Feature: Aider and VS Code

Purpose

Defines local-model usage for aider and the VS Code surfaces (llama.vscode, VS Code Chat), so everyday editing, completion, and chat run against the house routers instead of a cloud provider.

Requirements

• AIDER-R001 — Use --edit-format whole; diff formats are unreliable on local models.
• AIDER-R002 — Editing sessions use the GPU-resident coder (coder14).
• AIDER-R003 — The 120B runs on the CPU router and is slow; reserve it for design work and long batch runs, not routine edits.
• AIDER-R004 — Architect mode pairs a planner with an editor, but aider applies one --openai-api-base per run, so both models must live on the same router.
• AIDER-R005 — Avoid R1-distill models in the agent loop; they follow tool schemas poorly.
• AIDER-R006 — aider holds the whole conversation in context; keep sessions short, because context is the scarce resource on the GPU models.
• AIDER-R007 — Use the router alias (coder14, gpt-oss-120b) or the reported model id, never the UD- preset-key form.
• AIDER-R008 — The canonical llama.vscode settings fragment is sources/config/llama-vscode-settings.json; apply it with scripts/llama-vscode-config-apply.sh.
• AIDER-R009 — tests/03-llama-vscode-settings-validate.sh guards every referenced model id against the captures.
• AIDER-R010 — Do not include /v1 in an endpoint value; the extension appends /infill and /v1/chat/completions itself.
• AIDER-R011 — Completion must run against a single-model server, never a multi-model router. The extension’s FIM-priming and token-counting calls omit the model field, which a router rejects with 400 model name is missing from the request; completion entries therefore carry a localStartCommand that starts a single-model llama-server on 127.0.0.1:8012.
• AIDER-R012 — Chat and agent requests do carry model, so the routers serve those tiers.
• AIDER-R013 — Completion must use a FIM-capable model; the Qwen2.5-Coder family is the right one in this catalog.
• AIDER-R014 — Do not offer coder14 as a tools model; it does not emit parsed tool_calls.
• AIDER-R015 — Keep the MI25 tier to one model per working session; the GPU router runs --models-max 1 and two different models evict each other.
• AIDER-R016 — The Chat view must not be left on a Copilot cloud model by default. On a free SKU that model can be rejected with 400 model_not_supported and the chat appears to hang.
• AIDER-R017 — Do not run overlapping agent UIs on the same project at the same time; pick one surface per task.
• AIDER-R018 — Tool behavior changes; documents describing extensions or flags carry a verified <date> marker and cite the version checked.
• AIDER-R019 — Do not duplicate the router catalog here.

Behavior

• An edit session against the MI25 completes with --edit-format whole and the configured coder model.
• Inline completion is served by the extension’s own single-model server on 127.0.0.1:8012; chat and agent requests are served by the beast routers.
• The Chat view lists a local model only after the llama-vscode env is selected.

Dependencies

• prompts/features/01-endpoints-and-models.md

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 replaces that for private and free work.

Local-model invocation against the beast routers:

# routine edits — GPU-resident coder on the MI25
    aider --openai-api-base http://beast.lan:2001/v1 --openai-api-key sk-local \
          --model coder14 --edit-format whole
    
    # long planning or batch run — 120B on the CPU router
    aider --openai-api-base http://beast.lan:2002/v1 --openai-api-key sk-local \
          --model gpt-oss-120b --edit-format whole

llama.vscode

The official llama.cpp VS Code extension (ggml-org.llama-vscode) provides inline completion, chat, an agent, embeddings, and tool selection, and can use a remote llama.cpp server. It does not need to start llama-server itself: localStartCommand is optional, so the routers serve it directly.

The tiered mapping:

VS Code Chat View

The Chat view is provided by GitHub Copilot Chat, which is installed and authenticated on athena. Local models reach it through the llama-vscode language-model provider that the extension registers, so the env must be selected before any local model appears in the picker. Options for local or OpenAI-compatible providers, in order of preference for this setup:

1. The llama.vscode provider: select an env, then pick one of its models in the Chat view’s model picker.
2. GitHub Copilot BYOK (bring your own key): use the customoai provider pointed at a router base URL, since Copilot’s BYOK list has no llama.cpp-router entry.
3. Third-party gateway extensions that route Copilot Chat to self-hosted models (for example github-copilot-llm-gateway, AgenticProxy on the marketplace).
4. Skip the Chat view entirely: the llama.vscode chat and agent panels cover chat and agentic editing without Copilot.