🐙 Autopus-ADK

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

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

Start with one real task in an existing repository. After setup, run auto doctor in your terminal to check the installed platform wiring.

These examples use the Codex plugin router. Native Codex skills use $codex-auto-fix, $codex-auto-review, and the corresponding route names; Claude Code and OpenCode use /auto ....

Autopus is most useful when a team needs consistent delivery rules across coding tools and repositories. A single, short task may need only the coding tool's native workflow. Start with the default split skill catalog and expand it when needed. See the harness comparison and evaluation protocol.

To inspect overhead, start with auto skill audit --dir . --platform codex --json. Use explicit selection policies and replay cases before changing skill exposure, and auto telemetry harness to compare supplied native/current/reduced runs. The diagnostics guide includes runnable examples, unknown-value semantics, and read-only verification planning.

See multiagent coordination for current platform boundaries, dependency-aware runner behavior, and worker ownership validation. Native subagent availability alone does not establish a team lifecycle or workspace isolation in the active session.

auto doctor agents inventories installed runtimes without model calls. Explicit live collectors for Codex and OpenCode check native lifecycle evidence; auto telemetry team attributes supervisor/worker usage without adding parent rollups twice. See lifecycle checks and usage for execution boundaries, examples, and the recorded live results.

What is checked: CLI gates validate specific artifacts, configuration, and results when invoked and configured. Skill instructions guide planning, TDD, and review; they do not prove that an agent followed every step. A passing gate is evidence for its checked scope, not a guarantee that an application is ready for production.

# 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 with disjoint file ownership
  ✓ Phase 3:   Testing          coverage: 62% → 91%
  ✓ Phase 4:   Review           TRUST 5: APPROVE | Security: PASS
  ───────────────────────────────────────────────────────
  ✅ 5/5 tasks │ 91% coverage │ 0 security issues │ 4m 32s

The pipeline output above is illustrative, not a benchmark. Completion depends on the project’s acceptance criteria and executed checks.

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.

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.

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["Project-defined limits · opt-in @AX\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.

File size alone does not establish cohesion or correctness. Autopus reports large source files, but enforces a ceiling only when the project sets a positive architecture.max_file_lines. Absent or 0 is advisory; keep cohesive code together and split by real responsibility. Repository-specific CI limits remain authoritative.

❌ 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.

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 workflow instruction)
✅ 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

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-model read-only planning advisory → one SPEC writer → independent review
/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)

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.

The worktree execution path isolates executors in separate Git working directories. Codex native workers share one cwd/filesystem: assign disjoint files and serialize overlapping edits. The following example applies to the worktree path.

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

The worktree path supports up to 5 concurrent worktrees. File ownership and ordered integration reduce conflicts; merges still require verification. Native worker capacity depends on the platform and available runtime limits.

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 <noreply@autopus.co>

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

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.

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.

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.

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.

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 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

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

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.

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.

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.

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.

On pane-capable team runtimes, each team member can get a 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.

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.

auto init   # auto-detects supported installed AI coding CLIs

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

Antigravity discovers the workspace plugin directly. auto init and auto update do not import a project-rendered bundle globally, and no longer emit the unused .agents/commands/ mirror. Inspect the generated bundle with agy plugin validate .agents/plugins/autopus; this checks structure, not whether runtime permissions or hooks have executed. Native user permission settings are preserved.

The 2026-09-13 compatibility checks used Claude Code 2.1.263, Codex CLI 0.153.4, Antigravity CLI 1.1.26, OpenCode 1.18.7, and OMP 18.1.19. These are observed test versions, not a claim that every newer release or feature is verified. Codex 0.154.0 worktree support remains experimental; the harness does not silently enable experimental topology or replace user permission settings.

