cohere-models

name	cohere-models
description	Update model metadata for the pi-cohere extension. Use when refreshing Cohere model lists, checking Cohere model availability, or syncing hardcoded model metadata with Cohere API and models.dev.

Update Cohere models

Update src/cohere/models.ts from live Cohere and models.dev data, not guesses.

Default behavior

Take initiative:

1. Check public pricing and limits in https://models.dev/api.json under provider cohere.
2. If COHERE_API_KEY is available, fetch https://api.cohere.com/v1/models to confirm API-visible chat models.
3. Remember: Cohere's model list endpoint requires an API key. Without one, use models.dev and documented/runtime evidence.
4. Probe https://api.cohere.ai/compatibility/v1/chat/completions only when availability is unclear. Never print the API key.
5. Update src/cohere/models.ts.
6. Run model tests plus typecheck and lint.
7. Create a changeset for user-visible metadata changes.
8. Commit only relevant files when asked.

Do not push.

Sources of truth

Use these in order:

1. https://models.dev/api.json provider cohere for pricing, context, output limits, modalities, and reasoning flags.
2. Cohere model list endpoint, when COHERE_API_KEY is available: https://api.cohere.com/v1/models.
3. Cohere compatibility endpoint runtime checks: https://api.cohere.ai/compatibility/v1/chat/completions.
4. Cohere docs: https://docs.cohere.com/docs/compatibility-api and https://docs.cohere.com/docs/models.
5. Existing hardcoded definitions for fields live sources do not expose.

Required checks

Run:

pnpm test -- src/cohere/models.test.ts
pnpm typecheck
pnpm lint

Field mapping

From models.dev:

• id -> id
• name -> name where appropriate, otherwise keep existing display formatting
• modalities.input containing image -> input: ["text", "image"]
• reasoning -> reasoning
• limit.context -> contextWindow when models.dev is the source for that model
• limit.output -> maxTokens
• cost.input -> cost.input
• cost.output -> cost.output

Provider compatibility defaults:

compat: {
  maxTokensField: "max_tokens",
  supportsUsageInStreaming: true,
}

Runtime probes

Use these only when needed. Never echo COHERE_API_KEY.

Check model list

curl -sS https://api.cohere.com/v1/models \
  -H "Authorization: bearer $COHERE_API_KEY" \
  -H 'accept: application/json'

Check compatibility chat

curl -sS https://api.cohere.ai/compatibility/v1/chat/completions \
  -H "Authorization: Bearer $COHERE_API_KEY" \
  -H 'Content-Type: application/json' \
  -d @- <<'JSON'
{
  "model": "command-a-03-2025",
  "messages": [{"role": "user", "content": "Reply exactly ok"}],
  "max_tokens": 5
}
JSON

Decision rules

• Add a model when models.dev or runtime confirms it is available and it belongs to Cohere chat models.
• Keep hardcoded known models that are callable even if absent from GET /v1/models, such as north-mini-code-1-0.
• Do not remove a model only because GET /v1/models omits it if runtime or models.dev says it is callable.
• Leave cost at zero only when no pricing source exposes a price.
• Prefer api.cohere.ai for compatibility endpoint docs and provider base URL.

Commit workflow

When asked to commit:

1. Run checks.
2. Check git status.
3. Stage only relevant files. Never use git add . or git add -A.
4. Use a concise conventional commit message.