Bug: agent maxTokens forwarded raw (camelCase) instead of translated to max_tokens for custom @ai-sdk/openai-compatible providers #40151

Summary

When an agent config sets maxTokens, OpenCode forwards it raw and untranslated (literal camelCase maxTokens) into the outbound JSON request body for a custom @ai-sdk/openai-compatible provider. The upstream endpoint (Azure AI Foundry / Fireworks-hosted models) rejects this with a strict-schema validation error, since it expects the OpenAI-standard snake_case max_tokens field instead.

This affects any custom npm: "@ai-sdk/openai-compatible" provider pointed at an endpoint with strict request validation (e.g. Azure AI Model Inference API routes).

Environment

• OpenCode version: 1.18.11 (also reproduced conceptually against v2 beta docs/behavior, not yet independently verified there)
• OS: macOS (darwin)
• Provider: Custom @ai-sdk/openai-compatible provider pointed at Azure AI Foundry (Fireworks-hosted models: Kimi K3, DeepSeek V4, etc.)

Repro config (opencode.json, abbreviated)

{
  "agent": {
    "generalist": {
      "model": "azurefoundry/FW-Kimi-K3",
      "maxTokens": 12000
    }
  },
  "provider": {
    "azurefoundry": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "apiKey": "{env:AZURE_FOUNDRY_API_KEY}",
        "baseURL": "https://<resource>.services.ai.azure.com/models",
        "queryParams": { "api-version": "2024-05-01-preview" },
        "headers": { "api-key": "{env:AZURE_FOUNDRY_API_KEY}" }
      },
      "models": {
        "FW-Kimi-K3": {
          "modelID": "fw-kimi-k3",
          "name": "Kimi K3 (Microsoft Foundry)"
        }
      }
    }
  }
}

Steps to reproduce

opencode run --model azurefoundry/FW-Kimi-K3 "Say OK"

Actual result

Error: Extra inputs are not permitted, field: 'maxTokens', value: 12000

Expected result

The request should succeed, with maxTokens translated to the provider's expected field name (max_tokens) — or simply not being forwarded literally into the request body at all for providers/packages that don't recognize the raw config key.

Isolation / verification performed

I verified this is an OpenCode-side bug, not an endpoint/auth issue, by sending the raw HTTP requests directly with curl:

Works (no max_tokens field):

curl --url 'https://<resource>.services.ai.azure.com/models/chat/completions?api-version=2024-05-01-preview' \
  -H 'content-type: application/json' \
  -H "api-key: $AZURE_FOUNDRY_API_KEY" \
  --data-raw '{"model":"fw-kimi-k3","stream":false,"messages":[{"role":"user","content":"Say OK"}]}'
# -> 200 OK

Works (snake_case max_tokens):

curl ... --data-raw '{"model":"fw-kimi-k3","stream":false,"max_tokens":12000,"messages":[...]}'
# -> 200 OK

Fails (camelCase maxTokens — matches what OpenCode sends):

curl ... --data-raw '{"model":"fw-kimi-k3","stream":false,"maxTokens":12000,"messages":[...]}'
# -> 400 {"error":{"message":"Extra inputs are not permitted, field: 'maxTokens', value: 12000"}}

This confirms OpenCode is sending the literal maxTokens key (from the agent's config field name) directly in the request body, rather than mapping it to the AI-SDK/OpenAI-standard max_tokens parameter, for this provider package path.

Workaround

Removing the agent-level maxTokens field entirely allows the request to succeed. This is not scoped per-provider/per-model, so it removes the output-token cap globally for the agent across all models, which is not ideal.

Suggested fix

Translate agent/model maxTokens config into the standard max_tokens (or package-appropriate) field name before merging into the outbound request body for @ai-sdk/openai-compatible (and similar) provider packages, instead of forwarding the raw config key verbatim.

Related issues that may share root cause:

• cloudflare-ai-gateway: Azure GPT-5.x models fail (max_tokens vs max_completion_tokens); per-model compat not honored #32776 (Azure GPT-5.x max_tokens vs max_completion_tokens not honored per-model for custom openai-compatible providers)
• Bug: incorrect max_tokens calculation when "thinking" is enabled & "budgetTokens" is set for Anthropic models #2055 (max_tokens calculation issue for Anthropic thinking budget)