Codex note:

• Use $codex-auto-plan ..., $codex-auto-go ..., or another $codex-auto-<route> skill immediately after auto init or auto update
• Install the generated local plugin from .agents/plugins/marketplace.json (.autopus/plugins/auto) to enable the friendlier @auto ... syntax
• The plugin provides the @auto ... router. Detailed workflows are unique native skills under .codex/skills/codex-<name>/SKILL.md; Autopus does not generate repository .codex/prompts/ or markdown .codex/rules/
• Multi-agent mode uses [features.multi_agent_v2] with the six current collaboration tools and a shared cwd/filesystem; it does not use legacy send_input, resume_agent, or close_agent
• The requested spawned-worker ceiling is codex.agents.max_concurrent_threads in autopus.yaml (default 4, range 1–64); the coordinator is not included. auto update writes the requested value to .codex/config.toml. Namespace selection must be backed by CLI compatibility evidence; an unverified version uses the documented [agents] max_concurrent_threads_per_session with the assumption stated. Explicit harness settings survive regeneration. Host and account limits still take precedence.
• auto doctor separates the requested count and the value inspected on disk from the loaded-session and effective limits. Reading a project or user config file does not prove that an active session loaded it. Unobservable session limits remain unknown, with the inspection limitation reported. Start a new session after changing configuration; Autopus never interrupts active agents or equates this setting with local build/test concurrency.
• .codex/hooks.json is generated by default, while structural TOML merging preserves unrelated user config
• features.multi_agent is still a live native switch. Explicit true and false values, including user comments, survive regeneration and cleanup; equality with a current default is not ownership evidence.

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/
• The default skills.compiler.mode: split publishes core skills and their required references. Opted-in long-tail bundles use .opencode/skills/; workflow-only rules remain readable without loading into every initial prompt.
• 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

Oh My Pi note:

• Rules are written flat as .omp/rules/autopus-<name>.md. OMP scans each rule root non-recursively, so a nested rules/autopus/ directory would never reach the session; the autopus- filename prefix is the namespace instead
• Only the manifest-recorded autopus- files are ADK-owned. Your own rules in the same directory (.omp/rules/mine.md) are left untouched by auto update and auto platform remove omp
• auto init ignores .omp/rules/ as a directory pattern. A filename glob such as .omp/rules/autopus-*.md would silently remove every generated rule from OMP discovery, which is why the directory form is used. To track your own rule despite the ignore, run git add -f .omp/rules/mine.md (gitignore negation does not work inside an ignored directory) or keep it outside .omp/rules/
• OMP uses its priority-100 native .omp/skills/<name>/SKILL.md and .omp/commands/*.md roots. It does not register .agents/skills as a custom directory and does not create a base .omp/config.yml unless project-managed settings require one
• auto doctor verifies OMP 18.0.5 through version/help/config metadata and a provider-free RPC handshake. The handshake selects a bootstrap model locally but sends no prompt and performs no provider request

For everyday model setup, run:

auto quality

auto quality opens the agent model composition first (Per-agent model composition). Answer t there, or a target name such as claude-code, to configure one surface instead: the shared quality mode or one coding tool's agent models. Only tools whose generated agents carry a model are offered — Claude Code, Codex, and OMP; Antigravity CLI and OpenCode inherit the session model and are named as such instead of being silently omitted.

Claude Code / Codex: the wizard prints every agent with its relative tier and the concrete model that tier becomes on that tool, takes agent=tier edits (fable, opus, sonnet, haiku), previews the result, and on y stores the edits as a quality.presets.<name> entry bound through quality.providers.<tool>. Each tool keeps its own preset, so moving Claude Code's executor to fable leaves Codex where it was. quality.default never changes. A customized balanced or ultra preset is not overwritten: the editor refuses that name and cancels after three refusals. Run auto update (or auto quality --apply) to regenerate the agent files.

OMP: choose balanced/ultra → GPT/Claude, or custom to pick an installed model per bundled agent. The custom path still anchors a family for the agents you do not pin, lists the installed catalog with each model's thinking levels, and labels every prompt with the capability that agent's route requires. A model the catalog says cannot serve that capability is refused at the prompt with the models that can. Review the compact agent/model/thinking table and type y to apply. This line editor keeps OMP profiles in both modes; the composition step's OMP tab in the default autopus mode pins agents without a profile instead (Per-agent model composition).

