Verify the Claude Code adapter against a live Claude Code install

[leofang2007-maker]

The Claude Code adapter (clients/claude-code/, spec 0005) is built to the documented Claude Code hook API and validated against a mock ingest server — but has not yet been run end-to-end on a real Claude Code install. The Qoder adapter, by contrast, is verified in a real deployment. See clients/README.md for the status table.

This issue tracks promoting the Claude Code adapter from doc-based to verified.

What still needs real-environment confirmation

• UserPromptSubmit stdin JSON field names/shape match on a current Claude Code version (session_id, prompt, cwd, transcript_path, prompt_id, permission_mode).
• Command hook actually fires and the prompt lands in the audit log end-to-end.
• SessionStart drain re-delivers spooled records after an outage.
• Managed enforcement: managed-settings.json with allowManagedHooksOnly: true at the OS managed path is honored, and a user cannot remove/disable the audit hook.
• Managed paths correct on macOS / Linux / Windows.
• Native HTTP-hook variant (type: http, allowedEnvVars) works and its payload is accepted.
• UserPromptSubmit timeout behavior (~30s) doesn't clip the hook.

If you hit a problem

Please comment with your Claude Code version, OS, whether you used the shell hook or the HTTP hook, and what you observed (hook didn't fire / wrong field / not enforced / etc.). That's the fastest path to a fix — and lets us mark the adapter verified.

Follow-ups tracked separately under #5: Cursor / Copilot / Codex adapters.