply to explicitly remove a known generated root profile, or auto quality supervisor quality --apply to opt an unchanged Autopus-managed primary config into the Sol profile for the selected quality mode. User-owned project model or effort assignments remain preserved and take precedence. Start a new Codex session after applying changes so managed agent definitions are reloaded.

GPT/Codex Ultra support in the CLI and Ultra activation in a project are separate. Installing a new binary does not enable Ultra. Run auto update to refresh the project's generated files, then opt in with auto quality ultra --apply and start a new Codex session. Any Ultra compact rollout or promotion remains separate and is not activated by these update commands.

In Codex Ultra, a quality-managed supervisor and orchestra use Sol+ultra; planner, architect, and security-auditor use Sol+max; every other managed agent uses Sol+xhigh. The Opus labels below describe model-tiered platform behavior; Codex uses this role-selective effort profile.

flowchart LR
    subgraph Ultra ["🔥 Ultra — All Opus"]
        U1["Planner\nOpus"] --> U2["Executor\nOpus"] --> U3["Validator\nOpus"]
    end

    subgraph Balanced ["⚖️ Balanced — Adaptive"]
        B1["Planner\nOpus"] --> B2["Executor\nby complexity"]
        B2 -->|HIGH| BH["Opus"]
        B2 -->|MEDIUM| BM["Sonnet"]
        B2 -->|LOW| BL["Sonnet"]
    end

    style Ultra fill:#fff3bf,stroke:#f08c00
    style Balanced fill:#d0ebff,stroke:#1971c2

* HIGH complexity → Opus · MEDIUM/LOW → Sonnet

Execution Modes

📐 The Workflow

⚡ The Fast Path — Two Commands

For most features, you only need two commands:

# 1. Brainstorm — multi-provider debate + deep analysis
/auto idea "Add webhook delivery with retry" --multi --ultrathink

# 2. Build & Ship — full autonomous pipeline
/auto dev "Add webhook delivery with retry"

/auto idea runs multi-provider brainstorming (Claude × Codex × Gemini debate) with deep sequential thinking, scores ideas with ICE, and saves the result.

/auto dev does the rest — plan → go → sync in one shot with all the power flags on by default:

💡 Don't want the full power? Use --solo for single-session mode, --no-multi to skip multi-provider review, or call plan / go / sync individually for fine-grained control.

📋 The Manual Path — Three Commands

For more control, run each stage separately:

flowchart LR
    PLAN["📋 plan\nDescribe"] -->|SPEC created| GO["🚀 go\nBuild"]
    GO -->|Code + Tests| SYNC["📦 sync\nShip"]

📋 Step 1 · /auto plan — Describe What You Want

Turn a plain-English description into a full SPEC — requirements, tasks, acceptance criteria, and risk analysis.

/auto plan "Add webhook delivery with retry and dead letter queue"

The spec-writer agent produces 5 documents:

.autopus/specs/SPEC-HOOK-001/
├── prd.md          # Product Requirements Document
├── spec.md         # EARS-format requirements
├── plan.md         # Task breakdown + agent assignments
├── acceptance.md   # Given-When-Then criteria
└── research.md     # Technical research + risks

Options: --multi for multi-provider review · --prd-mode minimal for lightweight PRDs · --skip-prd to go straight to SPEC

🚀 Step 2 · /auto go — Build It

Feed the SPEC to 16 agents that plan, scaffold tests, implement in parallel, validate, annotate, test, and review — all automatically.

/auto go SPEC-HOOK-001 --auto --loop

Phase 1    │ 🧠 Planner         │ SPEC → tasks + agent assignments
Phase 1.5  │ 🧪 Tester          │ Failing test skeletons (RED)
Phase 2    │ ⚡ Executor ×N      │ TDD in parallel worktrees
Phase 2.5  │ 📝 Annotator       │ @AX documentation tags
Gate  2    │ ✅ Validator        │ Build + lint + vet
Phase 3    │ 🧪 Tester          │ Coverage → 85%+
Phase 4    │ 🔍 Reviewer + 🛡️    │ TRUST 5 + OWASP audit

Options: --team for Agent Teams · --solo for single-session TDD · --quality ultra for the premium execution path · --multi for multi-model review

📦 Step 3 · /auto sync — Ship and Document

Update SPEC status, regenerate project docs, manage @AX tag lifecycle, and commit with structured Lore history.