Enter, n, or EOF at confirmation cancels without changes; --apply is not required. Existing agent overrides and multi-provider review settings are preserved. An explicitly defined custom profile keeps its own model families. Start a new OMP session after applying.

Advanced and automation commands remain available:

auto platform omp models                 # installed model catalog
auto platform omp profile apply balanced --family gpt --plan
auto platform omp profile apply balanced --family gpt
auto platform omp explain
auto status --platform omp

When the installed OMP catalog lacks family, capability, or authorization metadata, models still returns exact native selectors and thinking support as a degraded display-only result. Automatic profile generation and strict routing remain blocked with catalog_metadata_insufficient; Autopus does not infer those fields or inspect credentials. A pre-existing explicit profile may instead select catalog_trust: operator-attested: Autopus first runs the same bounded strict probe, then uses only the exact intersection of native selectors and operator declarations while keeping auth/keyless unobserved.

Activation verifies every projected @role through an OMP RPC get_state session loaded with the generated overlay. It sends no prompt and makes no model-provider request; the resulting provider/model/thinking map is bound into the receipt and independently rechecked by explain/doctor.

From v0.50.126 OMP serves the 16 Autopus agents by default (SPEC-MODELUX-003). Generation writes .omp/agents/<agent>.md for every canonical agent, owned by the OMP manifest and carrying no model field; thinkingLevel is the agent's declared effort and appears only when no task.agentModelOverrides route names that agent. It also claims task.disabledAgents: [scout, security-reviewer, sonic, designer, librarian] in the project .omp/config.yml, which plain omp reads: every bundled agent but task across the omp releases Autopus supports (18.6.1 bundles the first three, 17.2.7 also designer and librarian), and a name the installed omp does not bundle is an inert filter entry. The bundled task agent stays enabled because an unnamed dispatch runs it, and the generated reviewer replaces the bundled one with the output schema of the omp 18.6.1 reviewer, so /review keeps its verdict. With a role profile, routes are written as task.agentModelOverrides keyed by the 16 names and task (task follows a task pin, then a planner pin, then the profile's planner route) through the existing activation, attestation, readback, and rollback; without a profile, pins alone (below) route only the pinned agents, and task through its own or planner's pin, and with no pin every agent inherits the session model.

auto update previews the move once — + .omp/agents/<agent>.md lines, task.disabledAgents: scout, security-reviewer, sonic, designer, librarian, and OMP now uses the 16 Autopus agents; set role_model_policy.agent_source: native to go back during v0.50.126. — and performs it in the OMP adapter's one transaction without writing autopus.yaml. An .omp/agents/<agent>.md the manifest does not record is a user file and stays byte-identical (.omp/agents/<agent>.md is a user file; left unchanged). A task.disabledAgents you set to another value is user-owned: Autopus leaves it, prints task.disabledAgents in .omp/config.yml is user-owned; scout, security-reviewer, sonic, designer, and librarian stay enabled, and doctor reports agents:disabled_agents_user_owned as degraded; only auto clean drops that mark.

Doctor's OMP agent catalog turns degraded with agents:autopus_agent_user_file while a canonical .omp/agents/<name>.md is a user file, because OMP then runs that file instead of the Autopus definition. A project .omp/config.yml the managed parser cannot read safely — a symlink, invalid YAML, several documents, an alias, a flow-style mapping, or a lone CR, NEL, LS, or PS line break — is never written: auto update prints task.disabledAgents was not written: .omp/config.yml cannot be read safely (a symlink, invalid YAML, several documents, an alias, a flow-style mapping, or a lone CR, NEL, LS, or PS line break); scout, security-reviewer, sonic, designer, and librarian stay enabled instead of the task.disabledAgents: … preview line, and doctor reports agents:project_config_unclaimable, a state distinct from a user-owned value. When a ledger already claims the key and the file then gains such a line break, the file and the ledger stay as they are, auto update prints task.disabledAgents is left as last written: …, and auto clean refuses with project_config_unclaimable until the file reads cleanly. Once no ledger claims .omp/config.yml, ADK keeps none of the backups its transactions recorded of that file except one: the rollback artifact of the newest committed transaction that released the ledger, applied to the config as that release left it. It stays under .autopus/backup/ until the next release replaces it, so a release can still be undone by hand; a release that deleted a config ADK had created leaves none, and auto clean restores from the verified preimage and leaves no rollback artifact of the config.

