pi-extension-wandb

A pi coding agent extension that adds Weights & Biases Inference as a model provider. Model list is discovered dynamically from W&B's OpenAI-compatible /v1/models endpoint, so new models show up without code changes.

Install

Get an API key from https://wandb.ai/authorize, then:

export WANDB_API_KEY=...               # required
export WANDB_PROJECT=team/project      # optional, recommended

pi install https://github.com/kirangadhave/pi-extension-wandb

Verify:

pi --list-models | grep wandb

Use

pi --provider wandb --model deepseek-ai/DeepSeek-V3.1 "hello"

Configuration

| Env var | Required | Notes | | --- | --- | --- | | WANDB_API_KEY | yes | From https://wandb.ai/authorize. If unset, the provider is silently skipped (no wandb/* models appear in --list-models). | | WANDB_PROJECT | no* | Sent as the OpenAI-Project header. Format: team/project. If you belong to multiple W&B teams, set this explicitly — otherwise W&B picks a default team for attribution, which can land usage on the wrong team. Single-team users can safely omit. | | WANDB_DEBUG | no | 1 enables stderr logging from the extension (fetch errors, cache hits, etc.). | | WANDB_NO_CACHE | no | 1 bypasses the model-list cache and fetches fresh on every pi startup. |

How it works

On every pi startup the extension:

1. Reads ~/.cache/pi-extension-wandb/models.json if it's fresh (< 1 hour old).
2. Otherwise fetches https://api.inference.wandb.ai/v1/models (5s timeout) and writes the result to the cache.
3. Registers each returned model under the wandb provider with api: "openai-completions".

If the fetch fails and there is no usable cache, the provider is not registered — pi --list-models will not show wandb models, and pi falls back to its other configured providers cleanly.

Known limitations

• Context windows are guesses. W&B's /v1/models doesn't expose limits. The extension hardcodes known windows for a handful of popular models (KNOWN_CONTEXT_WINDOWS in index.ts); everything else gets a default of 128K. PRs to extend the map are welcome.
• Reasoning detection is heuristic. Substring match against the model id (REASONING_MODEL_PATTERNS). Edit the list for new families.
• No cost tracking. All models report zero cost. W&B publishes per-token pricing — PRs to encode it are welcome.
• No provider-specific compat flags. Some reasoning models (Qwen3 thinking, gpt-oss, DeepSeek-R1) may need provider-specific quirks for thinking mode. If you hit a problem, open an issue with the model id and the exact failure.

Contributing

PRs welcome. The only file you typically need to touch is index.ts:

• KNOWN_CONTEXT_WINDOWS — add accurate context sizes for new models.
• REASONING_MODEL_PATTERNS — add model family substrings that support reasoning/thinking.

For editor type-checking, after cloning:

pnpm install
pnpm check

pi loads index.ts directly at runtime (no build step).

License

MIT — see LICENSE.