Added

• Added a dedicated support for vLLM backend option.

Improved

• Ollama is now the default backend when no backend is explicitly configured,
  with http://localhost:11434 as the default address.
• Improved backend labels, descriptions, and connection messages
• Updated the README, documentation, and Marketplace metadata to reflect the expanded Selfcoder 1.3 backend support and improve discoverability.

Added

• Plan Mode — Added a new mode alongside Chat and Agent for exploring and understanding a repository without making changes. Plan sessions have dedicated interface messaging and history metadata, and can seamlessly continue in Agent when you're ready to start implementing.
• OpenAI-compatible backends — Added support for vLLM, llama.cpp, and other OpenAI-compatible servers that expose /v1/models and streaming /v1/chat/completions .
• Backend authorization — Added the optional machine-scoped Selfcoder.backendAuthorizationToken setting for bearer-token authentication with OpenAI-compatible servers, LM Studio, and Ollama.

Improved

• Renamed the backend endpoint setting from Selfcoder.backendURL to
  Selfcoder.backendAddress and updated settings and connection messages to use
  the clearer Backend Address wording.

Fixed

• Agent change tracking now attributes files from explicit edit events instead of
  scanning the workspace for unrelated differences. It rechecks Git ignore
  rules, removes ignored and no-op entries, and keeps safe session-start
  baselines for review and revert.
• Fixed native @Selfcoder requests not receiving sidebar pins and recent editor
  history, including files visited before the sidebar was opened.
• Fixed Agent change summaries and revert controls being lost when switching to
  Plan and then back to Agent.

Added

• Added support for built-in slash commands, including /help, /context, /review, /models, /export, /new and /compact, /summarize, and /init.
• Added support for user-defined custom slash commands.
• Added a What's New view to the Selfcoder menu.
• Added the Elevated permission preset. Autonomous now runs without permission prompts, while Elevated provides broad autonomy with safeguards for destructive actions.
• Added support for discovering AGENTS.md instruction files in the workspace.

Improved

• Native participant now fully recognizes instruction file types: AGENTS.md, AGENTS.MD, CLAUDE.md, CLAUDE.MD, local-instructions.md and .github/copilot-instructions.md.
• Improved the accuracy of the context window indicator in Agent Mode.
• Agent change tracking now respects Git ignore rules when scanning or recording changed files

Fixed

• Fixed diagnostics and linting issues appearing as regular context files in the composer.
• Fixed escaping of model detail data in webview scripts.
• Fixed the positioning of the agent approval dialog.

Added

• Follow-up messages can now be queued while a response is still generating. They are sent automatically once the current response finishes.
• Added agent permission presets (Selfcoder.permissionMode): choose Autonomous, Balanced (default), or Restricted to decide how much the agent can do on its own before asking for approval.
• Added an in-app Support button for users who want to support Selfcoder’s ongoing development.

Improved

• Agent mode is now selected by default when the extension starts.
• Agent mode now shows the model's reasoning process inline as it works
• Improved agent activity messages with clearer file details and completed-action states.
• Agent mode now keeps its session warm between turns, so follow-up messages start responding faster.
• The agent permission approval prompt has been restyled to look and feel more native.
• Tooltips across the chat now use a single editor-themed style.

Fixed

• Fixed Ollama models not showing their reasoning while working in agent mode.
• Fixed the agent change view listing files with no actual changes, and newly created files sometimes missing from the diff.
• Fixed the agent's external-directory approval prompt not showing which directory was being requested
• Fixed approving a flagged agent command producing no further output.

Added

• Selfcoder now warns when the selected model's context window is too small for the current mode — agent mode needs at least 32K tokens, chat mode at least 8K — and explains how to raise it on the active backend.

Improved

• The context-window indicator now shows the window the model is actually running at instead of its theoretical maximum.
• The model-loading progress bar in chat now appears for all LM Studio models, not only reasoning-capable ones.
• Agent sessions start faster in large workspaces.

Fixed

• Fixed agent mode silently finishing without a response when another OpenCode installation exists on the same system.
• Fixed out-of-memory failures with some Ollama models on smaller GPUs, and an unnecessary model reload when switching between chat and agent mode
• Fixed Ollama unloading the model after a few minutes of inactivity, which caused a slow reload on the next message.
• Fixed reasoning output leaking into the answer as plain text when reasoning was turned off on the LM Studio backend
• Fixed stopping or clearing the chat while an agent run was still starting up — the run is now cancelled immediately instead of continuing in the background.
• Fixed the model-loading progress display of the native @Selfcoder chat participant.

Added

• Added a configurable response timeout (Selfcoder.responseTimeout) for both chat and agent responses, defaulting to 30 minutes. When the limit is reached, Selfcoder aborts the response and shows a timeout message. This prevents runaway processes and gives users control over long-running responses.

Improved

• Improved chat and agent streaming performance by coalescing rapid token updates into a single webview render per frame, keeping the interface smooth during fast responses.
• Improved editor responsiveness during agent sessions
• Refined the appearance and placement of the model-loading indicator.
• Unified the styling of the copy buttons across the chat webview.
• The chat welcome screen now displays the Selfcoder logo.

Fixed

• Fixed agent mode silently producing no output when the backend responds with an error; the backend's own error message is now surfaced in the chat.
• Fixed old responses continuing to stream after the user starts a new message or clears the chat.
• Fixed the loaded-model status not updating correctly in agent mode.
• Excluded cloud-hosted models from the status bar's loaded-models list
• Fixed the Ollama backend emitting visible reasoning for some models even when reasoning was turned off

Improved marketplace metadata to make the extension easier to discover.
Updates readme badges and links to reflect the current state of the project

Added

• Agent mode is now installed automatically the first time it is used.
• A progress popup is shown while a local model is loading into backend memory.
• The status bar now lists all models currently loaded into memory.
• The sidepanel chat now indicates when the backend is disconnected and offers a reconnection action.
• The native chat participant now shares the same model selection as the sidepanel chat, removing duplicated configuration.
• Unified UI component for reasoning and tool-call output.
• Agent mode can now process images when the selected model supports vision.

Improved

• File diff visualization has been reworked for a cleaner review experience.
• Token usage calculation is more accurate, giving better predictions against the context window limit.
• Parameter display has been improved for cloud-hosted Ollama models.

Fixed

• Fixed an exception thrown by the stop button in agent mode.
• Fixed the response stream not stopping when a message is deleted mid-stream.
• Fixed the status bar not updating when the backend is reconnected from the sidepanel.

• Updated readme file
• Added link to github repository

• Initial release of Selfcoder extension.