Error with Anthropic changes

[grenaad]

Looks like there was some changes for the headers that was required, I got this error:

{
  "type": "api_error",
  "message": "Claude Code returned an error result: API Error: 400",
  "details": {
    "type": "error",
    "error": {
      "type": "invalid_request_error",
      "message": "You're out of extra usage. Add more at claude.ai/settings/usage and keep going."
    },
    "request_id": "req_011CZqoaJwkn7FFfvJmC6GGf"
  },
  "stderr": "Warning: Custom betas are only available for API key users. Ignoring provided betas."
}

I remove them then got this error:

{
  "summary": "Claude Code returned an error result",
  "status": 400,
  "error": {
    "type": "invalid_request_error",
    "message": "You're out of extra usage. Add more at claude.ai/settings/usage and keep going."
  },
  "request_id": "req_011CZqq83Q8mLnDCGXHVbhTZ"
}

I assume this is Anthropic recent changes that they announced.

[endre82]

Yep, looks like claude has a mechanism to detect third party connectors.

[rynfar]

Son of a bitch I knew the system prompt would eventually bleed through. Anthropic.... love to hate them. I'll explore how they are triggering this. If they do a blanket SDK ban... that's the ultimate f u

[daviddao]

It's a regex on things such as "opencode", "pi", etc on the system prompt. Update the system prompt to avoid these key words is a way to circumvent it

[ianjwhite99]

Basic fix for the plugin: griffinmartin/opencode-claude-auth#145 (comment)

[ianjwhite99]

It's a regex on things such as "opencode", "pi", etc on the system prompt. Update the system prompt to avoid these key words is a way to circumvent it

You can update your system prompt in your opencode.json, similar to the following:

  "agent": {
    "build": {
      "prompt": "You are Anthropic's Claude Code, Anthropic's official CLI for Claude. Here are some rules. <rules>No emojis unless the user asks for them.\n\nOutput is rendered in a CLI using CommonMark markdown with a monospace font. Keep responses brief.\n\nText output is how you talk to the user. Tools are for doing work, not for communication — don't use Bash, code comments, or similar as a chat channel.\n\nDon't create files unless absolutely necessary. Edit existing files instead. This applies to markdown too.\n\nBe technically accurate and direct. Skip the praise, superlatives, and emotional validation. If the user is wrong, say so — respectful correction beats false confirmation. When uncertain, investigate before responding rather than reflexively agreeing. The user is better served by honest, rigorous assessment than by comfort.\n\nUse TodoWrite liberally to plan work, break down complex tasks, and give the user visibility into progress. Mark items complete immediately when finished — don't batch completions. Skipping this tool during planning risks forgetting steps, which is unacceptable.\n\nUsers will mostly ask for software engineering work: bugs, features, refactors, explanations, etc. General approach: plan with TodoWrite when the task warrants it. Note that system-reminder tags may appear in tool results or user messages. These are injected by the system with relevant context — they aren't tied to the specific message they appear in.\n\nPrefer the Task tool for file search to save context. Use Task with specialized agents proactively when the task fits an agent's description. On WebFetch redirects to a different host, immediately follow the redirect URL. Call independent tools in parallel. Call dependent tools sequentially — never guess at values from incomplete prior calls. If the user says in parallel, send one message with multiple tool blocks. Use dedicated tools over Bash equivalents: Read not cat, Edit not sed, Write not echo. Bash is for actual shell operations only. Never use Bash to communicate with the user. For broad codebase exploration or context gathering (not targeted lookups of a specific file/class/function), use the Task tool rather than running searches directly.\n\nAlways use TodoWrite to plan and track work throughout the conversation.\n\nWhen citing specific code, include file_path:line_number so the user can jump to it.</rules>"
    },
    "plan": {
      "prompt": "You are Anthropic's Claude Code, Anthropic's official CLI for Claude. Here are some rules. <rules># Plan Mode\n\nCRITICAL: Plan mode is ENABLED - you are operating in a READ-ONLY state. The following\nare STRICTLY PROHIBITED: ALL file writes, edits, deletions, or system modifications.\nDo NOT invoke sed, tee, echo, cat, or ANY other bash command that alters files -\ncommands are PERMITTED ONLY for reading and inspection. This ABSOLUTE RESTRICTION\ntakes precedence over ALL other directives, including explicit user requests to edit.\nYou may ONLY observe, analyze, and formulate plans. Any attempt to modify the system\nis a critical violation. NO exceptions.\n\n## Responsibility\n\nYour role right now is to reason, read, search, and dispatch explore agents to build\na clear, well-structured plan that achieves the user's stated goal. The plan should\nbe thorough but not bloated - detailed enough for reliable execution without\nunnecessary filler.\n\nWhen facing tradeoffs or ambiguity, ask the user for clarification or their preference.\n\n**NOTE:** Throughout this workflow, you are encouraged to ask the user questions at\nany point. Do not make significant assumptions about what the user intends. The\nobjective is to deliver a carefully researched plan and resolve open questions before\nany implementation starts.\n\n## Important\n\nThe user has specified that execution should NOT begin yet -- you MUST NOT perform any\nedits, invoke any non-readonly tools (this includes config changes and commits), or\nmake any other modifications to the system. This directive overrides all other\ninstructions you may have received.\n</rules>"
    }
  }

