fix: prevent Ollama key from polluting global litellm state

[gvago]

Summary

• Root cause: LiteLLMAIHandler.__init__() sets litellm.api_key globally when OLLAMA.API_KEY is configured. Since litellm is a module-level singleton, this overwrites the API key for ALL providers (Groq, xAI, OpenAI, Anthropic, etc.). Any non-Ollama call then sends the Ollama key and fails authentication.
• Fix: Store the Ollama key on the handler instance (self.ollama_api_key) instead of globally, and inject it per-request via litellm's api_key kwarg only when the model uses the ollama/ or ollama_chat/ prefix.
• Tests updated: The test_ollama_and_groq_coexist test now verifies correct per-provider key isolation -- Ollama models get the Ollama key, non-Ollama models get the Groq key.

This builds on prior fixes #2288 and #2293 which addressed related symptoms (Gemini broken by Ollama key, dummy key forwarding) but left the root cause (global litellm.api_key mutation) in place.

Note on GitLab submodule content fetching

There is a separate issue in gitlab_provider.py where _expand_submodule_changes() creates synthetic diff entries with paths like submodule_path/file.py, but get_diff_files() still calls get_pr_file_content() against the parent project ID. These paths don't exist in the parent repo -- the content lives in the submodule project. This is a real bug but requires more invasive changes (threading submodule project resolution through content fetching) and is left for a separate PR.

Test plan

• All 7 existing tests in test_litellm_api_key_guard.py pass
• test_ollama_and_groq_coexist now correctly verifies Ollama key goes only to Ollama models and Groq key goes to non-Ollama models
• Manual verification with Ollama + another provider configured simultaneously

🤖 Generated with Claude Code

[qodo-free-for-open-source-projects]

Review Summary by Qodo

Prevent Ollama key from polluting global litellm state

🐞 Bug fix

Walkthroughs

Description

• Store Ollama API key on handler instance instead of global litellm state
• Inject Ollama key per-request only for ollama/ prefixed models
• Prevent Ollama key from overwriting keys for other providers (Groq, XAI, OpenAI, etc.)
• Update test to verify correct per-provider key isolation

Diagram

flowchart LR
  A["OLLAMA.API_KEY configured"] -->|"Previously: global mutation"| B["litellm.api_key overwritten"]
  B -->|"Breaks other providers"| C["Non-Ollama calls fail"]
  A -->|"Now: instance storage"| D["self.ollama_api_key"]
  D -->|"Per-request injection"| E["Only for ollama/ models"]
  E -->|"Preserves other keys"| F["Groq/XAI/OpenAI work correctly"]

Loading

File Changes

1. pr_agent/algo/ai_handlers/litellm_ai_handler.py 🐞 Bug fix +7/-1

Store and inject Ollama key per-request

• Add self.ollama_api_key instance variable initialized to None
• Store Ollama API key on handler instance instead of setting litellm.api_key globally
• Inject Ollama key per-request via kwargs only when model starts with ollama/ or ollama_chat/
 prefix
• Preserve global litellm.api_key for other providers (Groq, XAI, etc.)

pr_agent/algo/ai_handlers/litellm_ai_handler.py

---

2. tests/unittest/test_litellm_api_key_guard.py 🧪 Tests +10/-10

Verify per-provider key isolation in coexistence test

• Update test_ollama_and_groq_coexist test documentation to reflect new per-request injection
 behavior
• Verify litellm.api_key remains as Groq key after initialization (not overwritten by Ollama)
• Verify handler.ollama_api_key stores the Ollama key on the instance
• Verify Ollama models receive Ollama key and non-Ollama models receive Groq key

tests/unittest/test_litellm_api_key_guard.py

---

[qodo-free-for-open-source-projects]

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (0) 📎 Requirement gaps (0)

Great, no issues found!

Qodo reviewed your code and found no material issues that require review

ⓘ The new review experience is currently in Beta. Learn more