[iNicholasBE]

Why

JavaClaw's Claude Code support read the OAuth token straight out of the macOS Keychain (or ~/.claude/.credentials.json) and called the API itself, impersonating Claude Code with its beta headers and system-prompt prefix. Our own onboarding page warned that this "violates the terms of your subscription and may result in account suspension".

What

This PR replaces that with a port of the Hermes Claude Subscription DirectSDK provider (NousResearch/hermes-official-plugins, MIT). JavaClaw never reads or sends the subscription credential. Everything goes through the official claude CLI, which handles login, token refresh and authentication itself.

How it works (new providers/claude-code module, provider type claude-code):

• One CLI process per call: each model call starts a fresh claude -p process in stream-json mode. Native tools, skills, settings, memory and compaction are off, and --max-turns 1 limits it to one turn.
• Replay: the conversation is replayed as stream-json frames, and Claude Code generates exactly one assistant turn.
• JavaClaw keeps the agent loop: tool calls come back as normal Spring AI tool calls, so JavaClaw's own ToolCallingAdvisor runs the tools.
• Single-request relay: a request-scoped loopback AdmissionRelay is Claude Code's ANTHROPIC_BASE_URL. It forwards only the first Messages request upstream and denies Claude Code's own retries locally, so one JavaClaw call can never become several billed requests.
• Prompt-cache hygiene: RelayPayloads keeps the cached prefix stable across tool rounds. It fails open, ported from the Hermes admission transforms.
• Signed thinking replay: Claude Code's native assistant messages, signed thinking included, ride along in assistant-message metadata so unchanged turns replay byte for byte.
• Fail closed: inherited ANTHROPIC_API_KEY / ANTHROPIC_BASE_URL / Bedrock / Vertex overrides are refused rather than silently billing an API key, and a custom base URL is rejected.

Deviations from Hermes:

• Hermes advertises the tool inventory through a Python stdio MCP process. Here the relay serves it over loopback HTTP, which avoids a second process per call; it's verified that the CLI connects (initialize, tools/list).
• Not ported: request gzip, HTTPS proxy support in the relay, the live model picker, /usage plan limits, the telemetry setting, and reasoning-effort / structured-output passthrough.

Onboarding:

• AgentOnboardingProvider drops systemWideToken() and gains readinessProblem(), supportsBaseUrl() and models().
• The "Claude subscription" step checks claude auth status before saving, hides the base URL field, and shows the models as a dropdown. Full ids are stored, and a saved model that isn't in the list stays selectable.

Bug fix found while testing:

• The in-process restart after onboarding re-read classpath:application.private.yaml, which is a stale build copy when running from Gradle or the IDE. That left the app on the "no model configured" placeholder, which fails with streaming is not supported.
• application.yaml now also imports the file ConfigurationManager actually writes, so the restart picks up the new config.

Migration: a config that still holds api-key: <claude-code-bearer-token> under provider: anthropic now fails with a clear message. Switch it to provider: claude-code and remove the api-key.

Testing

• ClaudeCodeChatModelIntegrationTest drives the real claude CLI (2.1.287) against a fake loopback Anthropic upstream: no credentials, no inference. It covers text, tool calls, history replay with signed thinking, the full ChatClient + ToolCallingAdvisor loop (blocking and streaming), upstream error mapping (429), truncation and cancellation. It's skipped when claude isn't installed.
• RelayPayloadsTest holds 29 cases ported from the Hermes admission tests. ClaudeCodeRequestTest, SseCaptureTest, ClaudeCodeCliTest, ClaudeCodeModelsTest and three new OnboardingControllerTest cases cover the rest.
• Live: a tool round trip on a Max subscription through ChatClient + ToolCallingAdvisor worked. Onboarding and web chat were also tested by hand.
• Results: claude-code 67, anthropic 2 and app 49 tests pass, including the ArchUnit provider-independence rules.

Note: when Gradle runs on JDK 27, :ui:vaadinBuildFrontend fails with "Unsupported class file major version 71". That is unrelated to this PR; run Gradle on JDK 25.

[arefbehboudi]

Thanks @iNicholasBE, That's what I really need it 😄 💯