OptPilot Assistant

OptPilot Assistant is the optional participant inside a Studio Conversation. A Conversation is the durable discussion thread; the Assistant is not a separate kind of conversation. It helps users understand what OptPilot can do, find published Catalog capabilities, prepare Studies, operate interfaces, work with editable Workspaces, and understand Run evidence.

Conversation is the default Studio surface, not a permanent side panel. Users can still open Catalog and choose components directly. When a simulator, Run, Workspace, or other focused tool occupies the main area, Ask from this page opens the currently selected Conversation as an overlay without restarting the focused tool. Studio includes the visible page and selection as bounded, read-only context. This action neither creates another Conversation nor grants access to a Workspace.

The Assistant is part of Studio, not the PyPI core package.

What A Conversation Coordinates

A Conversation can combine several existing OptPilot objects without becoming their source of truth:

• exact Environment, Method, and Resource references from Catalog
• an OptPilot Study, which configures an Environment, Method, objective, and budget
• active and completed Runs
• editable Workspaces explicitly made available to the Conversation
• interactive interface launches
• generated outputs and Candidate evidence
• approvals and Assistant execution events

These objects remain usable without the Assistant. Open work is a narrow process monitor for interfaces, Run setup launches, and running Runs. Durable Studies, completed Runs, and editable Workspaces stay in their named destinations, while Assistant activity and approvals stay with the Conversation. Closing the Conversation view does not stop work. Conversations, Workspaces, read-only viewers, saved Studies, completed Runs, and ordinary Assistant messages do not appear in Open work.

The Conversation list uses one short topic title plus only useful state, such as Working, Needs approval, or a nonzero Workspace count. A new Conversation receives an immediate title from its first substantive request; when the Assistant is available, the same Assistant turn may refine that title and update it later only if the primary goal changes. Greetings, confirmations, and “continue” do not rename a Conversation.

Recommendations And Cards

The Assistant may recommend Catalog entries or propose a Run setup in a structured card. Cards are rendered from a bounded Studio presentation contract and carry exact object identities. Studio validates card actions against a small allowlist before invoking existing launch, open, save, or review flows.

Ordinary model-authored Markdown is explanatory only. It cannot create a privileged Studio action. Starting a Run, writing files, running commands, or stopping work remains explicit and approval-aware.

For a Run setup, the Assistant can propose an Environment, a compatible Method, an objective, and a budget. Users can edit these values or open the detailed Study configuration. The visible UI calls this saved configuration a Run setup, while the underlying study schema, API, route, and command-line concept keep the study name.

Runtime Modes

Studio can run the Assistant in several modes:

Mode	What works	What it needs
Disabled or unreachable	Studio keeps the local Conversation and shows status, but no model/tool execution occurs.	No runtime.
Model chat	Chat-style answers grounded in Studio context.	Configured model/API key, for example OpenRouter or an OpenAI-compatible chat-completions endpoint.
OpenHands agent server	Assistant tool execution through the Studio bridge.	OpenHands-compatible agent server plus model/API key.
Workspace tools	Read/write files, run shell commands, and open previews in Workspaces made available to the Conversation.	OpenHands bridge and a Workspace runtime.

The OpenHands bridge has been checked with openhands-agent-server==1.29.0. OpenHands currently expects Python 3.12, so run it from a Python 3.12 environment when enabling tool execution. Studio gives OpenHands a small native inspection/planning tool set for codebase search. Terminal and file-editing calls use OpenHands-compatible Studio tool names, but they are still Studio client tools: Studio executes them through Conversation Workspace-access checks, editable-Workspace rules, runtime execution, and approvals instead of letting OpenHands edit files or run shell commands directly.

Install the runtime packages in the source-checkout environment:

uv pip install -U openhands-sdk openhands-tools openhands-workspace openhands-agent-server

Start OpenHands:

OPENHANDS_SUPPRESS_BANNER=1 uv run --no-sync agent-server --host 127.0.0.1 --port 8781

Start Studio:

uv run optpilot ui --host 127.0.0.1 --port 8765

Configure the Assistant in Studio Settings, or use environment variables:

OPTPILOT_OPENHANDS_URL=http://127.0.0.1:8781
OPTPILOT_OPENHANDS_SESSION_ENDPOINT=/api/conversations
OPTPILOT_OPENHANDS_MODEL=deepseek/deepseek-v4-flash
OPTPILOT_OPENHANDS_API_KEY=...