/auto sync SPEC-HOOK-001
auto sync verify --spec SPEC-HOOK-001 --strict

Before committing a multi-repo sync, auto sync verify creates a read-only Phase A/Phase B staging plan. It excludes generated/runtime, tracked-but-ignored, unclassified, and shell-unsafe paths; --spec limits the plan to workspace-relative files owned by exactly one SPEC host, and --strict fails on anything excluded or unrelated.

╭────────────────────────────────────╮
│ 🐙 Pipeline Complete!              │
│ SPEC-HOOK-001: Webhook Delivery    │
│ Tasks: 5/5 │ Coverage: 91%         │
│ Review: APPROVE                    │
╰────────────────────────────────────╯

That's it. Three commands: describe → build → ship. Every decision recorded. Every test enforced.

🎯 TRUST 5 Code Review

Every review scores across 5 dimensions:

📊 Multi-Model Orchestration

Providers: Claude · Codex · Gemini · OpenCode — with graceful degradation.

Interactive debate with real-time pane visualization (cmux/tmux). Hook-based result collection for structured JSON output. WebSearch fallback when Context7 docs are unavailable.

📖 All Commands

⚙️ Configuration

mode: full                    # full or lite
project_name: my-project
platforms:
  - claude-code

architecture:
  auto_generate: true
  enforce: true

lore:
  enabled: true
  required_trailers: [Why, Decision]
  stale_threshold_days: 90

spec:
  review_gate:
    enabled: true
    strategy: debate
    providers: [claude, gemini]
    judge: claude

methodology:
  mode: tdd
  enforce: true

orchestra:
  enabled: true
  default_strategy: consensus
  providers:
    claude:
      binary: claude
    codex:
      binary: codex
    gemini:
      binary: agy
    opencode:
      binary: opencode

🏗️ Architecture

autopus-adk/
├── cmd/auto/           # Entry point
├── internal/cli/       # 28 Cobra commands (110+ total with subcommands)
├── pkg/
│   ├── adapter/        # 4 platform adapters (Claude, Codex, Gemini, OpenCode)
│   ├── arch/           # Architecture analysis + rule enforcement
│   ├── browse/         # Browser automation backend (cmux/agent-browser routing)
│   ├── config/         # Configuration schema + YAML loading
│   ├── constraint/     # Anti-pattern scanning
│   ├── content/        # Agent/skill/hook/profile generation + skill activator
│   ├── cost/           # Token-based cost estimator
│   ├── detect/         # Platform/framework/permission detection
│   ├── e2e/            # E2E scenario generation, execution, verification
│   ├── experiment/     # Autonomous experiment loop (metric, circuit breaker)
│   ├── issue/          # Auto issue reporter (context collection, sanitization)
│   ├── lore/           # Decision tracking (9-trailer protocol)
│   ├── lsp/            # LSP integration
│   ├── orchestra/      # Multi-model orchestration (4 strategies + brainstorm + interactive debate + hooks)
│   ├── pipeline/       # Pipeline state persistence + checkpoint + team monitor
│   ├── search/         # Knowledge search (Context7/Exa) + hash-based search
│   ├── selfupdate/     # CLI binary self-update (publisher signature, SHA256, transactional replace)
│   ├── setup/          # Project doc generation + validation
│   ├── sigmap/         # AST-based API signature extraction (Go + TypeScript)
│   ├── spec/           # EARS requirement parsing/validation
│   ├── telemetry/      # Pipeline telemetry (JSONL event recording)
│   ├── template/       # Go template rendering
│   ├── terminal/       # Terminal multiplexer adapters (cmux, tmux, plain)
│   └── version/        # Build metadata
├── templates/          # Platform-specific templates
├── content/            # Embedded content (16 agents, 40 skills)
└── configs/            # Default configuration

