* feat(quickstart): add Local (Ollama) LLM provider option
- Detect Ollama via 'ollama list' in quickstart.sh and quickstart.ps1
- Add 'Local (Ollama)' menu option with interactive model picker
- Save provider=ollama, model=<selected> to ~/.hive/configuration.json
- Omit api_key_env_var for Ollama (no API key required)
Refs #5154, #5231
* feat: add local Ollama support and resolve native tool calling
This integrates Ollama as a first-class local provider choice during quickstart, and patches several configuration barriers preventing local models from safely executing the framework's agent graphs.
* **Quickstart Integration**: Added `Local (Ollama)` to the provider menu in both quickstart.sh and quickstart.ps1. When selected, it automatically queries `ollama list` and allows the user to pick an installed model without prompting for an API key.
* **Routing & Configuration**: Automatically sets `"api_base": "http://localhost:11434"` so LiteLLM routes correctly to the local daemon, and increases the default max_tokens config.py allocation to `32768`.
* **Native Tool Calling**: Normalized Ollama models to strictly use the ollama_chat provider prefix inside litellm.py and registered them as `supports_function_calling: True`. This forces native structured function calling and fixes the infinite loop caused by JSON-mode text fallbacks.
* **Context Truncation Fix**: Updated config.py to explicitly pass `"num_ctx": 16384` to Ollama. This prevents the local daemon from silently truncating the Queen agent's ~9,500 token system prompt (Ollama defaults to 2048 `num_ctx`).
* **UX Warnings**: Added terminal notices warning users to select high-parameter models (e.g., `qwen2.5:72b+`) to ensure sufficient contextual reasoning abilities.
Resolves#6027Resolves#6028
* test: add unit tests for Ollama helper functions
Cover _is_ollama_model(), _ensure_ollama_chat_prefix(), and num_ctx
injection in get_llm_extra_kwargs() as requested in PR review.
Fix existing test_init_ollama_no_key_needed assertion to expect the
normalised ollama_chat/ prefix.
Made-with: Cursor
* chores: fixed merge conflict
* fix(ollama): address PR review comments and normalize provider config
* fix(ollama): align quickstart defaults and add tool_choice comment
* fix(ollama): enforce OLLAMA_DETECTED logic and resolve quickstart script syntax errors
* fix(ollama): align quickstart logic and cleanup test imports
Anthropic tightened OAuth validation on 2026-03-17, requiring a
specific User-Agent header and a billing integrity system block for
subscription-authenticated requests. Without these, all OAuth calls
return HTTP 400 with a generic "Error" message.
Changes:
- Add billing integrity system block (SHA-256 hash derived from first
user message content) prepended to system messages on OAuth requests
- Set User-Agent to claude-code/<version> for OAuth sessions
- Fix OAuth header patch to detect tokens in x-api-key (not just
Authorization) and add required beta/browser-access headers
- Set litellm.drop_params=True to prevent unsupported params like
stream_options from leaking to Anthropic (causes 400)
- Skip stream_options entirely for Anthropic models
- Honour LITELLM_LOG env var for debug logging instead of hardcoding
LiteLLM logger to WARNING
- add OpenRouter chat completion validation to key checks for quickstart flows
- improve OpenRouter compat parsing to convert plain textual tool calls into real tool events
- prevent tool-call text from leaking into assistant responses
- add regression tests for OpenRouter key checks and LiteLLM tool compat parsing
Closes#5753
_patch_litellm_anthropic_oauth and _patch_litellm_metadata_nonetype
silently return when litellm internal modules change. This adds
logger.warning() calls so operators are alerted when patches cannot be
applied, instead of encountering cryptic 401 or TypeError at runtime.
Co-authored-by: GowthamT-1610 <gowthamt@umd.edu>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Richard Tang