Bug Description
When MiniMax is the configured provider, the compression/auxiliary client sends requests to https://api.minimax.io/anthropic. That endpoint only accepts Anthropic Messages format. The auxiliary client uses the OpenAI SDK (chat completions format), which needs https://api.minimax.io/v1.
Every compression call returns 404. The fallback model handles compression instead — not guaranteed to be equivalent in quality or cost.
Steps to Reproduce
- Set MiniMax as primary provider (
model.provider=minimax, model.base_url=https://api.minimax.io/anthropic)
- Run a conversation long enough to trigger context compression
- Check gateway log for:
Non-retryable client error: <html><head><title>404 Not Found</title></head>
Expected Behavior
Compression hits https://api.minimax.io/v1/chat/completions and succeeds.
Actual Behavior
POST https://api.minimax.io/anthropic/chat/completions returns 404. Context summary is skipped. Fallback model activates.
Affected Component
Agent Core (conversation loop, context compression, memory)
OS
Ubuntu 24.04
Python Version
3.11.x
Hermes Version
(v0.7.x)
Relevant Logs / Traceback
2026-04-07 07:25:47,445 ERROR root: Non-retryable client error: <html>
<head><title>404 Not Found</title></head>
<body><center><h1>404 Not Found</h1></center><hr><center>nginx</center>
</body></html>
Root Cause Analysis
hermes_cli/auth.py line 148:
"minimax": ProviderConfig(
inference_base_url="https://api.minimax.io/anthropic",
...
)Correct for the main agent loop. The main loop sends Anthropic Messages format to /anthropic/v1/messages.
agent/auxiliary_client.py uses the same inference_base_url to construct an OpenAI SDK client:
return OpenAI(api_key=api_key, base_url=base_url), model
The SDK appends /chat/completions, so requests go to https://api.minimax.io/anthropic/chat/completions — 404.
The /v1 path works for chat completions. Verified with curl:
POST https://api.minimax.io/anthropic/chat/completions → 404 (nginx)
POST https://api.minimax.io/v1/chat/completions → non-404 (processes request)
Proposed Fix
Option A (auxiliary_client.py): Strip /anthropic from base URLs before constructing OpenAI SDK clients:
if "/anthropic" in (base_url or "").rstrip("/"):
base_url = base_url.rstrip("/").replace("/anthropic", "/v1")Option B (auth.py + auxiliary_client.py): Add auxiliary_base_url to ProviderConfig. Auxiliary client checks it before falling back to inference_base_url:
"minimax": ProviderConfig(
inference_base_url="https://api.minimax.io/anthropic",
auxiliary_base_url="https://api.minimax.io/v1",
...
)Option C (auxiliary_client.py): When summary_base_url is set in config.yaml, use it directly for compression tasks instead of going through the provider registry.
Related Issues
Are you willing to submit a PR for this?
Bug Description
When MiniMax is the configured provider, the compression/auxiliary client sends requests to
https://api.minimax.io/anthropic. That endpoint only accepts Anthropic Messages format. The auxiliary client uses the OpenAI SDK (chat completions format), which needshttps://api.minimax.io/v1.Every compression call returns 404. The fallback model handles compression instead — not guaranteed to be equivalent in quality or cost.
Steps to Reproduce
model.provider=minimax,model.base_url=https://api.minimax.io/anthropic)Non-retryable client error: <html><head><title>404 Not Found</title></head>Expected Behavior
Compression hits
https://api.minimax.io/v1/chat/completionsand succeeds.Actual Behavior
POST https://api.minimax.io/anthropic/chat/completionsreturns 404. Context summary is skipped. Fallback model activates.Affected Component
Agent Core (conversation loop, context compression, memory)
OS
Ubuntu 24.04
Python Version
3.11.x
Hermes Version
(v0.7.x)
Relevant Logs / Traceback
Root Cause Analysis
hermes_cli/auth.pyline 148:Correct for the main agent loop. The main loop sends Anthropic Messages format to
/anthropic/v1/messages.agent/auxiliary_client.pyuses the sameinference_base_urlto construct an OpenAI SDK client:The SDK appends
/chat/completions, so requests go tohttps://api.minimax.io/anthropic/chat/completions— 404.The
/v1path works for chat completions. Verified with curl:Proposed Fix
Option A (auxiliary_client.py): Strip
/anthropicfrom base URLs before constructing OpenAI SDK clients:Option B (auth.py + auxiliary_client.py): Add
auxiliary_base_urltoProviderConfig. Auxiliary client checks it before falling back toinference_base_url:Option C (auxiliary_client.py): When
summary_base_urlis set inconfig.yaml, use it directly for compression tasks instead of going through the provider registry.Related Issues
Are you willing to submit a PR for this?