Logics Ollama Specialist

Platform scope:

• ollamainstallmacos.sh is intentionally macOS-specific.
• ollama_check.sh is a POSIX-shell helper and should not be treated as a Windows workflow entrypoint.
• On Windows, follow the Ollama product's native install path first, then adapt the integration guidance in references/.

Normalize common user variants such as deepseeker-coder-v2, deepseek coder v2, or deepseek-coder to deepseek-coder-v2. Treat Qwen coder, qwen coder, or qwen-coder requests as a Qwen-family coding profile and confirm the exact tag before changing config.

Quick start

1. Clarify OS, target model, editor surface, and whether Ollama runs locally or on another host.
2. If installation is needed, ask for explicit permission before running install commands.
3. For local coding, prefer the repo's curated DeepSeek default profile deepseek-coder-v2:16b unless the user explicitly asks for another supported family or tag.
4. Verify Ollama, daemon reachability, and model availability before editing editor config.
5. If the app is a frontend, prefer a same-origin proxy (/ollama/*) to avoid CORS.
6. If LAN URL causes 403, remove Origin in the proxy or set OLLAMA_ORIGINS.

Do not default to deepseek-coder-v2:236b unless the user clearly has server-grade hardware.

Supported profile guidance:

• DeepSeek default profile: deepseek-coder-v2:16b
• Qwen coding profile example: qwen2.5-coder:14b
• If the user asks for qwen3, qwen3.5, or another Qwen-family coder tag, confirm the exact local Ollama tag and treat the repo profile as overrideable rather than forcing one hard-coded tag.

Install Ollama (macOS-only helper)

• Ask for explicit approval before running installs.
• Use the script for safe and consistent setup:

DRY_RUN=1 logics/skills/logics-ollama-specialist/scripts/ollama_install_macos.sh deepseek-coder-v2:16b

If approved, run again without DRY_RUN=1.

DeepSeek Coder V2 workflow

Prefer the explicit 16B tag for local development:

ollama pull deepseek-coder-v2:16b

Validate the model path:

ollama list | grep -E '^deepseek-coder-v2'
ollama run deepseek-coder-v2:16b "Write a Python hello world"

Validate the local HTTP API:

curl -fsS http://127.0.0.1:11434/api/chat -d '{
  "model": "deepseek-coder-v2:16b",
  "messages": [{"role": "user", "content": "Reply with the word ready"}],
  "stream": false
}'

Use deepseek-coder-v2:latest only if the user explicitly wants the floating tag.

Qwen coding profile workflow

Use a Qwen-family coding profile when the operator explicitly prefers Qwen over DeepSeek for bounded local coding or hybrid assist work.

Curated example:

ollama pull qwen2.5-coder:14b

Validate the profile path:

ollama list | grep -E '^qwen'
ollama run qwen2.5-coder:14b "Write a Python hello world"

If the user wants another Qwen-family coder tag, confirm that exact tag first and then use it consistently in runtime config, Continue, and checks.

Continue in VS Code

When the user wants a local VS Code coding assistant and does not name another extension, prefer Continue.

The user-level config file is:

~/.continue/config.yaml

Minimal working model entry:

name: Local Assistant
version: 0.0.1
schema: v1

models:
  - name: DeepSeek Coder V2
    provider: ollama
    model: deepseek-coder-v2:16b
    apiBase: http://localhost:11434
    roles:
      - chat
      - edit
      - apply

Rules for Continue edits:

• Patch the existing file instead of overwriting unrelated models.
• Keep apiBase: http://localhost:11434 unless the user says Ollama is remote.
• Prefer explicit tags instead of latest.
• Keep the selected model family consistent across runtime config and editor config instead of mixing DeepSeek and Qwen tags accidentally.
• If the user wants faster inline completion, pair the main DeepSeek model with a smaller completion model instead of forcing one large model to do everything.

Roo Code

If the user explicitly wants Roo Code, keep the same Ollama baseline and wire the editor after validation:

• API provider: Ollama
• Base URL: http://localhost:11434
• Model ID: deepseek-coder-v2:16b

Validate in this order:

1. ollama --version
2. curl -fsS http://127.0.0.1:11434/api/tags
3. ollama list | grep -E '^deepseek-coder-v2'
4. Roo Code provider, base URL, and model selection match the local setup

Dedicated local autocomplete

If the user wants low-latency inline completion, prefer a split-model setup:

• deepseek-coder-v2:16b for chat, edit, and apply workflows
• a smaller local completion model such as qwen2.5-coder:1.5b for autocomplete

Use this pattern when:

• the main coding model is responsive enough for chat or edit work but too heavy for inline completion
• the user wants a local setup that feels closer to Copilot-style suggestions

Do not collapse these roles into a single larger model by default when responsiveness is the operator's primary concern.

Start and stop Ollama server

• Start in foreground:

ollama serve

• Start as Homebrew service:

brew services start ollama

• Stop foreground server: Ctrl + C
• Stop Homebrew service:

brew services stop ollama

• If Ollama.app auto-restarts the server, quit the app:

osascript -e 'quit app "Ollama"'

• Last-resort force stop:

pkill -f 'ollama serve'
pkill -x Ollama

• Verify server status:

lsof -iTCP:11434 -sTCP:LISTEN -n -P

Verify Ollama

Run the check script to validate prerequisites, reachability, model presence, and optional Continue config hints:

logics/skills/logics-ollama-specialist/scripts/ollama_check.sh deepseek-coder-v2:16b
logics/skills/logics-ollama-specialist/scripts/ollama_check.sh qwen2.5-coder:14b

Integrate with a frontend (Vite/React)

1. Use a proxy to call /ollama/* from the browser.
2. Set VITEOLLAMAHOST to the local Ollama URL.
3. If LAN access returns 403, remove the Origin header in the proxy or set OLLAMA_ORIGINS.

See references/ollama-integration.md for snippets and common pitfalls.

Add PWA support (Vite)

Use vite-plugin-pwa, register the SW, and add icons or manifest fields. See references/pwa-vite.md.

Troubleshooting

Common failures and fixes:

• ollama: command not found

Install Ollama first.

• curl ... /api/tags fails

Start the Ollama app or run ollama serve.

• Model missing from ollama list

Run ollama pull deepseek-coder-v2:16b.

• Wrong model family in config

Re-check that the selected DeepSeek or Qwen tag matches the intended runtime profile and editor config.

• Continue or Roo Code cannot answer

Re-check provider, model, and base URL against the local Ollama setup.

• Browser app on LAN returns 403

Remove Origin in the proxy or set OLLAMA_ORIGINS.

• Requests are very slow or fail under load

The machine likely does not have enough RAM or VRAM for the chosen model; move to a smaller model or a smaller completion model where appropriate.

Resources

scripts/

• ollama_check.sh: POSIX-shell helper to verify ollama, curl reachability, local model availability, and optional Continue config hints for the requested DeepSeek or Qwen tag.
• ollamainstallmacos.sh: macOS-only helper to install or start Ollama with an optional model (supports DRY_RUN=1).

references/

• ollama-integration.md: DeepSeek and Qwen profile notes, editor-integration notes, proxy or CORS guidance, env vars, and endpoint notes.
• pwa-vite.md: PWA steps for Vite or React.