OPTPILOT_OPENHANDS_API_KEY can fall back to LLM_API_KEY or OPENAI_API_KEY.

flowchart TB
  Conversation["Studio Conversation"]
  Assistant["OptPilot Assistant\nparticipant"]
  Catalog["Catalog\nEnvironment + Method + Resource"]
  Card["reviewable card\nor Study"]
  Approval["explicit action\nand approval"]
  Active["running Run\nor interface"]
  Durable["Studies, Runs,\nand Workspaces"]
  ActiveWork["Open work"]

  Assistant --> Conversation
  Conversation --> Catalog
  Catalog --> Card
  Card --> Approval
  Approval --> Active
  Approval --> Durable
  Active --> ActiveWork
  ActiveWork --> Conversation

Settings And Local Variables

Studio settings have two scopes:

Settings area	Purpose
OptPilot	OpenHands URL, model, API key, OptPilot capabilities, and approval defaults.
Local environment variables	Machine-local environment variables that component configs may request through envFromHost.

Values are write-only in the browser. Studio can show that a value is configured, but it does not echo the value back into the page. They are stored as plaintext in OptPilot's local settings file, with mode 0600 where the platform supports it. This local file is not a secret vault.

Components should declare the environment variables they need. For example, an LLM Method can declare OPENROUTER_API_KEY in its runtime environment requirements, and Studio can inject the locally configured value only when that name is requested.

For direct CLI Runs, envFromHost reads from the shell process environment. Studio settings are separate local values used only by Studio-managed setup, interface launch, Assistant, and Study-launch paths.

Workspaces In A Conversation

The Assistant works with editable Workspaces that the user explicitly makes available to a Conversation. The Conversation's right-hand Workspaces in this conversation panel lists those projects and is the single place to add, open, choose the default, or remove a project from the Conversation. This grants file context without copying the project or transferring ownership. Creating a Workspace or adding a local folder from that panel makes it available to the current Conversation immediately. Removing a Workspace from the Conversation does not delete it.

The visible page is separate from the Conversation's Workspaces. Ask from this page can include the current Catalog item, Study, Run, Candidate, Workspace screen, or interface as bounded read-only context with requests sent while that page is open. That does not let the Assistant edit files. File editing and shell commands are available only for an editable Workspace that the user explicitly makes available to the Conversation.

It can inspect read-only context such as:

• visible Studio selection and exact object coordinates
• available Workspace code through native OpenHands search tools
• Catalog entries
• Study configs
• Run summaries and evidence files
• OptPilot documentation

It can act on editable available Workspaces when allowed:

• read files
• write files
• run shell commands in the Workspace runtime
• open Workspace previews
• prepare package plans
• draft or save Study YAML

The Assistant should not modify immutable Catalog source directly. Editing package code requires an editable Workspace. Launching a declared Catalog interface is different: Studio keeps its source read-only and gives the process private launch-scoped runtime and output storage.

Approvals

Higher-impact actions are approval-gated in Studio. This includes:

• writing files
• running shell commands
• launching Studies and Runs
• stopping jobs
• applying package plans

OpenHands-native tools are limited to low-risk inspection and planning, such as grep, glob, and task_tracker. Studio exposes OpenHands-compatible optpilot_terminal and optpilot_file_editor as client tools so the model gets familiar software-engineering interfaces while OptPilot keeps control of paths, Workspace runtime, and approvals.

Approval records are stored under .optpilot-ui/ with the local Conversation state. Pending approvals remain visible in their Conversation; the top-level Open work shelf is reserved for interfaces and Runs.

Per-Launch Inputs

A Run setup can declare inputs — per-launch values such as the plain-language problem statement a one-shot solving Run setup expects. The Assistant reads the declared names, types, and descriptions from the Run setup's validation and passes their values when it launches, exactly as the Studio launch form does.

If a required input (one declared without a default) has no value, the launch is blocked before any Realm work with the code study_inputs_required, which names the unbound inputs so the Assistant can ask you for them rather than guessing. Because input values are the problem payload and are retained in Run evidence, the approval card shows the values themselves — you approve the exact problem that will run. Never put a credential in an input; secrets belong in Studio Settings environment values, which stay out of retained evidence.

When OpenHands Is Not Available

If OpenHands is disabled or unreachable, Studio still keeps local Conversations, Catalog browsing, Open work, and normal manual interfaces. It shows a clear status instead of pretending that Assistant execution is available. Tool execution, Workspace edits, shell commands, and Assistant-initiated Study launches require the OpenHands-backed tool path. Regular Studio Launch run actions still use the local OptPilot runner and do not require OpenHands.