🔒 Security

🛡️ Supply Chain Attack Protection

"A popular Python package with tens of millions of monthly downloads was injected with malicious code. A simple pip install could steal SSH keys, AWS credentials, and DB passwords — not from the package you installed, but from somewhere deep in its dependency tree." — Andrej Karpathy

AI coding environments make this worse: agents auto-install packages, expand dependency trees, and execute code — all without human review. Autopus builds defense into the pipeline itself.

How Autopus Protects Your Development Workflow

For Non-Go Projects

The same principles apply when Autopus manages Python, Node.js, or other ecosystems:

# autopus.yaml — configure per-ecosystem security scans
security:
  scanners:
    go: "govulncheck ./..."
    python: "pip-audit && safety check"
    node: "npm audit --audit-level=high"

Best practices enforced by the harness: - Version pinning — Lock all dependencies to exact versions (go.sum, package-lock.json, requirements.txt) - Minimal dependencies — The 300-line file limit and single-responsibility rule naturally reduce unnecessary imports - Isolation — Parallel executors run in isolated git worktrees; no cross-contamination between tasks - No blind installs — Security Auditor agent flags unknown or unvetted packages before they enter the codebase

Binary Distribution Safety

Every binary release from v0.50.73 includes SHA256 checksums (checksums.txt) and an ECDSA P-256 publisher envelope (checksums.txt.signatures). The current POSIX installer, Windows installer, and the self-updater shipped in v0.50.73 or later authenticate the checksum manifest before downloading or extracting an archive, then verify the archive checksum. They do not fall back to checksum-only installation. The installers reject unsigned v0.50.72-or-earlier releases.

The POSIX path requires OpenSSL plus sha256sum or shasum; missing verification tools fail closed. The Windows path uses the platform CNG implementation and Get-FileHash.

Recommended: Inspect before you install

# 1. Download the script first — review it before running
curl -sSfL https://raw.githubusercontent.com/Insajin/autopus-adk/main/install.sh -o install.sh
less install.sh          # Read what it does
sh install.sh            # Run only after review

Or verify manually:

# Download binary + checksums separately
VERSION=$(curl -s https://api.github.com/repos/Insajin/autopus-adk/releases/latest | grep tag_name | sed 's/.*"v\(.*\)".*/\1/')
curl -LO "https://github.com/Insajin/autopus-adk/releases/download/v${VERSION}/autopus-adk_${VERSION}_$(uname -s | tr A-Z a-z)_$(uname -m | sed 's/x86_64/amd64/;s/aarch64/arm64/').tar.gz"
curl -LO "https://github.com/Insajin/autopus-adk/releases/download/v${VERSION}/checksums.txt"

# Verify SHA256 integrity only (this does not authenticate the publisher)
shasum -a 256 -c checksums.txt --ignore-missing

For an authenticated install, use the inspected current installer or the updater shipped in v0.50.73 or later; both require a trusted publisher signature before accepting checksums.txt.

Trust boundary: the one-line installers are served from the repository's main branch. Given trusted installer or updater bytes, publisher verification protects release assets delivered by GitHub and its CDN. It does not authenticate the installer bootstrap itself. A compromise of repository main or raw-main delivery can replace the verifier, pins, or download target without also compromising the release assets. An independently pinned installer origin is not provided yet.

What We Don't Do

• No telemetry or analytics collection
• No network calls except explicit commands (orchestra, search, update --self)
• No access to your AI provider API keys — Autopus orchestrates CLI tools, not API calls

🤝 Contributing

Autopus-ADK is open source under the MIT license. PRs welcome!

make test       # Run tests with race detection
make lint       # Run go vet
make coverage   # Generate coverage report