snapshot: preserve central wiki state 2026-08-15 (7 modified + 1 new file)

2026-08-15 22:11:30 +08:00
commit 92e8028bb8
71 changed files with 9897 additions and 0 deletions
@@ -0,0 +1,77 @@
+---
+title: Model Providers
+type: system
+status: active
+created: 2026-07-22
+updated: 2026-08-14
+verified_on: 2026-08-14
+confidence: high
+tags: [model, provider, inference, routing]
+sources: [live-probe, config, web-search]
+---
+
+# Model Providers
+
+## Purpose
+Inference providers available to Hermes, their routing paths, and notable model updates.
+
+## Current Provider Stack
+| Provider | URL | Role |
+|---|---|---|
+| `custom:litellm` | `http://litellm:4000/v1` | Main inference (slower path, 6–26s/call) |
+| `custom:omni` | `http://omni:20128/v1` | Auxiliary roles (vision, skills, approval, mcp, web_extract, compression) — factory-reset 2026-08-14 |
+| `opencode-zen` | (internal) | Free-tier model access for cron jobs |
+| Guanaco | `ollama:11434` | LLM proxy/router on CT205 (Ollama stack) |
+
+## Default Model
+- `resilient-free` — resolves through OmniRoute (changed from `auto/best-free` on 2026-08-14 after factory reset)
+- Previous default `auto/best-free` was misconfigured — had no combo definition and routed to paid `gpt-5.6-sol`
+- `freellmapi-auto` (LiteLLM path) still available but slower
+
+## Notable Model Updates
+
+### GPT-5.6 Luna — Model Research (2026-08-14)
+- **Community consensus:** Luna is cleaner/more polished; DeepSeek V4 Flash is more persistent/cheaper
+- **Luna strengths:** visual/UI work, final code review, bounded high-stakes tasks, cleaner structured output
+- **DeepSeek strengths:** automation, persistent executor work, high-volume subtasks, long-context economics
+- **Key caveat:** Provider route matters more than model name — `openai-codex/gpt-5.6-luna`, `openrouter/openai/gpt-5.6-luna`, `deepseek/deepseek-v4-flash`, and `openrouter/deepseek/deepseek-v4-flash` are four different tests
+- **Pricing:** Direct OpenAI $0.20/M input, $1.20/M output; OpenRouter $0.10/M input, $0.60/M output
+- **Context:** 1.05M tokens, 128K output, function calling, structured outputs, image input, configurable reasoning effort
+- **Recommendation:** Luna plans and reviews; DeepSeek executes — but test on a clean direct route before judging
+
+### Guanaco v0.8.14 (2026-08-14)
+- Updated from v0.8.11 on CT205 — three incremental bug-fix releases focused on DSML/tool-call output hygiene
+- DSML tag leakage fixes: malformed tool-call output no longer leaks raw syntax junk into model context
+- Dashboard performance fix: cached content size, fixed wrong column names in history query
+- No new features, no breaking changes — low-risk update
+- See [[systems/guanaco]]
+
+### DeepSeek V4 Flash 0731 (2026-07-31)
+- **Same model ID:** `deepseek-v4-flash` — no ID change, just a post-training bump
+- **Same architecture:** 284B total params, 13B active per token, 1M context window
+- **Same pricing:** $0.14/M input, $0.28/M output
+- **Benchmark jump:** Terminal Bench 2.1: 82.7 (up from 61.8 preview — 34% relative gain)
+- **Now surpasses V4-Pro preview** (72.1) on agent/coding benchmarks
+- **Stricter tool name validation:** Rejects tool names not matching `^[a-zA-Z0-9_-]+$` — may cause issues with OmniRoute-transformed tool names
+- **Impact:** Your existing cron jobs using `deepseek-v4-flash` get the upgrade automatically. The empty-stream fallback issue (`opencode-go/deepseek-v4-flash` returning empty 200s) is a provider plumbing problem, not a model quality problem.
+
+### OmniRoute Free-DeepSeek Silent-Stop Mitigation (2026-08-04)
+- Free-DeepSeek upstreams returning empty-after-tool responses since Aug 1 → silent turn stops on `deepseek-flash-free` and `smart-route`/`good-free` across both instances
+- Reworked both combos: strategy `priority` → `fill-first`, dropped proven culprits (`238d5978/deepseek-v4-flash` + `-0731`, `nvidia` deepseek), healthy free models promoted, deepseek demoted to last resort
+- See [[incidents/2026-08-04-omniroute-free-deepseek-silent-stops]]
+
+## Context Enforcement
+- Three layers: model native → combo `data.config.context_length` → Hermes client cap
+- Effective context is the minimum of the three
+
+## Combo Behavior
+- `priority` retries same provider up to `maxRetries`, then errors; it does NOT fall through
+- `fill-first` walks the list on failure
+- `auto`/`lkgp` recalculates per request using health scoring
+- Same-provider entries in a combo share rate-limit buckets
+- Session affinity caches provider in `session_model_history`
+
+## Related
+- [[concepts/model-routing]]
+- [[systems/omniroute]]
+- [[systems/opencode-go]]