mini-SWE-agent changelog: what's new each month
Every stable mini-SWE-agent release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 12 months covered; the current month updates daily.
July 2026
3 releases: 2.4.4 → 2.4.6
July brought stability improvements to mini-SWE-agent, with fixes addressing import crashes, configuration handling, and cost tracking accuracy.
Highlights
- Fixed startup crashes on systems without UTF-8 support
- Configuration files now read consistently as UTF-8 across all locales
- Tool-call errors are now reported accurately instead of being misidentified as truncation
- Failed API calls are now correctly counted toward cost limits
Improved
- Benchmark environment setup no longer causes thread safety issues with shared configuration
Fixed
- Startup banner emoji caused import failures on non-UTF stdout environments
- Configuration file encoding depended on system locale instead of using UTF-8
- Malformed tool-call errors were incorrectly masked as truncation errors
- Failed API calls that couldn't be parsed were not counted against cost limits
June 2026
5 releases: 2.3.1 → 2.4.3
June brought refinements to environment handling and error management, along with improvements to the inspector interface. The month focused on stability fixes and better visibility into model reasoning.
- Removed default T=0 temperature setting in swe-bench configurations; explicitly set temperature if needed.
Highlights
- Inspector now displays reasoning content with toggle and hot-reload for easier debugging.
- Capped consecutive format errors to prevent runaway error states.
- Fixed timeout handling to properly clean up local process groups.
- Improved environment activation and startup command handling.
New
- Expose finish reason to format error template for better error context.
- Display reasoning content in inspector with toggle control.
Improved
- Persist full model response in extra dict on format errors for debugging.
- Cap consecutive format errors to improve robustness.
- Source bashrc via BASHENV to properly activate testbed environment.
- Program bench updated to use v6 images.
- Reject empty config override keys to prevent invalid configurations.
Fixed
- Fixed environment activation by sourcing bashrc through BASHENV in swe-bench.
- Killed local process group on timeout to ensure proper cleanup.
- Fixed startup command passing to env.execute() function as dict.
May 2026
1 release: 2.3.0
May brought stability improvements to mini-SWE-agent, focusing on environment handling and execution limits to prevent runaway processes.
Highlights
- Added wall-clock time limits to properly terminate long-running agents.
- Fixed cost limit enforcement when set to zero.
- Improved directory creation reliability in sandboxed environments.
Improved
- Wall-clock time limits now properly kill agents that exceed their execution window.
- Directory creation in BubblewrapEnvironment now handles existing directories gracefully.
Fixed
- Cost limit enforcement now works correctly when set to zero.
- Agents could previously run indefinitely; wall-clock time limits now prevent this.
March 2026
3 releases: 2.2.6 → 2.2.8
March brought compatibility updates and security improvements to mini-SWE-agent. The month focused on fixing environment integrations and addressing a critical dependency issue.
Highlights
- Fixed double command execution in ContreeEnvironment
- Excluded compromised litellm versions from dependencies
- Improved content string representation across the codebase
- Resolved compatibility with new contree-sdk version
Improved
- Generalized content string representation
- Enhanced ContreeEnvironment documentation
Fixed
- Added missing exceptioninfo to SwerexModalEnvironment output
- Fixed double command execution in ContreeEnvironment
- Fixed ContreeEnvironment compatibility with new contree-sdk version
February 2026
8 releases: 2.0.0 → 2.2.5
February brought major architectural changes to mini-SWE-agent, including native tool calling support and multimodal input capabilities. The month focused on expanding model compatibility and fixing edge cases across the new infrastructure.
Highlights
- Native tool calling API is now supported alongside text-based action parsing
- Images and other content types can now be provided as input
- Parallel tool calls are now enabled with plain text observation templates
- Startup time has been improved
- Better exit behavior in interactive mode
New
- Multimodal input support for images and other content types
- Native tool calling API
- Parallel tool calls capability
- ContreeEnvironment added
- Plain text observation template
Improved
- More flexible model support in general
- Better behavior on exit in interactive mode
- Startup time of mini improved
- Made 'actions' available in all format error templates
- Only ask for task if unset
Fixed
- Count of actions to confirm
- SwerexModalEnvironment compatibility with v2 Environment protocol
- Bug in ContreeEnvironmentConfig.gettagbyimageurl
- Config reload after setup
- Arguments validation to avoid exceptions
- Inspector handling of ANSI escape sequences and null bytes in content
January 2026
2 releases: 1.17.4 → 1.17.5
January brought bug fixes to the swebench runner and configuration handling, with a major v2 release arriving at month's end. Users should prepare for backwards incompatible changes coming in v2.
