feat: centralized provider router, call_llm API, unified /model command

[teknium1]

Summary

Complete provider routing infrastructure overhaul. Every LLM call now flows through a centralized router that handles auth, request formatting, and provider-specific quirks in one place.

Core Architecture

resolve_provider_client(provider, model, async_mode, raw_codex)

Central entry point for creating LLM clients. Handles:

• Auth lookup (env vars for API-key providers, OAuth tokens for Nous/Codex)
• Base URL resolution with env var overrides
• API format (Chat Completions vs Responses API adapter for Codex)
• Provider-specific headers (OpenRouter attribution, Kimi User-Agent)
• raw_codex mode for main agent (direct responses.stream access)

call_llm(task, provider, model, messages, ...) / async_call_llm(...)

Full request lifecycle:

1. Resolve provider+model from task config or explicit args
2. Get/create cached client via router
3. Build request kwargs (max_tokens handling, provider extra_body)
4. Make API call with max_tokens/max_completion_tokens retry
5. Return response

Config slots (config.yaml v7)

Every auxiliary task has a provider:model config slot:

auxiliary:
  compression: {provider: auto, model: ""}
  vision: {provider: auto, model: ""}
  web_extract: {provider: auto, model: ""}
  session_search: {provider: auto, model: ""}
  skills_hub: {provider: auto, model: ""}
  mcp: {provider: auto, model: ""}
  flush_memories: {provider: auto, model: ""}

Changes by Area

Auxiliary Consumers Migrated (all use call_llm/async_call_llm)

• context_compressor.py — compression summaries
• vision_tools.py — image analysis (fixed Codex bypass)
• web_tools.py — page summarization
• session_search_tool.py — session summarization
• browser_tool.py — snapshot summarization + browser vision
• mcp_tool.py — MCP sampling
• skills_guard.py — skill audit
• run_agent.py flush_memories — memory flush
• trajectory_compressor.py — trajectory summarization
• mini_swe_runner.py — SWE benchmark runner
• openrouter_client.py — shared OpenRouter client

Main Agent (run_agent.py)

• __init__: Uses router when no explicit creds provided
• _try_activate_fallback: Replaced duplicated _FALLBACK_API_KEY_PROVIDERS / _FALLBACK_OAUTH_PROVIDERS dicts with single resolve_provider_client() call
• Removed _resolve_fallback_credentials method

Provider Fixes

• Nous Portal: Don't send OpenRouter provider preferences (only, ignore, etc.) — caused 404. Don't send reasoning: {enabled: false} — Nous requires reasoning enabled
• Codex vision bypass: vision_tools.py was constructing raw AsyncOpenAI bypassing Codex adapter → crash. Fixed with get_async_vision_auxiliary_client()
• Vision errors: Clear message when model doesn't support vision

Config & Auth

• Removed LLM_MODEL env var — config.yaml is sole source of truth. Avoids conflicts in multi-agent setups. Removed from cli.py, auth.py, setup.py, gateway, cron
• Removed nous-api provider — Nous Portal is OAuth only

UX

• Unified /model and /provider — both show same view with all authenticated providers, their models, current selection, and switch examples
• Added curated model lists for Nous Portal and OpenAI Codex

Test Results

3251 passed, 2 pre-existing unrelated failures

Live Testing

• Tested hermes model switching between OpenRouter, Nous Portal, OpenAI Codex
• Tested /model provider:model mid-conversation switching (history preserved)
• Tested tool calls, vision, web search across all providers
• Tested auxiliary routing (vision/compression/web_extract) with each primary provider