pi-model-discovery

A Pi extension that auto-discovers local model providers (Ollama first) and registers their models with Pi on startup — no /reload required.

What it does

On every Pi session start, this extension:

1. Queries your configured local provider(s) for installed models
2. Pulls per-model metadata (context window, vision/reasoning/tools capabilities) from Ollama's /api/show endpoint
3. Registers all discovered models via Pi's registerProvider() so they're immediately available
4. Optionally adds them to enabledModels in settings.json so they appear in the model picker
5. Caches capability data to avoid hammering /api/show on every startup

Install

pi install npm:@kylebrodeur/pi-model-discovery

Update an existing install:

pi update npm:@kylebrodeur/pi-model-discovery

Commands

| Command | Description | |---------|-------------| | /providers status | Show current config, registered models, and capability counts | | /providers sync | Re-run the discovery sync (uses cache) | | /providers sync --force | Bypass the capability cache and refetch from Ollama | | /providers info [model] | Show model card (defaults to current model) | | /providers card | Open the model card popup for the current model | | /providers footer on/off | Toggle the footer status indicator | | /providers init | Create a default local-providers.json config file | | /providers reload | Reload config without restarting Pi | | /providers debug on/off | Toggle debug logging | | /providers help | Show command help |

Footer status

A compact indicator in pi's footer that complements (not duplicates) the built-in model display. It shows only what pi doesn't already show — capabilities and special variant badges, plus context window size:

☁ ◉◆⏵  ctx 524K

• ☁ cloud/remote model
• ⚡ QAT (Quantization-Aware Training, e.g. Gemma 4)
• ◇ embedding model
• ◉ vision (dim when off)
• ◆ thinking (dim when off)
• ⏵ tools (dim when off)
• ctx 524K — context window size (not usage; pi's bar shows usage)

Icons are dim when the capability is absent, green when present. Full capability words appear only in the model card.

Toggle with /providers footer on/off or set in config.

Model card popup

Full details for the current model, opened as a popup. Triggered by:

• Keyboard shortcut: ctrl+i
• Slash command: /providers card (current model) or /providers info [model] (any model)

minimax-m3:cloud ☁ cloud

minimax · 550B

context     524K tokens
size        384B (remote)
digest      6d55374b63bb

● vision   ● thinking   ● tools

Press Escape to close. /providers status output looks like:

Local Providers Status:
Sync on startup: yes
Add to scope: yes
Cleanup stale: no
Debug: off

Providers:
  ollama: watching (http://127.0.0.1:11434)

Registered Ollama models: 21
  5 vision | 7 reasoning | 21 tools
  text-only: 0

Configuration

Config file: ~/.pi/agent/local-providers.json (global) or ./.pi/local-providers.json (project, overrides global).

{
  "debug": false,
  "syncOnStartup": true,
  "addToScope": true,
  "providers": {
    "ollama": {
      "enabled": true,
      "baseUrl": "http://127.0.0.1:11434",
      "cleanupStale": false,
      "cacheTtlHours": 24
    }
  }
}

Provider fields

| Field | Type | Default | Description | |-------|------|---------|-------------| | enabled | boolean | true | Set to false to skip this provider | | baseUrl | string | http://127.0.0.1:11434 | Ollama HTTP API root | | cleanupStale | boolean | false | Remove this provider's entries from enabledModels if the model is no longer in Ollama | | cacheTtlHours | number | 24 | How long to cache /api/show results. 0 = always refetch |

Top-level fields

| Field | Type | Default | Description | |-------|------|---------|-------------| | debug | boolean | false | Log extra info on session start | | syncOnStartup | boolean | true | Run discovery during the async factory (before session_start) | | addToScope | boolean | true | Push discovered models into settings.json enabledModels | | showFooterStatus | boolean | true | Show the footer status indicator with capabilities and context. Use /providers footer off to disable. |

Capability detection

Starting with Ollama 0.30, all metadata is available directly from /api/tags — no separate /api/show call needed:

• context window — from details.context_length (1M+ for models that support it)
• vision — from capabilities: ["vision"]
• reasoning — from capabilities: ["thinking"] or ["reasoning"]
• tools — from capabilities: ["tools"]
• embedding — from capabilities: ["embedding"]
• parameter size, family, quantization, format — from details.*
• size, digest, modified_at — from top-level tag fields
• remote/cloud — when remote_model/remote_host are set
• QAT — detected from -qat suffix (Gemma 4 quantization-aware training)

Cache

Capability data is cached at ~/.pi/agent/ollama-model-cache.json. Entries older than cacheTtlHours are refetched on the next sync. Use /providers sync --force to bypass the cache.

Troubleshooting

Models show in --list-models but not in the model picker

addToScope: true writes to settings.json enabledModels during session_start. If you started a session before installing the extension, restart the session (or run /providers sync and reload).

"Ollama not reachable" warnings on startup

Ollama isn't running, or the baseUrl is wrong. Check with:

curl http://127.0.0.1:11434/api/tags

Capabilities seem wrong for a model

Run /providers sync --force to bypass the cache. If still wrong, check what /api/show returns directly:

curl -X POST http://127.0.0.1:11434/api/show \
  -H 'content-type: application/json' \
  -d '{"name": "gemma4:12b"}'

Stale entries in enabledModels after deleting models from Ollama

Set cleanupStale: true in the provider config, then /providers sync.

How it works

The extension uses Pi's async extension factory. On startup:

1. Async factory queries Ollama's /api/tags and registers the provider via pi.registerProvider() — models are available immediately with no /reload needed.
2. Per-model /api/show calls populate contextWindow, input (text/image), reasoning, and tools from the authoritative API.
3. Capability results are cached to ~/.pi/agent/ollama-model-cache.json (TTL configurable; default 24h).
4. During session_start, the extension writes the discovered models into settings.json enabledModels so they appear in the model picker.

License

MIT