Skip to content

settings-and-privacy.md

Latest commit

 

History

History
170 lines (109 loc) · 8.13 KB

settings-and-privacy.md

File metadata and controls

170 lines (109 loc) · 8.13 KB
Raw

Settings And Privacy

Selfcoder is local-first: requests are sent to the backend address you configure, usually LM Studio, Ollama, vLLM, or llama.cpp running on your own machine.

This page explains the main settings and how they affect behavior.

Privacy Model

By default, Selfcoder sends model requests to your configured backend address.

For a typical local setup, that means:

  • LM Studio at http://localhost:1234, or
  • Ollama at http://localhost:11434, or
  • an OpenAI-compatible server at http://localhost:8000

If you configure Selfcoder.backendAddress to point to a remote server, requests go to that server. Treat the backend address as the boundary for where prompts, attached files, selected workspace context, and tool-call results are sent.

Privacy Protection Mode

Selfcoder.privacyProtection controls native integration behavior.

Value Behavior
standard Enables native VS Code chat integrations and normal Selfcoder model listings.
strict Disables native chat integrations and hides cloud-tagged models from Selfcoder model lists when that metadata is available.

Use strict when you want to avoid potential data sharing with VS Code's native chat features or models tagged as cloud-hosted. This setting cannot infer whether an arbitrary OpenAI-compatible address is local or remote; verify the address yourself.

Main Settings

All settings are under the Selfcoder namespace.

Setting Default What it does
Selfcoder.backend LMStudio Selects LMStudio, Ollama, or OpenAICompatible.
Selfcoder.backendAddress backend default Sets the backend server address. Compatible addresses may be a server root or end in /v1.
Selfcoder.backendAuthorizationToken empty Sends an optional bearer token to the selected backend. Machine-scoped, but stored as an ordinary unmasked setting.
Selfcoder.privacyProtection standard Controls native chat integration and cloud-tagged model visibility.
Selfcoder.systemPrompt helpful coding assistant prompt Sets the base instruction sent with requests.
Selfcoder.enableVision enabled Enables image attachment support for vision-capable models.
Selfcoder.reasoning medium Sets reasoning effort for models that support it.
Selfcoder.historyScope global Shows all saved conversations or only conversations for the current repository.
Selfcoder.historyRetention 100 Sets the maximum number of conversations kept in history.
Selfcoder.contextMode enabled Enables or disables automatic editor and workspace context.
Selfcoder.responseTimeout 30 Response timeout in minutes for Chat, Plan, and Agent responses.
Selfcoder.permissionMode balanced Permission tier for Plan and Agent: autonomous, elevated, balanced, or restricted.

Backend Address And Authentication

Use Selfcoder.backendAddress when your backend is not running on the default address.

Examples:

http://localhost:1234

http://localhost:11434

http://localhost:8000

http://localhost:8080/v1

Make sure the address matches the selected backend. For OpenAICompatible, Selfcoder accepts either the server root or its /v1 base and then calls /models and /chat/completions itself. Do not enter a complete route such as /v1/models.

If you previously set a custom Selfcoder.backendURL, copy that value to Selfcoder.backendAddress after upgrading. The old key is no longer used.

Set Selfcoder.backendAuthorizationToken only when the selected server requires bearer authentication. Selfcoder sends it as Authorization: Bearer <token> to LM Studio, Ollama, or an OpenAI-compatible backend. Blank values omit the header.

The setting has machine scope to reduce accidental Settings Sync or repository sharing, but VS Code does not encrypt or password-mask it. Keep real tokens out of workspace settings, screenshots, and checked-in examples. Selfcoder keeps the raw token out of the generated OpenCode configuration and does not pass it to terminal commands started by the model. Selfcoder does not currently support custom authentication headers.

System Prompt

The system prompt is the base instruction Selfcoder sends with each request.

Use it for stable preferences, such as:

  • concise answers
  • preferred language
  • testing expectations
  • code style
  • security posture
  • "ask before making broad changes" rules

Keep the system prompt short enough that it does not waste too much model context.

Context Mode

Selfcoder.contextMode controls automatic context collection.

Value Behavior
enabled Selfcoder may include active editor content, diagnostics, recent files, git diffs, or repository snippets when useful.
disabled Selfcoder skips automatic context providers. Explicit context still works.

Even when automatic context is disabled, Selfcoder can still use pinned files, manual attachments, workspace instructions, system prompt, and chat history.

Response Timeout

Selfcoder.responseTimeout sets how long a Chat, Plan, or Agent response can run before Selfcoder stops it automatically, the same as pressing Stop.

The value is measured in minutes and is clamped from 1 to 1440. The default is 30 minutes.

Plan And Agent Permissions

Selfcoder.permissionMode controls which Agent actions run immediately and which Plan or Agent actions require approval.

Value Behavior
autonomous Runs without permission prompts.
elevated Allows almost everything, but asks before destructive or irreversible commands, .env access, external directories, and runaway-loop protection.
balanced Recommended default. Lets the agent handle reads, edits, and basic shell commands, while requiring approval for destructive commands, network access, and sensitive file edits.
restricted Asks before most non-allowlisted shell commands, denies reading secrets, and blocks external-directory access.

In Agent mode, file edits are allowed by default in every tier, change tracking remains active, and the current session supports one-click revert.

Plan starts with the selected tier's read rules, but hard-denies file changes, terminal commands, delegated work, and model-controlled exit into Agent in every tier. For example, balanced Plan can ask before a web request or external-directory read, but approving that request cannot unlock editing or terminal access. See Plan Mode for the full safety model.

Plan And Agent Runtime

Plan and Agent use a managed OpenCode runtime. On first use, Selfcoder downloads the platform-matched runtime into extension storage, verifies it, and reuses the cached copy later.

History

Selfcoder can save completed sidepanel conversations so you can resume them later.

Selfcoder.historyScope controls what appears in history:

  • global: show conversations across workspaces
  • repo: show conversations created for the current repository

Selfcoder.historyRetention controls how many conversations are kept.

Image binary content is not kept in conversation history, which helps avoid storing large image payloads in VS Code state.

Workspace Instructions

Workspace instruction files let you define project-specific behavior without repeating it in every prompt.

Selfcoder checks for:

  1. local-instruction.md (or local-instructions.md)
  2. .github/copilot-instructions.md
  3. AGENTS.md (or AGENTS.MD)
  4. CLAUDE.md (or CLAUDE.MD)

Agent mode also uses deeper discovery and includes instructions from the folders which are closer to the current context, while chat mode uses only the root-level discovery. This allows you to have different instructions for different subfolders in a repository.

Use these files for project rules and stable context that should apply across requests.

Sensible Defaults

For most users:

  • use standard privacy mode
  • keep contextMode enabled
  • keep enableVision enabled
  • use medium reasoning
  • use balanced Agent permissions
  • keep history scope as global unless you work across many repositories

Switch to stricter settings when you need a narrower, more controlled workflow.