Feature: Aider and VS Code

Defines local-model usage for aider and the VS Code surfaces (llama.vscode, VS Code Chat).

Aider

Aider is the CLI pair-programmer used across the workspace projects. Every project shares the same rule-injection pattern: .aider.conf.yml reads the project’s prompts/01-contract.md, 02-workflow.md, 03-conventions.md, the common/ files, the semantic-sort flavor, and tools/aider-rules.md. Historical sessions used cloud models via OpenRouter (openrouter/openrouter/auto) through the LiteLLM proxy; the local-model path is the goal of this feature.

Local-model invocation against the beast router:

aider --openai-api-base http://beast.lan:2001/v1 --openai-api-key sk-local \
          --model <router model id or alias> --edit-format whole

Rules:

• Use --edit-format whole; diff formats are unreliable on local models.
• Editing sessions use the GPU-resident coder (Qwen/Qwen2.5-Coder-32B-Instruct-GGUF:Q4_K_M).
• Architect mode splits planning from editing: --model <gpt-oss-120b id> --editor-model <coder id> --architect. The 120B runs on CPU and is slow; reserve it for design work, not routine edits.
• Avoid R1-distill models in the agent loop; they follow tool schemas poorly.
• aider holds the whole conversation in context; keep sessions short because context is the scarce resource on the GPU models.
• Use the router model id, or an alias once configured on the router (alias = coder32 in the preset). See prompts/features/01-endpoints-and-models.md.

llama.vscode

The official llama.cpp VS Code extension (ggml-org.llama-vscode) provides inline completion, chat, an agent, embeddings, and tool selection. It can use a remote llama.cpp server. Best use against the local routers:

• Completion model: a FIM-capable small coder (Qwen/Qwen2.5-Coder-1.5B/3B on athena, or the 32B coder on beast).
• Chat model: the GPU-resident coder or Devstral on beast.
• Agent: gpt-oss-20b or gpt-oss-120b (both in the router catalogs).
• Embeddings: ggml-org/Nomic-Embed-Text-V2-GGUF:Q8_0 (athena).

VS Code Chat View

The Chat view is tied to GitHub Copilot. Options for local or OpenAI-compatible providers, in order of preference for this setup:

1. GitHub Copilot BYOK (bring your own key): Copilot Chat accepts an OpenAI-compatible base URL and key; point it at http://beast.lan:2001/v1. Requires a Copilot plan.
2. Third-party gateway extensions that route Copilot Chat to self-hosted models (for example github-copilot-llm-gateway, AgenticProxy on the marketplace).
3. Skip the Chat view entirely: llama.vscode chat and Cline/Roo Code already cover chat and agentic editing without Copilot.

Avoid running overlapping agent UIs on the same project simultaneously; pick one surface per task.

Requirement: usage notes stay verified

• Tool behavior changes; documents describing extensions or flags carry a verified <date> marker and cite the version checked.
• Do not duplicate the router catalog here; reference 01-endpoints-and-models.md.