When closing sessions with already-stopped containers, the stop/kill
operation would raise an exception, preventing container.remove() from
being called. This left stopped containers in Docker even though they
were removed from cubbi's session tracking.
The fix wraps stop/kill operations in their own try-except block,
allowing the code to always reach container.remove() regardless of
whether the container was already stopped.
- Added documentation for the new --no-cache flag in image build command
- Added documentation for local MCP server support (add-local command)
- Updated MCP server types to include local MCP servers
- Added examples for all three types of MCP servers (Docker, Remote, Local)
- Add LocalMCP model for stdio-based MCP servers
- Implement add_local_mcp() method in MCPManager
- Add 'mcp add-local' CLI command with args and env support
- Update cubbi_init.py MCPConfig with command, args, env fields
- Add local MCP support in interactive configure tool
- Update image plugins (opencode, goose, crush) to handle local MCPs
- OpenCode: Maps to "local" type with command array
- Goose: Maps to "stdio" type with command/args
- Crush: Maps to "stdio" transport type
Local MCPs run as stdio-based commands inside containers, allowing
users to integrate local MCP servers without containerization.
* fix: add crush plugin support too
* feat: comprehensive model management for all standard providers
- Add universal provider support for model fetching (OpenAI, Anthropic, Google, OpenRouter)
- Add default API URLs for standard providers in config.py
- Enhance model fetcher with provider-specific authentication:
* Anthropic: x-api-key header + anthropic-version header
* Google: x-goog-api-key header + custom response format handling
* OpenAI/OpenRouter: Bearer token (unchanged)
- Support Google's unique API response format (models vs data key, name vs id field)
- Update CLI commands to work with all supported provider types
- Enhance configure interface to include all providers (even those without API keys)
- Update both OpenCode and Crush plugins to populate models for all provider types
- Add comprehensive provider support detection methods
* feat: migrate container configuration from env vars to YAML config files
- Replace environment variable-based configuration with structured YAML config files
- Add Pydantic models for type-safe configuration management in cubbi_init.py
- Update container.py to generate /cubbi/config.yaml and mount into containers
- Simplify goose plugin to extract provider from default model format
- Remove complex environment variable handling in favor of direct config access
- Maintain backward compatibility while enabling cleaner plugin architecture
* feat: optimize goose plugin to only pass required API key for selected model
- Update goose plugin to set only the API key for the provider of the selected model
- Add selective API key configuration for anthropic, openai, google, and openrouter
- Update README.md with comprehensive automated testing documentation
- Add litellm/gpt-oss:120b to test.sh model matrix (now 5 images × 4 models = 20 tests)
- Include single prompt command syntax for each tool in the documentation
* feat: add comprehensive integration tests with pytest parametrization
- Create tests/test_integration.py with parametrized tests for 5 images × 4 models (20 combinations)
- Add pytest configuration to exclude integration tests by default
- Add integration marker for selective test running
- Include help command tests and image availability tests
- Document test usage in tests/README_integration.md
Integration tests cover:
- goose, aider, claudecode, opencode, crush images
- anthropic/claude-sonnet-4-20250514, openai/gpt-4o, openrouter/openai/gpt-4o, litellm/gpt-oss:120b models
- Proper command syntax for each tool
- Success validation with exit codes and completion markers
Usage:
- pytest (regular tests only)
- pytest -m integration (integration tests only)
- pytest -m integration -k "goose" (specific image)
* feat: update OpenCode plugin with perfect multi-provider configuration
- Add global STANDARD_PROVIDERS constant for maintainability
- Support custom providers (with baseURL) vs standard providers
- Custom providers: include npm package, name, baseURL, apiKey, models
- Standard providers: include only apiKey and empty models
- Use direct API key values from cubbi config instead of env vars
- Only add default model to the provider that matches the default model
- Use @ai-sdk/openai-compatible for OpenAI-compatible providers
- Preserve model names without transformation
- All providers get required empty models{} section per OpenCode spec
This ensures OpenCode can properly recognize and use both native
providers (anthropic, openai, google, openrouter) and custom
providers (litellm, etc.) with correct configuration format.
* refactor: model is now a combination of provider/model
* feat: add separate integration test for Claude Code without model config
Claude Code is Anthropic-specific and doesn't require model selection like other tools.
Created dedicated test that verifies basic functionality without model preselection.
* feat: update Claude Code and Crush plugins to use new config system
- Claude Code plugin now uses cubbi_config.providers to get Anthropic API key
- Crush plugin updated to use cubbi_config.providers for provider configuration
- Both plugins maintain backwards compatibility with environment variables
- Consistent plugin structure across all cubbi images
* feat: add environments_to_forward support for images
- Add environments_to_forward field to ImageConfig and Image models
- Update container creation logic to forward specified environment variables from host
- Add environments_to_forward to claudecode cubbi_image.yaml to ensure Anthropic API key is always available
- Claude Code now gets required environment variables regardless of model selection
- This ensures Claude Code works properly even when other models are specified
Fixes the issue where Claude Code couldn't access Anthropic API key when using different model configurations.
* refactor: remove unused environment field from cubbi_image.yaml files
The 'environment' field was loaded but never processed at runtime.
Only 'environments_to_forward' is actually used to pass environment
variables from host to container.
Cleaned up configuration files by removing:
- 72 lines from aider/cubbi_image.yaml
- 42 lines from claudecode/cubbi_image.yaml
- 28 lines from crush/cubbi_image.yaml
- 16 lines from goose/cubbi_image.yaml
- Empty environment: [] from opencode/cubbi_image.yaml
This makes the configuration files cleaner and only contains
fields that are actually used by the system.
