LLM Council (VS Code Extension)

Minimal, streaming “council” workflow using the VS Code Language Model API (e.g., Copilot models). Runs three stages, logs streams, and saves a markdown artifact per run.

Current state

• Command: LLM Council (editor context menu / llmCouncil.run).
• Flow: pick context (file/selection/none) → pick prompt template → edit prompt → pick models (defaults auto-resolved and remembered per workspace) → council runs (3 stages) with streaming output.
• Models: via vscode.lm.selectChatModels; defaults prefer gpt-5.1, sonnet-4.5, gemini-pro-3 when available.
• Output: streams tagged by stage/model; a markdown artifact (with front matter + transcripts) is saved to the workspace root and opened in a new tab. History kept in global storage (llmCouncil.historySize).
• Tests: unit/integration via Vitest.

Installation

Prereqs: VS Code ≥1.84, VS Code CLI code on PATH, Copilot (or other LM API provider) with chat models enabled.

Dev/local install

npm install                 # install deps
npm run package:vsix        # builds dist/llm-council.vsix
code --install-extension dist/llm-council.vsix --force  # install into VS Code
# Optional: npm pack  # to produce github-llm-council-0.0.1.tgz

User install after publish

npm install -g github-llm-council   # postinstall auto-installs the bundled VSIX via `code`

Upgrade

npm update -g github-llm-council    # or npm install -g github-llm-council@latest

Publish (maintainers)

npm install
npm run package:vsix   # produce dist/llm-council.vsix
npm pack               # produce github-llm-council-0.0.1.tgz
npm publish            # prepublishOnly rebuilds/package before publish

Usage

1. Right-click in an editor → LLM Council (or run via Command Palette).
2. Choose context (file/selection/none).
3. Pick a prompt template, then edit/enter your prompt.
4. Confirm model selection (defaults pre-selected and remembered per workspace).
5. Watch streaming output; a markdown artifact opens and saves to the workspace root.

Requires: VS Code with LM API-capable extension (e.g., GitHub Copilot) and access to chat models.

Configuration

• llmCouncil.defaultModels (array): preferred model ids; first available three are used when no stored selection exists.
• llmCouncil.historySize (number): maximum run summaries to retain.

Testing

npm run lint
npm run test:unit
npm run test:integration

Notes

• Model choices are remembered per workspace between runs.
• Prompts live in src/prompts.ts; edit and rebuild/package to distribute changes.