hermes - 💡(How to fix) Fix [FR] Auto-switch transport to anthropic_messages when DeepSeek provider points to /anthropic endpoint [1 comments, 1 participants]

Summary

When provider: deepseek is combined with a base_url ending in /anthropic, Hermes should automatically resolve transport to anthropic_messages instead of the hardcoded openai_chat from the deepseek overlay.

DeepSeek's official docs (https://api-docs.deepseek.com/guides/anthropic_api) document an Anthropic-compatible endpoint at https://api.deepseek.com/anthropic. Users who follow these docs and configure Hermes accordingly get silently degraded behavior rather than the expected full Anthropic-protocol support.

Current Behavior (verified in source)

1. ✅ _is_deepseek_anthropic_endpoint() already exists (agent/anthropic_adapter.py:438)
2. ✅ _endpoint_speaks_anthropic_messages() recognizes /anthropic URLs (agent/auxiliary_client.py:890)
3. ✅ Thinking block handling has a dedicated DeepSeek branch (PR #16748)
4. ❌ But hermes_cli/providers.py:123-126 deepseek overlay hardcodes transport: openai_chat, overriding the endpoint detection
5. ❌ agent/auxiliary_client.py transport resolution doesn't follow the main slot's auto-detection for aux slots

Why this matters (silent degradation)

Users following DeepSeek's official docs configure base_url: https://api.deepseek.com/anthropic expecting full Anthropic-protocol features, but get:

• Thinking blocks partially handled (saved by #16748 for the main agent)
• cache_control injection skipped → future prompt-cache support DOA
• Aux slot (compression / hindsight / summarizer) still on openai_chat → asymmetric protocol behavior across one logical session

Currently this configuration is silently degraded rather than rejected or auto-corrected, which is a footgun for users following DeepSeek's official Anthropic-compatible endpoint docs. There is no error or warning — it just runs in degraded mode.

Minimal repro config

model:
  provider: deepseek
  base_url: https://api.deepseek.com/anthropic
  model: deepseek-v4-flash
# Expected: transport=anthropic_messages
# Actual:   transport=openai_chat (from providers.py:123-126 overlay)

Proposed solutions (in order of preference)

A. DeepSeek overlay auto-detects transport from base_url suffix — when base_url ends with /anthropic, resolve transport=anthropic_messages instead of the hardcoded openai_chat. This mirrors how _endpoint_speaks_anthropic_messages() already works for the aux client.

B. DeepSeek overlay declares dual transport, routed by base_url.

C. Config validator warns on the mismatch and suggests provider: anthropic with custom base_url as a workaround.

Files involved

• hermes_cli/providers.py — deepseek overlay (primary fix location, line 123-126)
• agent/auxiliary_client.py — aux slot transport resolution must mirror main slot's auto-detection; otherwise the main conversation uses Anthropic protocol but compression/session-search/skills-hub still call openai_chat, negating the benefit
• agent/run_agent.py::_anthropic_prompt_cache_policy() — future extension point, no change needed today; flagged for when DeepSeek's /anthropic endpoint adds cache_control support (currently listed as "Ignored" in their compatibility matrix)

Out of scope

• Changes to other providers' transport selection
• Implementing prompt-cache pricing/discount logic (server-side concern)
• Modifying _anthropic_prompt_cache_policy() current behavior