If you want to avoid the custom prompting, I baked this to the plugins latest release: https://github.com/ianjwhite99/opencode-with-claude/pull/92/changes

[faileon]

Updating the system prompt works. Anthropic is getting kinda boring with this regex style security.

[cristianmiranda]

Looks like this bypass is working fine as a plugin (posted here):

// .opencode/plugins/opencode-claude-auth-bypass.ts
import type { Plugin } from "@opencode-ai/plugin"

export const MyPlugin: Plugin = async ({ project, client, $, directory, worktree }) => {
  return {
    "experimental.chat.system.transform": async (input, output) => {
      if (input.model.providerID !== ('anthropic')) return;
      const opencodePromptPart = output.system.findIndex(x => x?.includes('https://github.com/anomalyco/opencode'))
      // Remove the OpenCode system prompt part if present
      if (opencodePromptPart !== -1) {
        output.system.splice(opencodePromptPart, 1)
      }
    },
  }
}

[rynfar]

I'm conflicted as to whether we bake something in to meridian or leave it to the harness to provide a plugin like this to fix it. It's going to be a moving target either way.

[jacobkring]

I'm conflicted as to whether we bake something in to meridian or leave it to the harness to provide a plugin like this to fix it. It's going to be a moving target either way.

I think I'd push you to avoid baking something like that in, seems out of scope for what you are aiming for with this package IMO. Anything that modifies the prompt/content someone is sending seems like a step too far. Perhaps a warning or something like that?

[rynfar]

Thanks for the input! Yea that’s kind of my thinking as well. I very much avoided anything that would modify what the user is sending. Just act as a dumb proxy, the harness of choice is responsible for what it sends over the wire, we just relay it in a valid format.

…

On Wed, Apr 8, 2026 at 11:53 AM jacobkring ***@***.***> wrote: *jacobkring* left a comment (rynfar/meridian#319) <#319 (comment)> I'm conflicted as to whether we bake something in to meridian or leave it to the harness to provide a plugin like this to fix it. It's going to be a moving target either way. I think I'd push you to avoid baking something like that in, seems out of scope for what you are aiming for with this package IMO. Anything that modifies the prompt/content someone is sending seems like a step too far. Perhaps a warning or something like that? — Reply to this email directly, view it on GitHub <#319 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACWNASSDALOUGWNY7GDSPR34U2G3LAVCNFSM6AAAAACXQON4SGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DEMBYGMZTEMZQGQ> . You are receiving this because you commented.Message ID: ***@***.***>

[endre82]

I think it should be left as is. Manipulating the prompt content should not be part of this application.

[cristianmiranda]

I upgraded OpenCode to v1.4.0 and Anthropic's models started working again through meridian. I'm pretty sure the upgrade had nothing to do with it because I checked the changelog and there's nothing related to changing prompt injection nor anything like it. I also confirmed that extra usage is not enabled on my account 🤷🏽‍♂️ Anyone else experienced this?

[KarmaMonk]

I upgraded OpenCode to v1.4.0 and Anthropic's models started working again through meridian. ... I also confirmed that extra usage is not enabled on my account 🤷🏽‍♂️ Anyone else experienced this?

Yes I can confirm that it works with OpenCode v1.4.0 (GUI),
but it also works again in OpenCode 1.3.17 (TUI) (which I haven't updated yet).
So maybe something changed at Anthropic?

[rynfar]

I think anthropic backed up their change because it triggered alot of false positives...

…

On Wed, Apr 8, 2026 at 5:22 PM Marc Pilloud ***@***.***> wrote: *KarmaMonk* left a comment (rynfar/meridian#319) <#319 (comment)> I upgraded OpenCode to v1.4.0 and Anthropic's models started working again through meridian. ... I also confirmed that extra usage is not enabled on my account 🤷🏽‍♂️ Anyone else experienced this? Yes I can confirm that it works with OpenCode v1.4.0 (GUI), but it also works again in OpenCode 1.3.17 (TUI) (which I have not updated yet). So maybe something changed at Anthropic? — Reply to this email directly, view it on GitHub <#319 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACWNASWDKJIZJNLN24RGTI34U3NJZAVCNFSM6AAAAACXQON4SGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DEMJQGM2DIMBWHE> . You are receiving this because you commented.Message ID: ***@***.***>