- v2 arriving with backwards incompatible changes; pin your dependency to v1 if you need stability, then consult the migration guide when ready to upgrade
Highlights
- Fixed failed submission commands in swebench that were still being submitted despite errors
- Resolved environment configuration being ignored in custom config files
- Fixed a clash in the -c options for the swebench runner
- v2 release arriving with improved flexibility and performance
Fixed
- Failed submission commands in swebench no longer submit when they should fail
- Environment configuration in custom config files is now respected
- Removed clash between -c options in the swebench runner
December 2025
3 releases: 1.17.1 → 1.17.3
December brought stability improvements and fixes to message handling and error reporting. The agent now handles edge cases better and prevents context overflow issues.
Highlights
- Fixed message handling to avoid passing unnecessary information to the model
- Improved error reporting by reraising original exceptions when retries fail
- Made batch progress display stay visible at the top of the screen
- Fixed timeout template truncation to prevent context overflow
Improved
- Batch progress bar now stays pinned at the top for better visibility
- More robust whitelisted action confirmation flow
- Timeout template now truncates content to prevent exceeding context limits
Fixed
- Undefined variable in roulette config
- Model name not being set correctly for LiteLLM cost calculation
- Timestamp field being incorrectly included in OpenAI Response API requests
- Extra message information being passed to models unnecessarily
- Original exceptions not being reraised when retries were exhausted
November 2025
3 releases: 1.15.0 → 1.17.0
November brought configuration flexibility and new model provider support to mini-SWE-agent. The month focused on making the agent more customizable and expanding integration options.
Highlights
- Action parsing regex is now configurable for custom action formats
- Added support for Requesty and Portkey responses API model providers
- Cost tracking errors can be ignored via configuration
- Trajectory output now includes timestamps and parsed actions
New
- Configurable action parsing regex
- Requesty model provider support
- Portkey responses API support
- Responses API integration
- Timestamps in trajectory output
- Parsed actions included in trajectory output
Improved
- Cost tracking now has an ignore-errors mode for more flexible error handling
October 2025
6 releases: 1.13.4 → 1.14.4
October brought updates to the default model and improvements to Anthropic integration, including caching refinements and better retry configuration. Several bugs affecting cache control, cost tracking, and command palette were resolved.
- Cache-related options are deprecated as caching is now handled automatically at end of requests
Highlights
- Updated default model to Claude Sonnet 4.5 for improved performance
- Anthropic caching now applied at end of requests with simplified controls
- Added configurable retry attempts for model calls
- Fixed cost limit flag to work correctly with zero values
- Improved Anthropic integration using litellm with cache control support
New
- Configurable number of retries for model requests
Improved
- Anthropic caching now applied at the end of requests rather than during processing
- Switched to litellm class for Anthropic to enable cache control support
- Enhanced cost tracking for Portkey-based Anthropic calls
Fixed
- Command palette no longer crashes with TypeError in visual mode
- Cache control setcachecontrol function now properly clears cache points
- Cost limit flag now works when set to 0
- Portkey Anthropic cost calculation issues
- Type errors in Portkey cost tracking workaround
September 2025
7 releases: 1.11.0 → 1.13.3
September brought expanded model support and infrastructure improvements to mini-SWE-agent. The release focused on enabling more AI model providers and making the agent more flexible for different deployment scenarios.
Highlights
- Added support for OpenRouter and Portkey model providers, expanding available AI model options.
- Introduced SWE-bench startup command for benchmarking agent performance.
- Made docker pull timeout configurable to handle slow connections.
- Added timeout override argument for environment command execution.
