pi-dial

Pi provider extension for DIAL Core.

Configuration

In interactive Pi, store the DIAL API key with:

/login dial

Pi persists it in ~/.pi/agent/auth.json. The legacy environment variable is still accepted as a fallback, but is no longer required:

export DIAL_BASE_URL="https://your-dial-core.example.com"
# Optional fallback only:
# export DIAL_API_KEY="your-api-key"

# Optional deployment used by /dial-usage. If omitted, the first deployment
# from /openai/models is used.
export DIAL_USAGE_MODEL="gpt-4"

The extension discovers deployment names from GET /openai/models. DIAL requests use the documented deployment endpoint:

/openai/deployments/{deployment_name}/chat/completions

Use a discovered deployment in pi:

pi --model dial/<deployment-name>

If model discovery is unavailable, specify fallback deployment names:

export DIAL_MODELS="gpt-4o,claude-sonnet-4"
# or a single deployment:
export DIAL_MODEL="gpt-4o"

The API key is sent in the DIAL-specific Api-Key header. Model IDs are deployment names returned by DIAL and may differ from the upstream model names.

Startup & model discovery (non-blocking)

The provider registers immediately with the seed/fallback list (DIAL_MODELS / DIAL_MODEL, or empty) so pi is usable the instant the extension loads — it never blocks startup waiting for GET /openai/models.

Live model discovery runs in the background via pi's refreshModels callback:

• pi's cache-only startup phase (or a cancelled refresh) returns the already-known list without touching the network.
• When network is allowed, GET /openai/models is fetched with an 8s per-request timeout; on success the discovered list is persisted to pi's provider cache (publish({ persist })) and hot-swaps the catalog.
• On failure it keeps the previous list, so the user is never left without models.

If DIAL returns HTTP 401 or 403, Pi reports that the key may be expired or invalid and suggests running /login dial to replace the stored credential.

What DIAL Core exposes (capabilities)

This extension targets DIAL Core (epam/ai-dial-core). Per its OpenAPI spec the catalog API (GET /openai/models) returns deployments whose ModelType is only CHAT / COMPLETION / EMBEDDING, with a capabilities object of booleans (chat_completion, completions, embeddings, fine_tune, inference) and — for chat deployments — input_attachment_types (e.g. image/*). DIAL Core has no image-generation or video-generation endpoint, so this extension does not implement image/video generation routing (the capability flags for those stay false).

Each deployment is registered with the capabilities DIAL actually reports, read from the model item's type / ModelType, capabilities, input_attachment_types, and input_modalities / output_modalities (the detection tolerates all of these schema variants). The catalog entry carries:

• reasoning — extended-thinking models (reported via reasoning/supports_reasoning/reasoning_effort).
• input: ["text"] or input: ["text", "image"] — vision-capable deployments accept images as input (from input_attachment_types / input_modalities).
• a capabilities block (tools ← chat_completion, vision ← image input, image, video, audio, reasoning) surfaced in /list-models.
• display_name is read from the { plainValue: "…" } object DIAL returns.

At request time a single provider-level streamSimple routes only chat deployments to POST /openai/deployments/{name}/chat/completions (OpenAI completions). Non-chat deployments (embedding/…) are registered for visibility but return a clear “not streamed by this extension” error if selected.

So pi --model dial/<vision-deployment> can be shown images inline, while image/video generation is not supported by DIAL Core and is therefore not routed.

Commands

• /dial-prices [input|output|total|context] [desc] — list DIAL model prices per 1M tokens (sorted; desc reverses).
• /dial-usage — query the documented DIAL Core /v1/deployments/{deployment}/limits endpoint using Api-Key authentication. Set DIAL_USAGE_MODEL to choose the deployment; otherwise the first model from /openai/models is used. A valid DIAL_API_KEY is required.
• /dial-capabilities [image|video|audio|vision|reasoning|tools] — list each deployment's capabilities (vision / image / video / audio / tools / reasoning). An optional filter narrows the table to deployments that support that capability, e.g. /dial-capabilities vision shows only DIAL deployments that support that capability.

Install

Install the published package:

pi install npm:pi-dial

Or install directly from GitHub:

pi install git:github.com/pgciq/pi-dial

Run from a local checkout

From this directory:

pi -e ./extensions/dial.ts

Or add this directory/package to your pi configuration.

Development

Requires Node.js 22.19 or newer.

npm install --ignore-scripts
npm test
npm pack --dry-run

The tests use Node's built-in test runner and make no DIAL requests.