deeploop 0.1.8

DeepLoop autonomous research autopilot control plane.

DeepLoop

Structured research missions from a bundled starter or local project folder — with visible autonomy boundaries, durable mission state, and an explicit operator inbox.

DeepLoop helps researchers and operators run structured work from existing artifacts or a bundled starter instead of rebuilding everything around one long chat. It keeps the loop moving, pauses only at real safety, authority, or support boundaries, and makes the path legible when you need to inspect or redirect it.

DeepLoop owns behavior and orchestration; substrate repos own reusable domain or science rules.

Why it matters

• Start from a real folder or a bundled starter: use the same mission flow either way.
• Keep control visible: status, inbox, and resume make the operator inbox explicit when DeepLoop needs a real decision.
• Inspect the loop: operator-facing summaries expose runtime telemetry, inner-loop progress, stage-kernel activity, reroutes, and temporary gaps instead of hiding them in raw JSON.
• Keep evidence close to the work: your project folder stays focused on facts, docs, and outputs while DeepLoop keeps durable mission state.
• Use autonomy with governance: the shipped path includes explicit release boundaries, autonomy governance, and reviewed promotion surfaces.
• Separate platform from domain logic: DeepLoop runs the loop; substrate repos keep reusable methods, constraints, and science rules.

Getting started

1. Install DeepLoop

   Fastest supported path:

   pip install deeploop
   

   For the full install matrix — GitHub installs, editable contributor setup, two-clone hybrid workflows, and the documented Conda path — use Getting started.

2. Prepare the workspace

   export DEEPLOOP_WORKSPACE_ROOT="$HOME/Workspaces"  # optional; choose before init/run/start
   deeploop setup
   deeploop preflight
   

   DeepLoop prints the resolved workspace root during deeploop init and deeploop start. If DEEPLOOP_WORKSPACE_ROOT is unset, it uses an existing unambiguous ~/Workspaces, ~/workspace, or ~/workspaces directory, then falls back to ~/workspaces.

   If you are in a repo checkout and want the contributor validation path too:

   make setup
   make public-bootstrap-check
   

3. Prepare a provider

   • quickest first-run machine check:

     deeploop provider-ready --selection-profile control-plane-copilot-cli
     

   • Provider setup

   • Provider selection

   deeploop run now performs the same machine-level readiness check before kickoff for the default first-run control-plane profile. If setup is still missing, it stops with the exact missing requirement, one next step, and a resume command.

4. Run DeepLoop

   • start from nothing with the installed bundled starters:

     deeploop run --until-complete
     

     DeepLoop opens an interactive kickoff, asks for your mission idea, lets you choose a bundled starter, materializes a project under WORKSPACE_ROOT/projects/, compiles a mission, and launches the same operator loop as every other path.

   • start from your own folder with the same command surface:

     deeploop run --project-root <project-folder> --until-complete
     

   • start from the repo example if you also cloned DeepLoop:

     cp -R examples/translation-budget-ladder PROJECT_FOLDER
     deeploop run --project-root PROJECT_FOLDER --until-complete
     

     Repo-local examples/... paths only exist in a DeepLoop source checkout. Package installs still get bundled starters through plain deeploop run.

     Note: If <project-folder>/.deeploop/missions/*.yaml files exist, deeploop run automatically uses the first one instead of bootstrapping a blank mission. For a plain folder with no existing config, it bootstraps from the folder's facts. If the folder is rough but still recognizable, DeepLoop can initialize with disclosed clarifications/defaults and keep the project folder unchanged. To target a specific explicit config directly, use deeploop init --config <mission-config.yaml> followed by deeploop start --mission-state <mission-state.json>.

   • advanced discovery / operator path:

     deeploop init --project-root PROJECT_FOLDER --force
     deeploop init --discover --project-root PROJECT_FOLDER --force
     

     Use plain deeploop init --project-root ... when the folder is rough but already recognizable. Add --discover when you want DeepLoop to ask clarifying questions before kickoff.

   On a copied folder, substitute PROJECT_FOLDER in the commands above. If DeepLoop cannot safely bootstrap the folder yet, it stops with bounded bootstrap-repair guidance instead of mutating the project.

5. Use the operator CLI when a run pauses

   deeploop status --mission-state MISSION_STATE_PATH
   deeploop inbox --mission-state MISSION_STATE_PATH
   deeploop resume --mission-state MISSION_STATE_PATH
   

   Start with status. Open inbox only when DeepLoop pauses for you. Use logs, decisions, retry, reroute, or triage only when the surfaced handoff says you need more detail or a managed-mode override.

The deeploop CLI is the single entry point — run, init, status, inbox, resume, and more are all subcommands.

Readiness at a glance

• Unified first run: Linux, Python 3.11+, pip install deeploop, deeploop setup, deeploop preflight, then deeploop run --until-complete
• Same front door for existing work: use deeploop run --project-root <project-folder> --until-complete
• Repo-checkout validation path: make setup, make public-bootstrap-check, and direct access to examples/translation-budget-ladder/
• Messy starts are supported: rough plain-folder projects can initialize with disclosed clarifications/defaults, or you can use deeploop init --discover ... for a guided kickoff
• Repair stays bounded: if the folder is missing the plain-folder bootstrap contract, DeepLoop exits with bootstrap-repair guidance and suggested starter inputs instead of silently rewriting project files
• Current baseline release proof: make public-bootstrap-check, make docker-release-validate, and make docs-build

Best fit today

DeepLoop is best when you have:

• either a project folder on disk or a clear enough idea to start from a bundled starter
• a clear mission or question
• an operator who can check status and respond when the operator inbox opens
• a need for bounded autonomy, durable state, and evidence-aware summaries

Public alpha — best on Linux with Python 3.11; not claiming a fully automatic experience for everyone. See the roadmap for current scope.

Key capabilities

Operating modes

• sandboxed-yolo for the fastest bounded path when you want DeepLoop to keep moving inside the supported guardrails
• managed when you want intervention hooks before DeepLoop continues; managed mode can surface a bounded retry, reroute, or downscope step for review
• human-directed when you want to approve important choices yourself

What you can inspect

• operator-facing status surfaces runtime telemetry, inner-loop progress, ratchets, reroutes, and temporary-gap recovery hints
• stage-kernel execution stays visible instead of disappearing behind one opaque agent loop
• the operator inbox keeps handoffs explicit when DeepLoop reaches a real decision or support boundary

Reusable methods and governance

• keep skills for reusable methods and domain/science rules in substrate repos
• use Release posture for the current claim and Autonomy governance for current boundaries
• review the current multi-substrate proof as proof of a bounded contract, not a claim of broad portability

Documentation

• Docs home
• Getting started
• Examples
• Plain-folder starter
• Release posture
• Portable bootstrap
• Provider setup
• Provider selection
• Autonomy governance
• Multi-substrate proof
• Technical reference

Contributing

Contributions, bug reports, and discussion are welcome.

• Contributing guide
• Code of conduct
• Security policy
• Changelog