@cad0p/pi-deepinfra-provider

DeepInfra provider for pi: dynamic model catalog, full reasoning-effort thinking levels (max included), vision, monthly billing footer (fork of pi-deepinfra)

Packages

Package details

extension

Install @cad0p/pi-deepinfra-provider from npm and Pi will load the resources declared by the package manifest.

npm repohome report
$ pi install npm:@cad0p/pi-deepinfra-provider
Package
@cad0p/pi-deepinfra-provider
Version
0.1.5
Published
Oct 20, 2026
Downloads
411/mo · 224/wk
Author
cad0p
License
MIT
Types
extension
Size
31.6 KB
Dependencies
0 dependencies · 2 peers
Pi manifest JSON
{
  "extensions": [
    "./index.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

@cad0p/pi-deepinfra-provider

A pi extension that adds the DeepInfra provider: 90+ hosted chat models (DeepSeek, Qwen, Llama, Gemma, Mistral, plus hosted frontier models like Claude and Gemini), with dynamic model discovery, reasoning-effort control, vision, and a footer statusline showing DeepInfra monthly usage.

Fork of puetsua/pi-deepinfra — see Differences from upstream.

Install

Requires pi.

pi install npm:@cad0p/pi-deepinfra-provider@next

or from source:

pi install git:github.com/cad0p/pi-deepinfra

  1. Authenticate — either:

    • Interactive: inside pi, run /login deepinfra and paste your API key (from https://deepinfra.com/dash/api_keys). The key is stored in ~/.pi/agent/auth.json.
    • Environment: set DEEPINFRA_API_KEY in your shell.

    If a deepinfra api_key entry already exists in ~/.pi/agent/auth.json, it is picked up automatically.

  2. Restart pi (or /reload), then select a model:

    /model deepinfra/deepseek-ai/DeepSeek-V4-Flash-0731

What you get

Feature Notes
Model catalog Registered instantly with a curated fallback list — startup never blocks on the network. The live catalog from https://api.deepinfra.com/v1/openai/models is fetched lazily at session start, swapped in over the fallback, and cached in ~/.pi/agent/cache/deepinfra-models.json so later startups register the full catalog instantly (needed by clients that select a model right after launching pi, e.g. RPC mode) — no API key required (the catalog is public), so models are ready the moment a key is added. All chat-tagged models are registered (embeddings/image/TTS/STT excluded).
Streaming openai-completions API; SSE streaming with reasoning_content thinking deltas on reasoning models.
Thinking levels The full reasoning_effort vocabulary: off → none, and minimal/low/medium/high/xhigh/max pass through by name. DeepInfra validates one global enum for every chat model (an invalid value returns HTTP 422 listing none, minimal, low, medium, high, xhigh, max), so Ctrl+P / /model cycling offers every level including max.
Vision vision-tagged models accept image input.
Tool calling Standard OpenAI tools/tool_choice.
Cost tracking Input/output/cache prices from the catalog ($ per 1M tokens, cache_read_tokens when DeepInfra publishes it). Override per model in ~/.pi/agent/models.json if needed. Session tokens + cost are rendered by pi's own footer.
Footer statusline While a DeepInfra model is active: M: $12.34 / $50.00 — the DeepInfra monthly spend and the configured spending limit (via /payment/config), or M: $12.34 when no limit is set. Nothing else: pi's footer already shows session tokens and cost, so this extension does not repeat them. Refresh with /deepinfra-billing; disable with PI_DEEPINFRA_FOOTER=off.

Commands

  • /deepinfra-billing — refresh the monthly-usage figure in the footer (also refreshed on session start / model select; no background polling).

Configuration

  • API key: /login deepinfra, DEEPINFRA_API_KEY, or the auth.json credential.

  • Footer: PI_DEEPINFRA_FOOTER=off hides the statusline entirely.

  • Per-model overrides (prices, context, max tokens, thinking levels): standard pi ~/.pi/agent/models.json mechanism, e.g.:

    {
  "providers": {
    "deepinfra": {
      "modelOverrides": {
        "deepseek-ai/DeepSeek-V4.1-Flash-1017": {
          "contextWindow": 400000
        }
      }
    }
  }
}

    Note: models.json models entries replace the extension's models for that provider — set a full models list there if you use it.

Differences from upstream

  1. Thinking levels cover the whole enum. Upstream maps only off → none, so pi's getSupportedThinkingLevels() stops at high (pi only offers xhigh/max when thinkingLevelMap names them explicitly) and a max request is silently clamped to high. Verified against the live API: the validator accepts none/minimal/low/medium/high/xhigh/max for every chat model, so the map is complete.
  2. The footer no longer duplicates pi's. Upstream renders ↑in ↓out · $cost · M: $spent; pi's own footer already renders ↑in ↓out $cost, so the fork keeps only the monthly figure and drops the per-message recomputation.
  3. PI_DEEPINFRA_FOOTER=off to hide the statusline (upstream has no switch).

Everything else — catalog discovery, caching, auth, compat flags — is upstream's.

Development

pnpm install --frozen-lockfile
pnpm check   # tsc --noEmit (thinking.ts + its test; pi-importing files are excluded)
pnpm test    # node --test (offline)

See docs/verification.md for the CI gates and live-proof checklists.

License

MIT — see LICENSE. Original work © 2026 puetsua.