Thinking Path

• custom_providers[].models dict (YAML key-value) is parsed but never iterated for dropdown population
• Only the singular model field was read → users with multiple models in models dict see only one
• This compounds with SSRF false positives (SSRF check blocks legitimate local-network custom providers from model auto-detection #1105): when auto-detect is blocked, custom_providers is the only fallback but only picks up one model
• Fix: after reading model field, also iterate models dict keys and add each as a selectable model

What Changed

• api/config.py: _build_available_models_uncached() now collects model IDs from both model (singular) and models (dict) fields in custom_providers entries. Dict keys are deduplicated against the singular model field. Non-string keys are silently skipped.
• tests/test_issue1106_custom_providers_models.py: 7 regression tests covering dict iteration, no duplicates, overlap handling, empty dict, non-string keys, unnamed providers, and multiple named providers.

Why It Matters

Users with local inference servers (llama.cpp, llama-swap, vLLM, TabbyAPI) list their models in custom_providers[].models in config.yaml. Currently only one model appears in the dropdown. After this fix, all models from the dict are available for selection.

Verification

• pytest tests/test_issue1106_custom_providers_models.py -v — 7/7 pass
• pytest tests/test_custom_provider_display_name.py tests/test_model_resolver.py -v — 26/27 pass (1 pre-existing failure)
• Syntax check: py_compile api/config.py — OK

Risks / Follow-ups

• models dict values (metadata like context_length) are not yet used — could power context length display in dropdown in a future PR
• Relates to SSRF check blocks legitimate local-network custom providers from model auto-detection #1105 (SSRF false positive) — when both are fixed, local model servers will fully work

Model Used

• Provider: zai
• Model: glm-5-turbo
• Tools: Hermes Agent

Closes #1106