---
id: SMH-RESEARCH-RSCH0007
type: research
title: "Provider Normalization Research"
---

# Provider Normalization Research

Purpose: check whether Pi generated provider data, catwalk provider configs, and live
provider model APIs are similar enough to normalize for Smith provider suggestions.

## Sources Checked

- Pi generated registry clone: `https://github.com/earendil-works/pi-mono.git`
  at `3eb0027`, file `packages/ai/src/models.generated.ts`.
- Catwalk clone: `https://github.com/charmbracelet/catwalk.git` at `2fe397e`,
  files `internal/providers/configs/zai.json` and
  `internal/providers/configs/minimax.json`.
- Live Z.AI API probes using `ZAI_API_KEY` from environment. Key was not printed.
- Live MiniMax API probes using `MINIMAX_API_KEY` from environment. Key was not
  printed.

## Shape Comparison

Observed common normalized fields:

- provider id,
- provider name,
- API/base URL,
- API family (`openai-completions` / OpenAI-compatible, `anthropic-messages` /
  Anthropic-compatible),
- model id,
- model display name,
- context window,
- max output/default max tokens,
- input modality / attachment support,
- reasoning support,
- per-token or per-million-token costs,
- cache read/write costs when known.

Observed shape differences:

| Concept | Pi generated registry | catwalk config |
|---|---|---|
| API family | `api` values like `openai-completions`, `anthropic-messages` | `type` values like `openai-compat`, `anthropic` |
| Base URL | per model `baseUrl` | provider `api_endpoint` |
| Context | `contextWindow` | `context_window` |
| Max output | `maxTokens` | `default_max_tokens` |
| Reasoning | `reasoning` | `can_reason` |
| Modalities | `input: ["text", "image"]` | `supports_attachments` |
| Cost | `cost.input/output/cacheRead/cacheWrite` | `cost_per_1m_*` |
| Compat quirks | `compat` object | mostly absent |
| Defaults | not primary in model entries | `default_large_model_id`, `default_small_model_id` |

## Z.AI Observations

Pi direct `zai` entries observed:

- `glm-4.5-air`
- `glm-4.7`
- `glm-5-turbo`
- `glm-5.1`
- `glm-5v-turbo`

Catwalk `zai` entries observed:

- `glm-5.1`
- `glm-5-turbo`
- `glm-5`
- `glm-4.7`
- `glm-4.7-flash`
- `glm-4.6`
- `glm-4.6v`
- `glm-4.5`
- `glm-4.5-air`
- `glm-4.5v`

Live API probes:

```text
zai-coding-models: 200 → glm-4.5, glm-4.5-air, glm-4.6, glm-4.7, glm-5, glm-5-turbo, glm-5.1
zai-general-models: 200 → glm-4.5, glm-4.5-air, glm-4.6, glm-4.7, glm-5, glm-5-turbo, glm-5.1
zai-coding-chat-smoke: 200
```

Observed drift:

- Pi has `glm-5v-turbo`; live model listing did not show it.
- Catwalk has `glm-4.7-flash`; live model listing did not show it.
- Catwalk has pricing for Z.AI; Pi direct Z.AI costs are zero in generated data.
- Pi has provider-specific `compat.thinkingFormat = "zai"` and `zaiToolStream`
  flags; catwalk does not.

## MiniMax Observations

Pi direct `minimax` entries observed:

- `MiniMax-M2.7`
- `MiniMax-M2.7-highspeed`

Catwalk `minimax` entries observed:

- `MiniMax-M2.7`
- `MiniMax-M2.7-highspeed`
- `MiniMax-M2.5`
- `MiniMax-M2.5-highspeed`
- `MiniMax-M2.1`
- `MiniMax-M2.1-highspeed`
- `MiniMax-M2`

Live API probes:

```text
minimax-openai-models: 200 → MiniMax-M2.7, MiniMax-M2.7-highspeed, MiniMax-M2.5, MiniMax-M2.5-highspeed, MiniMax-M2.1, MiniMax-M2.1-highspeed, MiniMax-M2
minimax-anthropic-models: 200 → MiniMax-M2.7, MiniMax-M2.7-highspeed, MiniMax-M2.5, MiniMax-M2.5-highspeed, MiniMax-M2.1, MiniMax-M2.1-highspeed, MiniMax-M2
minimax-anthropic-smoke: 200
```

Observed drift:

- Catwalk coverage matches live model listing better than Pi direct MiniMax.
- Pi has Anthropic-compatible base URL and cache/cost fields for direct MiniMax.
- Catwalk context values are rounded (`200000`) while Pi uses `204800`.
- Live model listings confirm model IDs but not costs/context/compat behavior.

## Candidate Conclusion

Pi generated data and catwalk configs are similar enough to normalize into a
Smith suggestion schema for common model metadata.

They are not similar enough for silent replacement. Required review points:

- model coverage drift,
- context/max-token conflicts,
- pricing conflicts or missing pricing,
- provider-specific compat quirks missing from catwalk,
- live model endpoints listing IDs but not enough metadata for correctness.

Best fit: Pi as primary, catwalk as gap-fill, live provider APIs as optional
ID-availability evidence only. Current checked-in `providers.json` remains the
runtime authority.