fix(metadata): Some providers use requestbody(body.metadata.user_id) to enable cache features Fixes #8195

[tisoz]

What does this PR do?

Fixes #8195

allow user add a options to enable cache features.

opencode.json

    "provider": {
        "anthropic": {
            "name": "fox.anthropic.claude",
            "options": {
                "baseURL": "https://code.newcli.com/claude/aws/v1",
                "apiKey": "sk-ant-xxxxx",
                "setCacheKey": true,   
                "enableMeta":true    // fix provider cache
            },
.....

How did you verify your code works?

clone the project and

bun run dev

[github-actions]

Hey! Your PR title Fix Some providers cache futures doesn't follow conventional commit format.

Please update it to start with one of:

• feat: or feat(scope): new feature
• fix: or fix(scope): bug fix
• docs: or docs(scope): documentation changes
• chore: or chore(scope): maintenance tasks
• refactor: or refactor(scope): code refactoring
• test: or test(scope): adding or updating tests

Where scope is the package name (e.g., app, desktop, opencode).

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

No duplicate PRs found

[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.

[rektide]

user_session_id would be more clear. or, ideally, let's make user_id just be user_id and concatenate our values at the site of use.

[tisoz]

Currently, the purpose of this combination is to simulate the behavior in the Claude Code CLI, which is why this approach is adopted

[tisoz]

If it's a variable name, it could indeed be optimized to user_session_id to express the meaning more clearly

[rektide]

Perhaps "the train has already left the station", but I would prefer not adding opencode specific headers that give away where my traffic is coming from like this. Also, should we be using a metadata

[tisoz]

This addresses the issue where many model providers cannot properly enable caching functionality. I attempted to add parameters like user_id in the metadata.
However, the model providers cannot recognize them correctly, including AWS Claude Code and several third-party provided Claude models.

[70akaline]

Perhaps "the train has already left the station", but I would prefer not adding opencode specific headers that give away where my traffic is coming from like this. Also, should we be using a metadata

I think some API endpoint like AWS bedrock/Azure are alao require this fix to get better prompt cache support

[tisoz]

yeah, I ran into this issue while using AWS

[tisoz]

Change Description:

1. llm.ts:
   Check if provider.options.enableMeta is enabled, and only then proceed with metadata information passing (due to incomplete context parameters in provider.ts)

2. The core implementation of this feature requires manually enabling the provider.options.enableMeta configuration item to ensure third-party providers can properly implement the caching functionality for conversations.

3. The request parameters originate from simulating the native parameters of the Claude Code CLI, so this is configured as an optional setting, defaulting to off. Regarding privacy, there should be no issues.

This submission is primarily aimed at reducing tool conversation costs. Previously, without caching features, contextual conversations consumed a large number of tokens, leaving no choice but to adopt this method to maintain compatibility with the model's functionality.

[rekram1-node]

ideally we make everything just work without additional configs, it sounds like azure and bedrock are the ones that require this? Can you point to any docs that show it?

[tisoz]

ideally we make everything just work without additional configs, it sounds like azure and bedrock are the ones that require this? Can you point to any docs that show it?理想情况下所有功能都应在无需额外配置的情况下直接运行，听起来只有 Azure 和 Bedrock 需要这个设置？能否提供相关说明文档？

anthropic api docs:
https://platform.claude.com/docs/en/build-with-claude/prompt-caching

However, it should be noted that the documentation does not specify the exact function of metadata.user_id. I discovered through reverse engineering that claude code adds this parameter in requests.

I then enabled this parameter in opencode and tested it on the sonnet 4.5 and opus 4.5 models, which successfully implemented the caching functionality.

It might be that AWS has special caching identification methods for these models, but clearly, we all need to enable this feature to reduce our token input and output

[rekram1-node]

Oh if ur talking specifically about bedrock, there was a separate issue and has been addressed:
ca5e85d

[tisoz]

Oh if ur talking specifically about bedrock, there was a separate issue and has been addressed:哦，如果你特指 Bedrock 的问题，之前确实有个独立的问题，现在已经解决了： ca5e85d

I noticed this commit, but it's not based on the same principle as our current cache mechanism identification. I tried simulating these parameters, but the third-party endpoint I'm using didn't successfully enable it

[rekram1-node]

but the third-party endpoint I'm using didn't successfully enable it

Well if you are using a third party endpoint then it is separate, could u provide docs for ur other provider instead? I don't wanna add random headers and body fields unless we can have good reason for it.

This did fix caching for bedrock, at least directly

[tisoz]

but the third-party endpoint I'm using didn't successfully enable it但我使用的第三方端点未能成功启用它

Well if you are using a third party endpoint then it is separate, could u provide docs for ur other provider instead? I don't wanna add random headers and body fields unless we can have good reason for it.如果你使用的是第三方端点，那情况就不同了，能否提供你其他供应商的文档？我不想随意添加请求头和正文字段，除非我们有充分的理由。

This did fix caching for bedrock, at least directly这确实解决了 Bedrock 的缓存问题，至少是直接解决了

Currently, because the platform's scale is not that large, only a small number of developers will notice this issue.

They offer relatively low-cost endpoint channels, but they lack the capability to provide complete documentation. Therefore, I only have some open-source projects from fellow enthusiasts to address this issue.

Next, I will provide some open-source projects for your reference:
https://github.com/1034378361/foxcode-cache-proxy
https://github.com/[x7yue/opencode-foxcode-aws-cache](https://github.com/x7yue/opencode-foxcode-aws-cache)

So I think this should be a necessary feature for simulating the Claude CLI.

If the addition of the config.provider.options.enableMeta option is a carefully considered feature

I can consider modifying it into a more general form to allow users to add metadata themselves for reliable iteration of the caching functionality