Investigate: can Auto mode's permission classifier be routed through the proxy (and how)?

[cryptiklemur]

Goal

I'd like Claude Code's Auto mode to work through better-ccflare — including
its permission classifier traffic. I'm not sure that's possible, so this is to
investigate whether it can be done and, if so, how.

Background

Auto mode (released 2026-03-24) runs without routine permission prompts, but per
Anthropic's docs "every shell command or network action first passes through a
separate safety classifier — a second model that vets each action." That
classifier is an additional model call.

For the proxy to fully support Auto mode, that classifier call ideally routes
through better-ccflare too, so it benefits from the same load balancing, account
fallback, and analytics as normal turns. Right now I don't know whether that's
even where the call goes.

What to figure out

• Does it work at all today? With Claude Code pointed at the proxy via
  ANTHROPIC_BASE_URL, does Auto mode function, and does the classifier
  call go through the proxy or bypass it straight to Anthropic?
• If it bypasses — can we make it route through? Is the classifier
  endpoint controlled by the same base URL / env config, or is it hardcoded
  / separate? Is there any knob (env var, header, config) that points it at
  the proxy?
• If it routes through — does ccflare handle it correctly? What model is
  it, does it pick up account selection (session / leastUsed /
  session-affinity), and does its usage show up in analytics rather than
  getting dropped or mis-attributed?
• Rate-limit impact: if classifier calls do flow through, how much extra
  volume does Auto mode add per account, and does that change which strategy
  is advisable?
• What (if anything) would the proxy need to change to support routing
  the classifier call?

Notes

I haven't confirmed where the classifier call actually goes — that's the first
thing to pin down, since everything else depends on it. If it turns out the
classifier is server-side / out-of-band and genuinely can't be pointed at a
proxy, that's a valid (if disappointing) answer to capture here.

[github-actions]

🤖 Issue Triage

Severity: medium

Analysis:
This is an investigation issue into a specific feature of Claude Code (Auto mode) and its compatibility with the better-ccflare proxy. The author is asking whether the secondary 'safety classifier' calls made by Auto mode can be routed through the proxy. This requires research into the Claude Code client behavior to determine if these calls respect the ANTHROPIC_BASE_URL environment variable or if they bypass the proxy. Depending on the findings, this may result in a configuration documentation update, a new feature request, or a confirmation that it is not supported.

---

Thanks for opening this investigation @cryptiklemur. This is a great question regarding the compatibility of Claude Code's Auto mode with better-ccflare.

To help the community investigate this, could you please provide the following details:

1. better-ccflare version (e.g., v2.1.0, or commit hash if built from source)
2. Installation method (npm, bun, pre-compiled binary, built from source, Docker/Docker Compose)
3. Operating system and architecture

Additionally, if you have already attempted to use Auto mode with the proxy, please share your observed behavior (e.g., did Auto mode work? Did you see any errors in the proxy logs?).

We will keep this issue open to track the findings on whether the classifier endpoint respects the standard proxy configuration.

---

This automated triage was performed by the better-ccflare Issue Triage Agent using glm-4.7.

[cryptiklemur]

Looks as if this works now too. Please verify and close if it works for you as well