fix: use max_completion_tokens for Azure reasoning models

[junmediatek]

Fixes Azure reasoning models (GPT-5, o1) by using max_completion_tokens instead of max_tokens.

[junmediatek]

@rekram1-node Could you help to review this PR for #5421

[rekram1-node]

/review

[github-actions]

Suggestion: The nesting here could be reduced by using early returns/guards. This is a stylistic suggestion - the current code is functionally correct.

Suggested change

	// @ts-ignore see here: https://github.com/oven-sh/bun/issues/16682
	// Reasoning models require max_completion_tokens instead of max_tokens
	// when using openai-compatible provider
	// Check if model has reasoning capability OR has reasoningEffort option
	const isReasoningModel =
	model.capabilities.reasoning || (model.options && "reasoningEffort" in model.options)
	const shouldTransformMaxTokens =
	model.api.npm === "@ai-sdk/azure" &&
	isReasoningModel &&
	opts.body &&
	typeof opts.body === "string"
	if (shouldTransformMaxTokens) {
	try {
	const body = JSON.parse(opts.body)
	if (body.max_tokens !== undefined) {
	body.max_completion_tokens = body.max_tokens
	delete body.max_tokens
	opts.body = JSON.stringify(body)
	}
	} catch {
	// Ignore JSON parse errors
	}
	}

This flattens one level of nesting by combining the conditions into a single check. Also note: using catch without a parameter (instead of catch (e)) is cleaner when you do not use the error.

[rekram1-node]

only condition here should be:
model.capabilities.reasoning

[junmediatek]

I have updated it, please help to review it 7b5a00e

[junmediatek]

@rekram1-node is anything wrong with the new patch 7b5a00e

[rekram1-node]

@junmediatek feel free to resolve any bot comments that you addressed or want to ignore

[rekram1-node]

Also, im not entirely sure if this is the perfect fix and I need to test something w/ it so in the meantime (without requiring code changes from us, you should be able to add this:

import { Plugin } from "@opencode-ai/plugin"

export const AzurePatch: Plugin = async (ctx) => {
  return {
    auth: {
      provider: "gaisf-azure",
      loader: async (getAuth, provider) => {
        return {
          async fetch(input, init) {
            const opts = init ?? {}
            if (opts.body && typeof opts.body === "string") {
              try {
                const body = JSON.parse(opts.body)
                if (body.max_tokens !== undefined) {
                  body.max_completion_tokens = body.max_tokens
                  delete body.max_tokens
                  opts.body = JSON.stringify(body)
                }
              } catch (e) {}
            }
            return fetch(input, {
              ...opts,
              timeout: false,
            })
          },
        }
      },
    },
  }
}

to ~/.config/opencode/plugin/max_completion_tokens.ts

And assuming your azure proxy provider is still called gaisf-azure like it was in your config you showed me, then this should function as expected

[rekram1-node]

You can completely override fetch implementations for any provider via plugins and I just wanted to demonstrate that for you

[junmediatek]

Also, im not entirely sure if this is the perfect fix and I need to test something w/ it so in the meantime (without requiring code changes from us, you should be able to add this:

import { Plugin } from "@opencode-ai/plugin"

export const AzurePatch: Plugin = async (ctx) => {
  return {
    auth: {
      provider: "gaisf-azure",
      loader: async (getAuth, provider) => {
        return {
          async fetch(input, init) {
            const opts = init ?? {}
            if (opts.body && typeof opts.body === "string") {
              try {
                const body = JSON.parse(opts.body)
                if (body.max_tokens !== undefined) {
                  body.max_completion_tokens = body.max_tokens
                  delete body.max_tokens
                  opts.body = JSON.stringify(body)
                }
              } catch (e) {}
            }
            return fetch(input, {
              ...opts,
              timeout: false,
            })
          },
        }
      },
    },
  }
}

to ~/.config/opencode/plugin/max_completion_tokens.ts

And assuming your azure proxy provider is still called gaisf-azure like it was in your config you showed me, then this should function as expected

@rekram1-node

That is a good idea, I will try it. However, the standard solution is to modify the code,
so I still need your help to do more testing. If you encounter any other issues, feel free to contact me at any time.

[junmediatek]

hi @rekram1-node

This issue can be resolved by the plugin, However, the standard solution is to modify the code, because the API reference in the openai API doc.
Additionally, how is the testing progress for the solution to this issue?

[junmediatek]

hi @rekram1-node

This patch can be merged?

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[8-unlimit]

This solved my problem, thank you.