Add support for Claude Code Token

[egenvall]

Summary

• Add Claude Code OAuth token (sk-ant-oat*) support, enabling use of Anthropic subscription-based authentication alongside standard API keys
• Introduce a new src/llm/anthropic/ module handling auth routing, tool name normalization, prompt caching, and request building
• Add per-slot adaptive thinking effort configuration (auto/max/high/medium/low) for Claude 4.6 models
• Add frontend UI for thinking effort when a 4.6 model is selected

Details

Inspired & Ported from PI-AI OAuth token authentication.

Anthropic API keys prefixes are auto-detected and route differently depending on whether it stems from a claude setup-token

Tool name normalization - When using a claude setup-token, tool names are mapped to Claude Code canonical casing on the way out, and reverse-mapped back to original names on responses.

Prompt caching - System prompt blocks and the last tool definition get cache_control attached

Adaptive thinking - 4.6-generation models (opus-4-6, sonnet-4-6) get thinking: { type: "adaptive" } with configurable effort per routing slot. Opus defaults to max, others to high.

Test plan

• Run claude setup-token and use that as your API key for Anthropic.
• Verify standard API key auth (sk-ant-api*) still works as before
• Verify thinking effort dropdown only appears for 4.6 models in the admin UI
• Run cargo test — new unit tests cover auth detection, cache control, tool normalization, and reverse mapping

Known issues

• Different thinking levels for the same model needs a refactor to pipe per route.

[jamiepine]

Maybe not the best timing on this, given Anthropic have cracked down. Gonna sit on this one, curious what others think?

[insecurejezza]

I say give us the option at our own risk. Anthropic is intentionally vague about whether this is allowed or not.

[egenvall]

Maybe not the best timing on this, given Anthropic have cracked down. Gonna sit on this one, curious what others think?

It's technically against TOS but after their statement yesterday they went out on X and said the crackdown was misunderstood and due to "docs cleanup".

In my opinion, let people use it at their own risk. The reward greatly outweighs the risk as API pricing is simply not reasonable for individuals who want to use 4.6

[egenvall]

@jamiepine if we're aligned I can rebase to get it in.

[robertocarvajal]

why not use opencode to spawn claude instances? that would be fine with TOS. Basically run a claude -p ralph loop on it.

[ricorna]

Don't bother. They straight up changed something on their end yesterday evening. Anthropic only accepts OAuth based keys from Claude code now as far as I can tell?

Anyone still running an agent on their sub? I had to go to GLM on my OpenClaw.

[egenvall]

Don't bother. They straight up changed something on their end yesterday evening. Anthropic only accepts OAuth based keys from Claude code now as far as I can tell?

Anyone still running an agent on their sub? I had to go to GLM on my OpenClaw.

I've been running on my sub all day 🤷