Prompt-Level Detection? github.com/anomalyco/opencode Causes Billing Path Switch in Claude Max

[micro-xl]

Question

(Not question. Just sharing)

Sharing a reproducible observation that may be relevant to this project.

When using Anthropic’s Messages API with a Claude Max plan, including the string github.com/anomalyco/opencode (or its previous name github.com/sst/opencode) in the system prompt consistently correlates with requests being routed to “Extra Usage” billing—even when Max plan credits are still available.

Details

In my testing, prompt content alone appears sufficient to influence how billing is applied. When the system prompt contains a reference to this repository(github.com/anomalyco/opencode or github.com/sst/opencode), requests that would normally be covered by the Max plan instead require Extra Usage credits.

Reproduction Context

1. Clone the repository to:
   $HOME/github.com/anomalyco/opencode
2. Run claude (CLI) to summarize recent changes
3. Ensure Extra Usage is disabled (Max plan only)
4. Observe: request fails with 400 (Extra Usage required)
5. Remove or alter the repository string in the system prompt
6. Observe: same request succeeds under Max plan

I also verified via a local proxy that the only meaningful difference between requests was the presence or absence of the repository string (including partial matches).

I was aware that opencode couldn’t use Anthropic’s subscription plans, but it was unexpected that I couldn’t even modify the project using Claude Code under a subscription.

[github-actions]

Thanks for updating the issue! It now follows our contributing guidelines. We appreciate you providing the reproduction context and details.

[CGustavoFlores]

We can reproduce this locally and the behavior matches this issue very closely.

Environment:

• opencode version: 1.4.3
• Anthropic auth plugin: @ex-machina/opencode-anthropic-auth@1.5.1
• OS: Linux
• Auth method: Claude Pro/Max OAuth

What we verified:

1. claude CLI works fine on the same machine:

   claude -p "Respond with OK only"

   returns:

   OK
   

2. OpenCode has a valid Anthropic OAuth credential:

   opencode providers list

   shows Anthropic oauth.
3. We re-did the Anthropic login flow from OpenCode (Claude Pro/Max) and tested again.
4. The minimal OpenCode request still fails:

   opencode run --print-logs --model anthropic/claude-sonnet-4-5 "Respond with OK only"

   with:

   You're out of extra usage. Add more at claude.ai/settings/usage and keep going.
   

Important detail: this is happening on a minimal prompt, not a genuinely long-context request. So this looks closer to prompt / identity / billing-path detection than to the separate long-context error classification issue.

Another suspicious detail from the OpenCode log: the failed request still shows the full OpenCode-flavored system prompt under system[], including references like github.com/anomalyco/opencode / OpenCode docs text. If Anthropic is keying off prompt-level markers, that would explain why Claude CLI works while OpenCode does not on the same account.

So from our side, we can rule out:

• missing Anthropic credential
• stale login alone
• old plugin version alone
• Claude Max subscription being generally broken

Happy to provide the exact --print-logs output if useful, but the short version is: same machine, same account, Claude CLI succeeds, OpenCode fails with extra usage on a minimal request.

[ariane-emory]

Anthropic plans are not officially supported, making this issue totally off-topic. Recommend closing.