Problem

The auxiliary model system (auxiliary_client.py) handles side tasks (compression, vision, session search, web extraction, memory flush) but has multiple UX issues for users on non-OpenRouter providers — especially local-only users (Ollama, vLLM, llama.cpp).

This tracker consolidates the fixes.

Sub-issues and PRs

Fixes (merged)

• Silent cloud billing on local-only setups ([Bug] Auxiliary model silently falls back to Gemini Flash on OpenRouter even when user configured local-only #7559) — PR fix(auxiliary): harden fallback behavior for non-OpenRouter users #7594
  Explicit auxiliary.{task}.provider now treated as hard constraint (no fallback to cloud)
• model="default" sentinel crashes (auxiliary_client.py: Fallback from custom/Ollama provider sends model="default" to API, causing model_not_supported #7512 bug 1) — PR fix(auxiliary): harden fallback behavior for non-OpenRouter users #7594
  Providers without known aux models skipped instead of sending literal "default"
• async_call_llm missing fallback (auxiliary_client.py: Fallback from custom/Ollama provider sends model="default" to API, causing model_not_supported #7512 bug 2) — PR fix(auxiliary): harden fallback behavior for non-OpenRouter users #7594
  Async path now mirrors sync call_llm payment/connection fallback
• Misleading "payment error" logs (auxiliary_client.py: Fallback from custom/Ollama provider sends model="default" to API, causing model_not_supported #7512 bug 3) — PR fix(auxiliary): harden fallback behavior for non-OpenRouter users #7594
  _try_payment_fallback() uses accurate reason in log messages
• Cache hit bypasses model slug check (Auxiliary client cache hit bypasses model slug compatibility check #5809) — PR fix: drop incompatible model slugs on auxiliary client cache hit #5804
  _compat_model() guard on cache-hit paths drops incompatible slugs
• OPENAI_BASE_URL env var poisoning (OPENAI_BASE_URL env var not cleared on provider switch — silently poisons auxiliary clients #5161) — PR fix: warn and clear stale OPENAI_BASE_URL on provider switch #7601
  Warning on startup + auto-clear on provider switch

Fixes (in review)

• api_mode not honored by auxiliary client ([Feature]: auxiliary_client.py should honor api_mode flag (parallel to runtime_provider.py) #6800) — PR fix(auxiliary): honor api_mode in auxiliary client #7630
  _resolve_task_provider_model now returns api_mode; resolve_provider_client wraps clients in CodexAuxiliaryClient when api_mode=codex_responses or auto-detected from URL+model pattern
• Async response contract validation ([Bug]: auxiliary async model path can return invalid payloads and crash session_search with misleading AttributeError #7264) — PR fix(auxiliary): validate response shape in call_llm/async_call_llm #7631
  _validate_llm_response() checks .choices[0].message shape before returning; fails fast with clear RuntimeError instead of propagating misleading AttributeError
• Per-job compression model override for cron (Feature: Per-job compression model override for cron jobs #5438) — PR [Feature]: Optimize the display of Feishu message status. #7632
  Cron jobs auto-inherit compression provider from per-job settings via env vars

Related

• perf: bulk file sync for all remote backends (SSH, Modal, Daytona) #7467 — Bulk file sync tracking (performance, separate from auxiliary UX)
• Daytona backend: bulk upload, config bridge, silent disk cap #7362 — Original Daytona investigation that surfaced the auxiliary issue pattern