- Add ollama_num_ctx setting (default 4096) to config.py
- Pass num_ctx option to Ollama in agent.py and agents/base.py
- Add OLLAMA_NUM_CTX to .env.example with usage docs
- Add context_window note in providers.yaml
- Fix mock_settings in test_agent.py for new attribute
- qwen3:30b with 4096 ctx uses ~19GB vs 45GB default
Model upgrade:
- qwen2.5:14b → qwen3.5:latest across config, tools, and docs
- Added qwen3.5 to multimodal model registry
Self-hosted Gitea CI:
- .gitea/workflows/tests.yml: lint + test jobs via act_runner
- Unified Dockerfile: pre-baked deps from poetry.lock for fast CI
- sitepackages=true in tox for ~2s dep resolution (was ~40s)
- OLLAMA_URL set to dead port in CI to prevent real LLM calls
Test isolation fixes:
- Smoke test fixture mocks create_timmy (was hitting real Ollama)
- WebSocket sends initial_state before joining broadcast pool (race fix)
- Tests use settings.ollama_model/url instead of hardcoded values
- skip_ci marker for Ollama-dependent tests, excluded in CI tox envs
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: set qwen3.5:latest as default model
- Make qwen3.5:latest the primary default model for faster inference
- Move llama3.1:8b-instruct to fallback chain
- Update text fallback chain to prioritize qwen3.5:latest
Retains full backward compatibility via cascade fallback.
* test: remove ~55 brittle, duplicate, and useless tests
Audit of all 100 test files identified tests that provided no real
regression protection. Removed:
- 4 files deleted entirely: test_setup_script (always skipped),
test_csrf_bypass (tautological assertions), test_input_validation
(accepts 200-500 status codes), test_security_regression (fragile
source-pattern checks redundant with rendering tests)
- Duplicate test classes (TestToolTracking, TestCalculatorExtended)
- Mock-only tests that just verify mock wiring, not behavior
- Structurally broken tests (TestCreateToolFunctions patches after import)
- Empty/pass-body tests and meaningless assertions (len > 20)
- Flaky subprocess tests (aider tool calling real binary)
All 1328 remaining tests pass. Net: -699 lines, zero coverage loss.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: prevent test pollution from autoresearch_enabled mutation
test_autoresearch_perplexity.py was setting settings.autoresearch_enabled = True
but never restoring it in the finally block — polluting subsequent tests.
When pytest-randomly ordered it before test_experiments_page_shows_disabled_when_off,
the victim test saw enabled=True and failed to find "Disabled" in the page.
Fix both sides:
- Restore autoresearch_enabled in the finally block (root cause)
- Mock settings explicitly in the victim test (defense in depth)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Trip T <trip@local>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
- Add MultiModalManager with capability detection for vision/audio/tools
- Define fallback chains: vision (llama3.2:3b -> llava:7b -> moondream)
tools (llama3.1:8b-instruct -> qwen2.5:7b)
- Update CascadeRouter to detect content type and select appropriate models
- Add model pulling with automatic fallback in agent creation
- Update providers.yaml with multi-modal model configurations
- Update OllamaAdapter to use model resolution with vision support
Tests: All 96 infrastructure tests pass
Inspired by OpenClaw-RL's multi-model orchestration, this adds four
features for custom model management:
1. Custom model registry (infrastructure/models/registry.py) — SQLite-backed
registry for GGUF, safetensors, HF checkpoint, and Ollama models with
role-based lookups (general, reward, teacher, judge).
2. Per-agent model assignment — each swarm persona can use a different model
instead of sharing the global default. Resolved via registry assignment >
persona default > global default.
3. Runtime model management API (/api/v1/models) — REST endpoints to register,
list, assign, enable/disable, and remove custom models without restart.
Includes a dashboard page at /models.
4. Reward model scoring (PRM-style) — majority-vote quality evaluation of
agent outputs using a configurable reward model. Scores persist in SQLite
and feed into the swarm learner.
New config settings: custom_weights_dir, reward_model_enabled,
reward_model_name, reward_model_votes.
54 new tests covering registry CRUD, API endpoints, agent assignments,
role lookups, and reward scoring.
https://claude.ai/code/session_01V4iTozMwcE2gjfnCJdCugC
- YAML-based provider configuration (config/providers.yaml)
- Priority-ordered provider routing
- Circuit breaker pattern for failing providers
- Health check and availability monitoring
- Metrics tracking (latency, errors, success rates)
- Support for Ollama, OpenAI, Anthropic, AirLLM providers
- Automatic failover on rate limits or errors
- REST API endpoints for monitoring and control
- 41 comprehensive tests
API Endpoints:
- POST /api/v1/router/complete - Chat completion with failover
- GET /api/v1/router/status - Provider health status
- GET /api/v1/router/metrics - Detailed metrics
- GET /api/v1/router/providers - List all providers
- POST /api/v1/router/providers/{name}/control - Enable/disable/reset
- POST /api/v1/router/health-check - Run health checks
- GET /api/v1/router/config - View configuration