- Improved bash command handling to work correctly with whitespace in arguments.
New
- OpenRouter model support
- Portkey AI provider integration
- SWE-bench startup command for running benchmarks
- Timeout override argument for env.execute
- Support for different model classes via mini run script
Improved
- Bash command line now handles whitespaces correctly
- Docker pull timeout is now configurable
- Full litellm response now included in agent responses
- Agent config location is printed in run scripts
- Model name can be overridden in portkey models
- Undefined variable errors are now raised explicitly
Fixed
- Response serialization error
- Undefined 'processor' variable in template that could cause failures
- Version numbering issue in a previous release
August 2025
12 releases: 1.4.0 → 1.10.0
August brought significant usability improvements and environment management enhancements to mini-SWE-agent, with better interactive workflows, configuration flexibility, and support for additional execution environments like Singularity and Bubblewrap.
- Agent completion sequence changed from MINISWEAGENTFINALOUTPUT to echo COMPLETETASKANDSUBMITFINALOUTPUT.
- Default configuration no longer reads Claude.md and similar files automatically.
- TextualAgent now quits application at end of run.
- Docker containers are now removed by default (previous behavior required explicit cleanup).
Highlights
- Human mode for interactive command entry and exit confirmation in mini -v, enabling manual task assignment during agent runs.
- Support for Singularity and Bubblewrap environments, expanding deployment options beyond Docker.
- Enhanced configuration system with templating, persistent variables, and ability to specify config files without .yaml extension.
- Improved interactive mode with better keystroke display, help shortcuts, and visual mode customization.
- New roulette model class and better cost tracking integration for template formatting.
New
- Human mode for mini -v allowing manual command entry like a shell.
- Exit confirmation feature to reassign tasks when agent completes.
- Singularity environment support for running swebench.
- Bubblewrap environment for sandboxed execution.
- Roulette model class for randomized model selection.
- Support for Claude thinking blocks in output.
- Ability to load model registry from a fixed path.
- Minimal logging system for environments and batch runs.
Improved
- System prompt templating now enabled by default.
- Visual mode default can now be configured by users.
- Configuration files can be specified without .yaml extension.
- Progress bar now shows estimated time for swebench runs.
- Model cost information available in templates.
- Task variables persisted for template formatting across runs.
- Help shortcuts and Ctrl-Y for yolo mode added to textual interface.
- Docker containers now removed by default with --rm flag.
Fixed
- Custom model registry loading was broken and is now working.
- Human messages not being added to history in interactive mode.
- MINISWEVISUALMODE_DEFAULT behavior corrected.
- Config file model name preference now correctly overrides environment variables.
- Local environment variables now properly accessible as template variables in Docker runs.
- Double logging messages eliminated and logging simplified.
July 2025
7 releases: 1.0.0 → 1.3.0
July brought mini-SWE-agent to version 1.3.0 with refinements to configuration, better local model support, and improvements to how the agent reads project documentation. The month focused on stability fixes and making the agent more aware of project-specific guidance files.
- Configuration file renamed from local.yaml to mini.yaml
- Environment variable names changed for consistency
- Step limit changed to 250
Highlights
- Added --exit-immediately flag to stop the agent after finding a solution without further steps.
- Agent now automatically looks for and reads claude.md, copilot-instructions.md and similar guidance files in projects.
- Better support for running with local models through custom litellm registry configuration.
- Configuration now simplified so API keys don't need manual setup if already in environment variables.
New
- --exit-immediately flag for the mini command to exit after solution is found
- Guide documentation for setting up and using local models
Improved
- Agent now looks at claude.md and copilot-instructions.md to understand project conventions
- Agent now looks at additional instruction files like copilot-instructions.md
- Better handling of bash code in output
- Setup process no longer requires manual API key name configuration when key is in environment
- Improved compatibility with Python 3.10
- Better support for local models with custom litellm registry
Fixed
- Fixed extra invocation occurring during execution
- Fixed KeyboardInterrupt not being reraised properly
- Fixed output file handling for runtextual
- Fixed content type validation to ensure string values
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. mini-SWE-agent: pricing, features and alternatives · All changelogs