Configure an advisor and manage optional models

The deterministic advisor requires no model. The optional advisor panel lets you choose Ollama, a local or remote Chat Completions endpoint, or one MCP HTTP advice tool. It opens without contacting the provider or downloading weights.

VS Code

Run PerfChecker: Configure advisor and manage models, use the settings icon in Scenarios & advice, or choose Model settings in an investigation.

1. Select a provider and enter its endpoint. Remote connections require an explicit HTTPS opt-in. Enter only the name of a credential environment variable; never paste a token into the panel.

2. Choose Test connection / discover. This reads the available models or MCP tools, without sending project evidence or generating advice.

3. Choose Use beside a model/tool. For MCP, check the prompt argument and any other required arguments shown in its schema. The selection remains explicit; a tool name is not a guarantee that the server contains an assistant.

4. Customize the instructions and save. The panel updates the configured advisor file, or creates perf/advisor.json, then selects it in workspace settings. Existing custom provider fields are retained through the common validator.

5. Diagnose a scenario and choose Explain with configured model. The displayed saved evidence is used directly; otherwise the extension offers its history.

Choose Rule-based advice only and save to disable optional advice. This does not remove model files. Choosing experiments is separately enabled and bounded by count and duration. MCP requires structured responses for this mode; free text is displayed as unverified advice and never executed.

Optional local model files

Install/start Ollama through its official guide. The panel connects to that existing service; it does not install system software. With a local Ollama endpoint selected, the panel provides:

• Model inventory with sizes reported by the server.

• Explicit download of an exact model name, with a confirmation step.

• Unloading from memory while retaining the files.

• Deletion from the shared server after confirmation.

• Cancellation and a configurable deadline, including worker startup.

Size before download and free disk space can be unknown. Reported installed sizes include shared layers and must not be summed as unique disk usage. Cancellation stops the local worker; the server may retain partial downloads. Refresh its inventory before deciding whether another operation is needed. Deleting a model affects every project using that Ollama installation. Management is disabled for remote servers and other providers.

Web interface (Oxygen) and Pluto

The web interface links to the same panel. Saving applies to that studio and persists under its reports directory in advisor-settings.json. Investigation and model-management jobs cannot run concurrently in that studio. Requests use the studio session token and are restricted to the existing loopback service.

Generated Pluto notebooks have an Optional advisor setup section with provider, prompt, MCP fields, discovery, save and model lifecycle controls. Only the action button launches a request; changing fields never does. The confirmation checkbox is required for model mutations and overwriting an existing configuration. Refresh displays the returned models/tool schemas; copy the selected identifier into the configuration before saving. The advisor file is then used by the existing narration/investigation controls.

The REPL and CLI consume the same backend, for example:

After saving perf/advisor.json through the panel, the equivalent Julia calls are:

using PerfChecker
config = load_advisor_config("perf/advisor.json")
result = advisor_setup(config) # inventory only; no generation/evidence
job = launch_advisor_setup(config; action=:models)
investigation_status(job)
cancel!(job)

The advisor-setup CLI takes --source=request.json containing config, action, optional model, and confirmed for mutations. Its JSON output has explicit availability, evidence_sent=false, and generation_tested=false. The connection check cannot establish model quality, and some otherwise compatible model endpoints do not expose a model-list route.

See MCP advice for protocol limits and advisors for deterministic fallback and evidence verification.