Document migrated API-key providers and the free, fast, free-chat combos

OPERATIONS.md

@@ -50,12 +50,19 @@ The container disk is wiped on every rebuild or restart, so the DB lives in the
 Lives in the DB, so it survives restarts through the bucket backup.
 
 - **Login:** `INITIAL_PASSWORD` secret. **API key:** one key named `main` (the owner keeps its value).
-- **Combos** (use the combo name as the `model` field):
+- **Provider connections:** API keys migrated from the 9router DB on 2026-10-19: Gemini (2 keys),
+  OpenRouter, Ollama Cloud, Groq, NVIDIA NIM, Cerebras, Cloudflare Workers AI, Mistral, Cohere. OAuth
+  accounts (Copilot, Antigravity, Gemini CLI, Qoder, Kilo, Cline, Kiro, Codex, Claude) were not migrated;
+  re-login in the dashboard if needed.
+- **Combos** (use the combo name as the `model` field; all `priority`, `maxRetries: 0`, `failoverBeforeRetry: true`):
 
-  | Combo | Order (strategy `priority`, `maxRetries: 0`, `failoverBeforeRetry: true`) | Use |
+  | Combo | Order | Use |
   |---|---|---|
-  | `free` | `ddgw/mistral-small-2603` → `ddgw/tinfoil/gemma4-31b` → `ddgw/tinfoil/gpt-oss-120b` → `ddgw/claude-haiku-4-5` → `ddgw/gpt-5.4-mini` | Coding agents / tool calls. ~1–2 s, real `tool_calls` |
-  | `free-chat` | `ddgw/gpt-5.6-luna` → `cfp/deepseek-ai/deepseek-v4-pro-0813` → `cfp/zai-org/glm-5.2` → `cfp/moonshotai/kimi-k2.7-code` → `ddgw/tinfoil/gpt-oss-120b` → `ddgw/mistral-small-2603` | Plain chat |
+  | `free` | Gemini 3.8 Flash → Gemini 3.7 Flash → Ollama Cloud gpt-oss:120b → Cohere Command-A → Groq gpt-oss-120b → Codestral → Ollama gemma4:31b → Gemini 2.5 Flash → Gemini 3.1 Flash Lite → NVIDIA gemma-4-31b → DuckDuckGo (mistral-small, gpt-oss-120b, gemma4-31b) | Coding agents / tool calls, quality first |
+  | `fast` | Ollama Cloud gpt-oss:120b → Ollama gemma4:31b → Cohere Command-A → Gemini 3.1 Flash Lite → DuckDuckGo mistral-small, gemma4-31b | Lowest latency (~1 s) with tool calls |
+  | `free-chat` | Gemini 3.8 Flash → Cohere Command-A Reasoning → Ollama gpt-oss:120b → DuckDuckGo GPT-5.6 Luna → Cloudflare Playground DeepSeek V4 Pro, GLM-5.2 → DuckDuckGo mistral-small | Plain chat |
+
+  Built-in `auto/*` combos also work now that real API-key providers exist.
 
 - **Retention** (Settings → Storage): call logs 2 d; quota snapshots, compression analytics, MCP audit,
   A2A events, XP audit 7 d; usage history, memory, conversation nodes 30 d; internal DB backups keep 2,
@@ -71,9 +78,21 @@ Lives in the DB, so it survives restarts through the bucket backup.
 | ZCode (`zc/`), Augment (`aug/`), Codex App-Server (`cxa/`), Devin bridge (`dva/`) | Need their CLI binaries inside the container (`spawn … ENOENT`) |
 | UncloseAI (`unc/`) | 404 model not found upstream |
 
-Built-in `auto/*` combos skip no-auth providers ("no executable targets" / "skipped by pre-dispatch
-filters"); use the custom combos above. Adding real API keys or OAuth accounts (e.g. Google AI Studio
-keys under Free Tier) makes `auto/*` usable and is the best way to improve quality.
+### API-key providers (tested 2026-10-19, tool calls)
+
+| Provider | Working models | Not working |
+|---|---|---|
+| Gemini | 3.8 Flash, 3.7 Flash, 3.1 Flash Lite, 2.5 Flash | 3.1 Pro preview (429, free-tier cooldown) |
+| Ollama Cloud | gpt-oss:120b(-medium), gemma4:31b (~0.7 s) | glm-5.2, deepseek-v4-pro (402, Pro plan only); plain `gpt-oss:120b` id 404 |
+| Cohere | command-a-03-2025, command-a-reasoning-08-2025 | — |
+| Groq | openai/gpt-oss-120b | llama-3.3-70b, qwen3-32b (404) |
+| Mistral | codestral-latest | large/small/devstral (403/429 rate limit) |
+| NVIDIA NIM | google/gemma-4-31b-it | kimi-k3, deepseek-v4-pro, gpt-oss-120b (410) |
+| OpenRouter | `:free` models (slow, ~20 s) | — |
+| Cerebras | — | credits exhausted (402) |
+| Cloudflare Workers AI | not tested (no models listed) | — |
+
+Built-in `auto/*` combos skip no-auth providers; they work through the API-key providers above.
 
 ## Required Space secrets