* feat: implement environment variable forwarding for aider
Updates aider to automatically receive all relevant environment variables
from the host, similar to how opencode works.
Changes:
- Added environments_to_forward field to aider/cubbi_image.yaml with
comprehensive list of API keys, configuration, and proxy variables
- Updated aider_plugin.py to use cubbi_config system for provider/model setup
- Environment variables now forwarded automatically during container creation
- Maintains backward compatibility with legacy environment variables
Environment variables forwarded:
- API Keys: OPENAI_API_KEY, ANTHROPIC_API_KEY, DEEPSEEK_API_KEY, etc.
- Configuration: AIDER_MODEL, GIT_* variables, HTTP_PROXY, etc.
- Timezone: TZ for proper log timestamps
Tested: All aider tests pass, environment variables confirmed forwarded.
* refactor: remove unused volumes and init fields from cubbi_image.yaml files
Both 'volumes' and 'init' fields were loaded but never processed at runtime.
These were incomplete implementations that didn't affect container behavior.
Removed from all 5 images:
- volumes: List with mountPath: /app (incomplete, missing host paths)
- init: pre_command and command fields (unused during container creation)
The cubbi_image.yaml files now only contain fields that are actually used:
- Basic metadata (name, description, version, maintainer, image)
- persistent_configs (working functionality)
- environments_to_forward (working functionality where present)
This makes the configuration files cleaner and eliminates confusion
about what functionality is actually implemented.
* refactor: remove unused ImageInit and VolumeMount models
These models were only referenced in the Image model definition but
never used at runtime since we removed all init: and volumes: fields
from cubbi_image.yaml files.
Removed:
- VolumeMount class (mountPath, description fields)
- ImageInit class (pre_command, command fields)
- init: Optional[ImageInit] field from Image model
- volumes: List[VolumeMount] field from Image model
The Image model now only contains fields that are actually used:
- Basic metadata (name, description, version, maintainer, image)
- environment (loaded but unused - kept for future cleanup)
- persistent_configs (working functionality)
- environments_to_forward (working functionality)
This makes the data model cleaner and eliminates dead code.
* feat: add interactive configuration command
Adds `cubbi configure` command for interactive setup of LLM providers
and models through a user-friendly questionnaire interface.
New features:
- Interactive provider configuration (OpenAI, Anthropic, OpenRouter, etc.)
- API key management with environment variable references
- Model selection with provider/model format validation
- Default settings configuration (image, ports, volumes, etc.)
- Added questionary dependency for interactive prompts
Changes:
- Added cubbi/configure.py with full interactive configuration logic
- Added configure command to cubbi/cli.py
- Updated uv.lock with questionary and prompt-toolkit dependencies
Usage: `cubbi configure`
* refactor: update integration tests for current functionality
Updates integration tests to reflect current cubbi functionality:
test_integration.py:
- Simplified image list (removed crush temporarily)
- Updated model list with current supported models
- Removed outdated help command tests that were timing out
- Simplified claudecode test to basic functionality test
- Updated command templates for current tool versions
test_integration_docker.py:
- Cleaned up container management tests
- Fixed formatting and improved readability
- Updated assertion formatting for better error messages
These changes align the tests with the current state of the codebase
and remove tests that were causing timeouts or failures.
* fix: fix temporary file chmod
* fix: remove config override logging to prevent API key exposure
* feat: add network filtering with domain restrictions
- Add --domains flag to restrict container network access to specific domains/ports
- Integrate monadicalsas/network-filter container for network isolation
- Support domain patterns like 'example.com:443', '*.api.com'
- Add defaults.domains configuration option
- Automatically handle network-filter container lifecycle
- Prevent conflicts between --domains and --network options
* docs: add --domains option to README usage examples
* docs: remove wildcard domain example from --domains help
Wildcard domains are not currently supported by network-filter
Auto-detects connection type from URL (/sse -> sse, /mcp -> streamable_http)
or allows manual specification. Updates goose plugin to use actual MCP type
instead of hardcoded sse.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add mcps field to Session model to track active MCP servers and populate it from container labels in ContainerManager. Enhance MCP remove command to warn when removing servers used by active sessions.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Claude <noreply@anthropic.com>
* feat: implement Aider AI pair programming support
- Add comprehensive Aider Docker image with Python 3.12 and system pip installation
- Implement aider_plugin.py for secure API key management and environment configuration
- Support multiple LLM providers: OpenAI, Anthropic, DeepSeek, Gemini, OpenRouter
- Add persistent configuration for ~/.aider/ and ~/.cache/aider/ directories
- Create comprehensive documentation with usage examples and troubleshooting
- Include automated test suite with 6 test categories covering all functionality
- Update container.py to support DEEPSEEK_API_KEY and GEMINI_API_KEY
- Integrate with Cubbi CLI for seamless session management
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Fix pytest for aider
* Fix pre-commit
---------
Co-authored-by: Your Name <you@example.com>
* feat: add Claude Code image support
Add a new Cubbi image for Claude Code (Anthropic's official CLI) with:
- Full Claude Code CLI functionality via NPM package
- Secure API key management with multiple authentication options
- Enterprise support (Bedrock, Vertex AI, proxy configuration)
- Persistent configuration and cache directories
- Comprehensive test suite and documentation
The image allows users to run Claude Code in containers with proper
isolation, persistent settings, and seamless Cubbi integration. It
gracefully handles missing API keys to allow flexible authentication.
Also adds optional Claude Code API keys to container.py for enterprise
deployments.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Pre-commit fixes
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Your Name <you@example.com>
* refactor: rework how image are defined, in order to create others wrapper for others tools
* refactor: fix issues with ownership
* refactor: image share now information with others images type
* fix: update readme