role_model_policy.agent_source: native restores the behavior described below for v0.50.126 only; the key and the native path are planned for removal in v0.50.127. Any other value fails to load with role_model_policy.agent_source_invalid, and the key alone configures no role model policy. Roll back with this key, not with an older binary: auto older than v0.50.126 cannot read an autopus.yaml that sets agent_source or quality.presets.*.efforts, and v0.50.125 rejects the project ownership ledger (.autopus/omp-model-project-ownership-v1.json) once it carries the new agent-surface fields. The next two paragraphs describe agent_source: native.

auto init writes no agent definition for OMP. OMP registers its own bundled agents (task, scout, reviewer, security-reviewer, sonic) and resolves a task agent by exact name, so a generated .omp/agents/<name>.md would only shadow the bundled one. A role profile binds models through task.agentModelOverrides instead, and the 16 ADK work roles collapse onto those five agents. auto platform omp explain and auto status --platform omp show one row per bundled agent with its role/capability provenance and effective selector; a project file at .omp/agents/<bundled name>.md is reported as native_agent_shadowed.

Several role keys can land on one bundled agent. The representative role decides — task follows planner, scout follows explorer, sonic follows validator — and every plan row, explain row, and receipt entry reports the governing key, so a pre-cutover policy that names all 16 roles keeps working and the applied entry stays visible. A set with no representative row has nothing to rank it and is refused with omp_native_agent_conflict naming the entries to reconcile.

Select the OMP mode and family together with profile apply balanced --family gpt|claude. The stored family names are openai and anthropic; both canonical names are also accepted. --plan previews the bundled agents, their requested/effective model and thinking, candidate order, fallback attempts, and blockers without writing configuration or activating a profile. Apply verifies the same routes through the installed OMP catalog and provider-free RPC readback.

Ordinary reviewers follow the selected family. Multi-provider review remains a separate orchestra.providers policy: changing this profile preserves its models and judge. For top-model review, keep Fable max, Astra max, and the selected Gemini model's highest supported thinking level (high for Gemini 3.1 Pro; it does not expose max).

Balanced has no implicit lower-model fallback. Missing models or unsupported thinking block apply before writes, and a blocked preview returns nonzero with per-agent reasons. Native retry.modelFallback is false when the profile has no explicit fallback chains.

Override one agent without copying the full profile:

auto platform omp profile apply balanced --family gpt --agent executor=openai-codex/gpt-6-astra:max --plan
auto platform omp profile apply balanced --family gpt --agent executor=openai-codex/gpt-6-astra:max
auto platform omp profile apply balanced --agent executor=inherit

Repeat --agent for multiple agents. Pins are stored in role_model_policy.agents.<name>.candidates and remain explicit overrides across profile/family changes until cleared with =inherit. On a native catalog without semantic metadata, a pin must have an exact family declaration in the shipped built-in profiles or the selected custom profile; arbitrary unknown models are rejected.

This standard matrix is shared with native Claude Code and Codex. OMP's built-in routes remain independent of custom quality.presets.balanced.agents tiers; OMP overrides use role_model_policy.agents. Ultra and custom OMP profile definitions retain their existing behavior. An explicit role_model_policy.profiles.balanced definition still wins over the built-in; --family is rejected for such a custom definition instead of silently ignoring it. OMP profile selection does not change quality.default, standalone Claude/Codex settings, or the supervisor's native model roles. auto quality show reports OMP's independent selection.

