Bug Report

Version: OpenClaw 2026.2.26 (bc50708)

Summary

When the Anthropic subscription (OAuth) hits rate limits, the configured API key fallback (anthropic-api/claude-opus-4-6) is never triggered. Instead, the gateway immediately returns All models failed (1) — meaning only 1 model was attempted despite a fallback being configured.

Config

{
  "primary": "anthropic/claude-opus-4-6",
  "fallbacks": ["anthropic-api/claude-opus-4-6"]
}

Auth profiles:

• anthropic:default → provider: anthropic, mode: token (OAuth/subscription)
• anthropic:api → provider: anthropic-api, mode: token (API key)

Error from gateway.log

Error: All models failed (1): anthropic/claude-opus-4-6: Provider anthropic is in cooldown (all profiles unavailable) (rate_limit)

Note: (1) — only 1 model tried, fallback never attempted.

Expected Behavior

When anthropic/claude-opus-4-6 hits a rate limit, the gateway should attempt anthropic-api/claude-opus-4-6 (API key, separate billing account) before failing.

Actual Behavior

Gateway puts the entire anthropic provider in cooldown and fails immediately. It appears anthropic and anthropic-api share the same cooldown bucket despite being configured as separate providers — so when one is rate limited, both are blocked.

Impact

The subscription and API key are on separate Anthropic accounts with independent rate limits. When the subscription is exhausted, the API key should absorb the load seamlessly. This defeats the entire purpose of configuring a fallback.

Workaround

None currently. Both providers fail simultaneously, leaving the agent completely unavailable until the cooldown expires.

Steps to Reproduce

1. Configure primary as anthropic/claude-opus-4-6, fallback as anthropic-api/claude-opus-4-6
2. Trigger rate limit on subscription (run multiple parallel subagents)
3. Observe All models failed (1) despite fallback being configured