Tool Integration — verified 2026-08-26

How the local routers are consumed from the desktop tools. Model ids and contexts come from documents/08-model-catalog.md; this document does not restate the catalog. Router ids below are the tuned aliases (live since 2026-08-26); a full router id works anywhere an alias is shown.

aider

All workspace projects share 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 usage (observed in the workspace .aider.* files, model-router-setup sessions, 2026-08): cloud models via OpenRouter (--model openrouter/openrouter/auto, openrouter-gpt-4o) through the LiteLLM proxy, whole edit format, repo-map 4096 tokens. The local-model path replaces that for private and free work.

Local-model invocation against the beast router (aider 0.86.2):

# routine edits — GPU-resident coder on beast
    aider --openai-api-base http://beast.lan:2001/v1 --openai-api-key sk-local \
          --model coder32 --edit-format whole
    
    # architect mode — 120B plans on CPU, coder edits on GPU
    aider --openai-api-base http://beast.lan:2001/v1 --openai-api-key sk-local \
          --model gpt-oss-120b --editor-model coder32 --architect

Rules:

Use --edit-format whole; diff formats are unreliable on local models.
Editing sessions use the GPU-resident coder (coder32); reserve architect mode (120B planner, slow, CPU) for design work, not routine edits.
Do not put R1-distill models in the agent loop; they follow tool schemas poorly.
Keep sessions short: aider holds the whole conversation in context, and coder32 has 16384 context.
Athena variant for ultra-low-latency small work: same invocation with --openai-api-base http://athena.lan:2001/v1 and --model coder3.

llama.vscode

The official llama.cpp extension (ggml-org.llama-vscode, version 0.0.63 installed on athena) provides inline completion, chat, an agent, embeddings, and tool selection, and can use a remote llama.cpp server.

Best use against the local routers:

Purpose	Model
Inline completion (FIM)	`coder3` (athena, local) or `coder32` (beast)
Chat	`coder32` or `devstral` (beast)
Agent	`gpt-oss-20b` (beast GPU preset since task 01; the extension’s own recommended agent model) or `gpt-oss-120b` for heavy runs
Embeddings	`nomic-embed` (athena; `/v1/embeddings`)

Point the extension at http://beast.lan:2001/v1 (or http://athena.lan:2001/v1 for the small completions), and select the per-purpose model in its model picker. Inline completion requires a FIM-capable model — the Qwen2.5-Coder family is the right one in this catalog. MCP tools from installed VS Code MCP servers can be selected for the agent.

VS Code Chat View

The Chat view is tied to GitHub Copilot. Three options exist for local or OpenAI-compatible providers:

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.
Third-party gateway extensions that route Copilot Chat to self-hosted models (for example github-copilot-llm-gateway, AgenticProxy on the marketplace).
Skip the Chat view: llama.vscode chat and Cline/Roo Code already cover chat and agentic editing without Copilot.

Decision (2026-08-26): option 3. The owner suppresses Copilot and uses llama.vscode; no Copilot plan is needed. Suppression steps:

Extensions panel: disable or uninstall the Copilot extensions. On athena the installed Copilot-related extension was ms-azuretools.vscode-azure-github-copilot (1.0.231) — removed 2026-08-26 (extension directory deleted; llama.vscode 0.0.63 unaffected). The standard github.copilot / github.copilot-chat were not present.
Per-workspace fallback: "github.copilot.enabled": false in .vscode/settings.json keeps the extensions installed but off.

Do not run overlapping agent UIs on the same project simultaneously; pick one surface per task (DSH for agentic sessions, aider in the CLI, llama.vscode in the editor).

Tool-Against-Task Summary

Task	Tool
Agentic repo work, background goals	DeepSeek Harness (DSH) with the beast/athena providers
Quick CLI pair-programming in a project	aider with `coder32`
Inline completion and in-editor chat	llama.vscode
Chat-with-docs, RAG	Open WebUI (optional; embeddings endpoint live on athena) or llama.vscode chat

Version-Dependent versus Stable Facts

Version-dependent (recheck on tool upgrades): aider flags and version (checked 0.86.2); llama.vscode capabilities and version (checked 0.0.63); Copilot extension ids/versions; --edit-format behavior on local models.
Stable (live-state, verified 2026-08-26): router endpoints and aliases (http://beast.lan:2001/v1, http://athena.lan:2001/v1), model roles from documents/08-model-catalog.md, the routing policy (documents/10-routing-policy.md).