OMP policies are provider-neutral and inactive until a named profile is selected. Prefer the non-destructive overlay mode. The built-in profiles use shipped operator-attested declarations intersected with the installed catalog. For a custom profile, use strict mode when semantic catalog metadata is available; otherwise use operator-attested mode only after reviewing its exact selectors, families, capabilities, and thinking levels. A ${NAME} placeholder anywhere in a role_model_policy model candidate — its selector, thinking, family, or a key, also when reached through a YAML alias — fails the load with role_model_policy.candidate_placeholder_unsupported, which names the path only, while placeholders in the root profile and family still load. Placeholder values can no longer change the YAML structure of autopus.yaml: a value that would — for example one that starts with [, {, !, &, or *, or holds : , where its placeholder is unquoted — or that holds a line break fails the load with an error that names the variable and never the value, as does an explicit tag other than !!str on a placeholder; written as '${NAME}', a placeholder passes any printable one-line value without ' as the exact string. Custom profile example:

role_model_policy:
  version: v1
  profile: omp-balanced
  profiles:
    omp-balanced:
      config_mode: overlay
      catalog_trust: operator-attested
      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: [autopus_reviewer, autopus_security_auditor]

• Candidate order is the fallback order. In strict mode, selector, family, and thinking must match observed semantic catalog metadata. In operator-attested mode, selectors must exist in the bounded native catalog and family/capability/thinking come only from the explicit profile; this does not claim that authentication will succeed. 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 managed OMP bridge supports long-lived supervisor ACKs, exact canonical reinjection, and same-session compaction/rehydration through the current compaction.methodOrder contract. Generation alone still does not activate optimization: omp-active-probed remains fail-closed until a fresh task-owned runtime receipt proves the installed lifecycle and cleanup.
• 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.

Split compiler note:

• skills.compiler.mode: split is the default. It publishes the core/dependency set and all /auto routes; optional recipes remain available through auto skill list and auto skill info <name>.
• Select skills.compiler.mode: full to publish the complete compatible library, or choose bundles / explicit_skills. In split mode, opted-in long-tail skills use .opencode/skills/ or .autopus/plugins/auto/skills/; opencode_long_tail_target: shared and codex_long_tail_target: repo select native visible locations.

Pipeline entrypoints load their references/ details only when needed. Ordinary work stays inline; independent work, specialist review, or context isolation justifies delegation. auto pipeline run uses implement → validate → review only after one compact contract and the actual change set validate as low risk. Missing, ambiguous, malformed, or contradicted authorization retains all five phases. Resume requires the same authorized route. Annotation is opt-in through auto spec gates --annotation; validation, security, data-loss, and deterministic-oracle gates remain active.

The default workflow.coverage_threshold is 85. An explicit project or route threshold overrides it, and 0 turns the numeric gate off. Ordinary auto workflow context selects architecture through --conditional-profile architecture or --required-document; required core/SPEC bodies remain complete and verified. The signed OMP canonical-context and release-evidence paths are unchanged.

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

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/autopus-ai/autopus-adk/main/install.sh | sh

# Windows (CMD or PowerShell)
cd your-project
powershell -c "irm https://raw.githubusercontent.com/autopus-ai/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 to use @auto ...; otherwise invoke $codex-auto-<route> ...
• OpenCode: use /auto ... or /auto-<subcommand> ...
• Claude Code / Antigravity CLI / Oh My Pi: 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/autopus-ai/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

brew trust --cask insajin/autopus/auto
brew cleanup --dry-run

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

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/workflows/, CLAUDE.md
✓ Generated: .codex/skills/, .codex/agents/, .codex/hooks.json, .codex/config.toml, AGENTS.md
✓ Generated: .gemini/, GEMINI.md
✓ Generated: .opencode/, .agents/skills/, AGENTS.md, opencode.json
✓ Generated: .omp/rules/, .omp/skills/, .omp/commands/, optional .omp/extensions/
✓ Created: autopus.yaml

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
$codex-auto-setup  # Codex native skill 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.

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/*, .omp/*, OpenCode-owned .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.

