hermes - 💡(How to fix) Fix Context window auto-probe accumulates for providers without context_length metadata (DeepSeek V4)

Environment

• Hermes v0.13.0 (upgraded from pre-0.13 via git merge upstream/main 2026-05-08)
• Provider: DeepSeek (api.deepseek.com)
• Model: deepseek-v4-pro
• No global model.context_length set

Symptoms

When no explicit context_length is configured for DeepSeek, the CLI status bar displays a climbing context window value across turns:

deepseek-v4-pro · 35% · 2m
deepseek-v4-pro · 35% · 4m
deepseek-v4-pro · 35% · 6m
deepseek-v4-pro · 35% · 8m
deepseek-v4-pro · 35% · 10m

The fill percentage (35%) remains static while the displayed context window grows with every API call. After 5+ turns it can exceed 10M.

Root Cause

DeepSeek API /v1/models does not return context_length metadata:

{"id": "deepseek-v4-pro", ...}  // no context_length field

When get_model_context_length() cannot find a value via API probing, the result appears to accumulate rather than being cached at session start. Each turn triggers a re-probe, and the value compounds.

Fix (Workaround)

Setting provider-level context_length stops the accumulation:

hermes config set providers.deepseek.context_length 1048576

This bypasses the auto-probe chain entirely, matching the existing pattern used for custom providers (e.g., llama.cpp endpoints that also lack context_length in /v1/models).

Expected Behavior

The context window should be resolved once at session init and cached. The hardcoded fallback in model_metadata.py (deepseek-v4-pro: 1_000_000) should be used when API probing returns no value.

Impact

• Status bar shows false/misleading context window sizes
• If the inflated value exceeds real limits, compression may trigger too late or not at all
• Likely affects other providers without context_length in /v1/models (xAI Grok was mentioned in code comments as having the same issue)