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                    │
╰────────────────────────────────────╯

The workflow connects implementation, verification, and documentation. Confirm the project's acceptance criteria and recorded checks before deployment.

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:

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/*, .omp/*, and OpenCode-owned .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.

On machines with Autopus Desktop installed, the app owns the auto entry on your PATH (usually ~/.local/bin/auto) and replaces it with a small launcher script that brokers every call into the Desktop-managed CLI slot. When the app rejects the brokered call, auto exits with code 126 and prints:

Autopus Desktop managed ADK broker: managed_adk_broker_current_slot_rejected

The CLI itself is still installed. Invoke the managed binary directly — quote the path, it contains spaces — or alias it for the session:

"$HOME/Library/Application Support/co.autopus.desktop/managed-adk/current/auto" doctor
alias auto="$HOME/Library/Application Support/co.autopus.desktop/managed-adk/current/auto"

Then run auto update from that managed binary, or reinstall Autopus Desktop, so the launcher and the managed slot agree again. auto doctor reports the launcher as the doctor.launcher.desktop_shim check — it names the launcher path, the managed slot, and whether the managed binary is present.

/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)

Ordinary work stays inline; a validated compact contract uses the three-phase runtime route. This diagram illustrates the responsibilities in a full job, not a mandatory agent per phase:

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

    opt Explicit annotation request
        E->>A: Apply requested @AX tags
        A-->>E: Annotation result
    end
    E->>V: Gate 2: Build + lint + vet
    V->>T: Phase 3: Relevant tests + coverage gate (default 85%)
    T->>R: Phase 4: TRUST 5 + OWASP audit
    R-->>S: ✅ APPROVE

/auto go SPEC-ID --quality ultra      # Premium path for every role; Codex effort varies by role
/auto go SPEC-ID --quality balanced   # Top-model planning/review/debugging; lighter implementation

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
auto quality show --agents            # Also print the agent × platform model table, read-only
auto quality                          # Compose per-agent tiers by role group (interactive)

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 Astra 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.

Claude Code and Codex share the standard balanced role matrix with OMP:

The native files are .claude/agents/autopus/*.md (model, effort) and .codex/agents/*.toml (model, model_reasoning_effort). Apply both through auto quality balanced --apply, or use the provider-specific commands above. Complete historical default layouts receive the standard placement without rewriting their YAML; an explicit different agent tier or a custom quality preset keeps its existing interpretation. One custom agent does not change its siblings.

Codex retains the exact standard profile with an unverified diagnostic when its native catalog cannot be read. If an observed catalog rejects the requested model or effort, generation stops before writing files instead of substituting an older model or lower effort.

Ultra is unchanged: its seven-role core uses Fable/Astra, and remaining roles use Opus/Sol. Claude Ultra emits max for its Fable/Opus agents; Codex Ultra uses Astra/max and Sol/xhigh. Supervisors still inherit by default; quality-managed Codex supervisors use Astra/ultra in Ultra and Astra/xhigh in Balanced. Native multi-provider review defaults to Fable 5.1/max and Astra/max in both modes, while explicit provider model/effort pins remain untouched.

auto quality without arguments, every interactive auto update, and the auto init quality step share one composition step. It shows one row per agent and one column per installed platform:

A quality.providers.claude or quality.providers.codex override labels its column override: <preset>. Edits go through four role groups that exist only on screen and are never written to autopus.yaml:

At Change the composition?, g picks one tier per group (Enter keeps a group, mixed ones included), a takes agent=tier lines, and u (shown only when one exists) updates previous-version defaults to the current ones without changing quality.default. The preview marks every changed cell with * and lists every key it writes; only y writes. A result equal to the current balanced or ultra default binds that preset, writing the preset too when it is missing or a previous-version copy; any other result is saved as quality.presets.custom with quality.default: custom. Each written balanced and ultra preset is classified by its exact agent map: current default, previous-version default (can update) (a default an earlier release wrote), or customized (N differ) (anything else, partial maps included). A customized preset's agents and description are never modified; a composition that would bind it is saved as custom.

auto update asks once on every interactive run, after its project settings prompts. Enter, input end, or three invalid answers keep autopus.yaml byte-identical; a confirmed change is generated in the same run, so the agent files equal the preview. --yes/-y, --auto, --check, a CI value other than empty, 0, or false, non-TTY input, and --workspace neither ask nor change anything and print at most one line naming previous-version presets; --plan, --preview, and --dry-run skip the step. auto quality --apply regenerates right after a confirmed change, and auto quality show --agents prints the table read-only on any stdin.

The composition step and the per-tool editors refuse to write, and leave the file unchanged, while the quality block holds a ${NAME} value of a set environment variable or a YAML alias or merge key. They rewrite only the quality block: every byte outside it stays identical, comments inside it are not kept.

In a terminal the step runs as huh forms, like the init wizard: arrow keys and Enter choose, and scripted or piped input keeps the line prompts above. The menu offers keep, update to the latest defaults (only when a built-in preset is a previous-version default), edit the composition, and, in auto quality, choose the default preset. edit the composition opens a tool tab:

• same for all tools is the role-group composition above; every tier option names the cells each tool would get, for example opus — Claude Code: opus-5-5 · high | Codex: 6.1-sol · xhigh.
• Claude Code and Codex change only that tool. Options read <cells> [<tier>], such as 6.1-sol · xhigh [opus], and the result is saved as quality.presets.<provider>-custom (all 16 agents) with quality.providers.<provider>: <provider>-custom. Binaries older than v0.50.88 ignore quality.providers.
• OMP appears in auto update and auto quality when the platforms saved in autopus.yaml include omp, and in auto init when the platforms this init chose include omp. It pins one installed model and thinking level per native OMP agent, previews the plan's real routes, and writes through the existing OMP profile writer; init keeps the choice in its config and writes it with the platform files.

A Claude label names the model slug (opus-5-5), while the table cell is the frontmatter value the generator writes (opus, or an exact ID such as fable-5-1). auto update also shows a column for every platform it detected in this run and generates them from the same catalog read. Ctrl+C in a form ends auto update with a non-zero exit before any generation, cancels auto init before it writes any file, and leaves auto quality without an error.

On the Claude Code and Codex tabs every tier select, for a group or for one agent, is followed by that tool's effort select (SPEC-MODELUX-003). keep is the default; the other options are the levels every affected agent accepts, each labelled with the cells it leaves and (automatic) on the automatic one. Claude Code offers low, medium, high, xhigh, and max and leaves haiku agents out; Codex offers the levels the run's catalog reports for the agents' models, and with an unreadable catalog it asks nothing and prints Codex catalog unreadable; effort stays automatic. An agent given its automatic level stores nothing; the rest are written with the tiers to quality.presets.<provider>-custom.efforts, a map from canonical agent names to those five levels. The preview lists quality.presets.<provider>-custom.efforts written (N agents) (or rewritten, removed) and, for the file's first efforts map, auto older than v0.50.126 cannot read quality.presets.*.efforts. An explicit effort takes the agent off the standard balanced placement and is generated as Claude effort: (never for haiku) and as Codex model_reasoning_effort, resolved from the same catalog read. Later edits, u, and the per-tool line editor keep stored efforts, and same for all tools moves a binding preset's efforts into custom (quality.presets.custom.efforts carried (N agents)) instead of writing them onto another built-in. A file without efforts loads and generates exactly as before. The OMP tab picks a model and its thinking per role group or agent instead, as the next paragraph describes.

In the default autopus mode the OMP tab asks no profile or family (SPEC-MODELUX-003). It starts from the current pins, or from the current profile's effective routes (one pin-less plan, whose blocker returns to the menu with nothing written), and asks role groups (the four groups above) or per agent (the 16 agents, plus task while it has its own pin). Every model select offers keep — <current>, keep the OMP session model, and the installed models of every provider that Autopus can attest, as provider/model sorted by provider and model; a group lists only models every member accepts. A strict catalog attests by its family and capability metadata; a catalog without that metadata lists only the installed models the built-in profiles declare, after The installed OMP catalog has no family or capability metadata; only models Autopus can attest are listed. The thinking select offers only the levels the chosen model reports. keep the OMP session model drops the pin, and for task the option reads follow planner: task follows planner unless it has its own pin, so the Design·Judgment group moves planner but never an existing agents.task. The result is saved through the existing OMP plan, write, and activation as pins without a profile — role_model_policy: {version: v1, agents: {<agent>: {candidates: [{selector, thinking, family}]}}}, with config_mode and agent_source kept — and the preview adds auto older than v0.50.126 cannot read role_model_policy.agents without a profile. Only pinned agents get an override, and task gets one through its own or planner's pin; the rest inherit the session model, and a pinned model the catalog no longer has fails closed with required_route_unresolved. Applying over a profile-based policy prints Moving the routes of OMP profile <name> to per-agent pins; unchanged agents keep their routes. before the confirmation: every unchanged agent keeps all candidates of its route, task keeps its route (stored as agents.task only when it differs from planner's), the legacy scout, security-reviewer, and sonic keys are expanded into per-agent pins and removed, profile and family go away, and an ultra profile's link to quality.presets.ultra ends. Cancelling leaves the file unchanged. Native mode, the OMP line editor, and auto platform omp profile apply keep using profiles, and a profile with pins still loads. Agents without a profile are an autopus-mode policy: with agent_source: native the load fails with role_model_policy.agents_require_profile and says to set role_model_policy.profile or remove agent_source: native. On a catalog without family or capability metadata, the tab warns at start when the current profile's routes use models outside the built-in declarations: Autopus cannot attest these models on this catalog, so applying with them unchanged is blocked: <models>; the OMP line editor or auto platform omp profile apply keeps editing the profile.

OMP keeps these axes separate: --team selects the owner-omp native task/hub/todo topology, while --multi adds provider-diverse planning and review. --team --multi composes both.

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.

For more control, run each stage separately:

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

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 a read-only multi-provider planning advisory plus final SPEC review · --prd-mode minimal for lightweight PRDs · --skip-prd to go straight to SPEC

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
Optional   │ 📝 Annotator       │ @AX tags only when explicitly requested
Gate  2    │ ✅ Validator        │ Build + lint + vet
Phase 3    │ 🧪 Tester          │ Relevant tests + coverage gate (default 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

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, auto sync verify creates a read-only staging plan and prints the topology it detected: a multi-repo workspace splits into Phase A (module) and Phase B (meta), while a single Git repository holding autopus.yaml — including a linked worktree — yields one commit group. Both topologies exclude generated/runtime, tracked-but-ignored, unclassified, and shell-unsafe paths; a location that is neither fails with an unsupported topology: diagnostic under its own exit code. --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                    │
╰────────────────────────────────────╯

Three commands connect requirements, implementation, and delivery records. Check the recorded acceptance and verification results before treating the work as complete.

Every review scores across 5 dimensions:

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

Every provider runs as a headless subprocess (or through OMP for a backend: omp provider), and each command runs all rounds synchronously and prints its result directly. WebSearch fallback when Context7 docs are unavailable.

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
  forbidden_trailers: [Co-Authored-By]  # default; [] allows them

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

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 + debate judge, subprocess and OMP backends)
│   ├── 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, 53 skills)
└── configs/            # Default configuration

"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.

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 — Reuse existing code and native capabilities before adding dependencies or abstractions
• Isolation — Use native isolation when available; shared-workspace writers require disjoint ownership. Conversation forks alone do not isolate files.
• No blind installs — Security Auditor agent flags unknown or unvetted packages before they enter the codebase

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/autopus-ai/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/autopus-ai/autopus-adk/releases/latest | grep tag_name | sed 's/.*"v\(.*\)".*/\1/')
curl -LO "https://github.com/autopus-ai/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/autopus-ai/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.

• 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

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