NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI · #1154 most downloaded on PyPI
Mini SWE Agent - A simple AI software engineering agent
Last release 2 months ago
23 Jul 2026
Release timing varies
gaps range from 8 days to 2 months
Nearly every release is documented
notes for 56 of 57 stable releases
Nothing withdrawn
no release was ever pulled
1 years old
60 releases · first in 2025
One column per month.
fix(agents): count billed calls that fail to parse against cost_limit by @EnesYilmazcode in #915
Full Changelog: v2.4.5...v2.4.6
fix(config): read config files as UTF-8 regardless of locale by @EnesYilmazcode in #892
Full Changelog: v2.4.4...v2.4.5
fix: remove startup-banner emoji that crashes import on non-UTF stdout by @EnesYilmazcode in #886
Full Changelog: v2.4.3...v2.4.4
fix(run): pass the startup_command as a dict to env.execute() function. by @Ethan-Kuang-6666 in #880
Full Changelog: v2.4.2...v2.4.3
Small release to bump image versions in programbench runner.
Small release to bump image versions in programbench runner.
Full Changelog: v2.4.1...v2.4.2
fix: kill local process group on timeout by @Swing-G in #865
Change(swe-bench configs): Drop T=0 default in swe-bench by @klieret in #861
Full Changelog: v2.3.1...v2.4.0
Fix(swebench): source ~/.bashrc via BASH_ENV to activate testbed env by @seek-far in #858
Full Changelog: v2.3.0...v2.3.1
The main feature is compatibility with ProgramBench, a new and ultra-challenging software benchmark.
The main feature is compatibility with ProgramBench, a new and ultra-challenging software benchmark.
<img width="600" height="438" alt="programbench mini announcement" src="https://github.com/user-attachments/assets/ee861c54-1696-4096-b458-70c31b24f557" />
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.2.8...v2.3.0
Warning The dependency litellm in versions 1.82.7 and 1.82.8 was compromised . If you updated/installed litellm in the last few overs SSH and API keys
Warning
The dependency litellm in versions 1.82.7 and 1.82.8 was compromised. If you updated/installed litellm in the last few overs
SSH and API keys might have been compromised.
The compromised versions should have been removed from pypi, but the PR below also excludes them in the dependencies.
Full Changelog: v2.2.7...v2.2.8
Mostly fixes to ContreeEnvironment:
Mostly fixes to ContreeEnvironment:
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.2.6...v2.2.7
Thanks to @AlienKevin for fixing a bug in the exception handling with the SwerexModalEnvironment
Thanks to @AlienKevin for fixing a bug in the exception handling with the SwerexModalEnvironment
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.2.5...v2.2.6
Bugfix release fixing issues with FormatErrors especially for weaker/smaller models that make toolcalling mistakes, and a rare issue in the inspector
Bugfix release fixing issues with FormatErrors especially for weaker/smaller models that make toolcalling mistakes, and a rare issue in the inspector that could hang if encountering ANSI-sequences in trajectories.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.2.4...v2.2.5
Fix a bug that caused an exception about missing default model name if setup is invoked in the first mini invocation right after installation.
Fix a bug that caused an exception about missing default model name if setup is invoked in the first mini invocation right after installation.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.2.3...v2.2.4
Fixed a bug inside ContreeEnvironmentConfig.get_tag_by_image_url by @delphikettle in https://github.com/SWE-agent/mini-swe-agent/pull/751
Minor bug fix.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.2.2...v2.2.3
Enh: Make 'actions' available in all format error templates by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/748
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.2.1...v2.2.2
Small patch fixing SwerexModalEnvironment. Thanks to @arbipher!
Small patch fixing SwerexModalEnvironment. Thanks to @arbipher!
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.2.0...v2.2.1
Main feature is the new ContreeEnvironment develoepd by Nebius. Thanks @delphikettle for implementing it! You should now also be able to run on swe re
Main feature is the new ContreeEnvironment develoepd by Nebius. Thanks @delphikettle for implementing it!
You should now also be able to run on swe re-bench with the main swebench.py script.
We've also finally updated the official SWE-bench leaderboard with the results from mini-swe-agent v2:
<img width="2385" height="1033" alt="main_bar_chart" src="https://github.com/user-attachments/assets/a65b11f3-ad60-4bb3-9a73-53ecba9f05c2" />
You can find more information in this tweet.
Welcome ❤️
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.1.0...v2.2.0
This minor release improves the UX of the interactive agent (i.e., what is called when you run mini). We no longer prompt with the multiline prompt at
This minor release improves the UX of the interactive agent (i.e., what is called when you run mini). We no longer prompt with the multiline prompt at the end of the agent run. You can still submit a multiline new task by entering /m first.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.0.0...v2.1.0
> This is the first full release of mini-swe-agent v2. This includes lots of breaking changes.
[!warning] This is the first full release of mini-swe-agent v2. This includes lots of breaking changes.
This is the largest new release since the launch, focusing on supporting more diverse models and switching to toolcalls by default (as opposed to manual parsing of actions).
I only use the mini CLI with default configs
No changes needed.
I use custom configs
You might need to move some config keys. See Config changes.
I parse/analyze trajectories
Some metadata fields moved to extra. See Trajectory format.
I use the python bindings or built custom subclasses
You will need to refactor. An agent can refactor your code based on the instructions below.
Read the full migration guide.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.17.3...v2.0.0
Feat(agent): Override agent/env class from config/cli by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/730
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.0.0a2...v2.0.0a3
Fix(swebench): Do not use docker login flags on swebench by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/724
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v2.0.0a1...v2.0.0a2
> This is the prerelease of the mini-swe-agent v2. This includes lots of breaking changes.
[!warning] This is the prerelease of the mini-swe-agent v2. This includes lots of breaking changes.
This is the largest new release since the launch, focusing on supporting more diverse models and switching to toolcalls by default (as opposed to manual parsing of actions).
[!warning] This is a pre-release, there might still be a few changes.
To install from pypi, you need to specify the version explicitly:
pip install --pre mini-swe-agent
I only use the mini CLI with default configs
No changes needed.
I use custom configs
You might need to move some config keys. See Config changes.
I parse/analyze trajectories
Some metadata fields moved to extra. See Trajectory format.
I use the python bindings or built custom subclasses
You will need to refactor. An agent can refactor your code based on the instructions below.
Read the full migration guide
> v2 coming on Monday! More flexibility & performance, but backwards breaking changes. Pin your dependency to v1 to avoid interruption and then check…
[!warning] v2 coming on Monday! More flexibility & performance, but backwards breaking changes. Pin your dependency to v1 to avoid interruption and then check the migration guide.
This small patch release fixes a clash in the -c options for the swebench runner.
Thanks to @bofia for the fix in https://github.com/SWE-agent/mini-swe-agent/pull/720
[!note] v1.x releases are now based on the
v1branch. Thev2branch will soon become themainbranch.
> v2 coming soon! More flexibility & performance, but backwards breaking changes. Pin your dependency to v1 to avoid interruption.
[!warning] v2 coming soon! More flexibility & performance, but backwards breaking changes. Pin your dependency to v1 to avoid interruption.
This is (hopefully) one of the last mini v1 releases, mostly fixing two bugs:
mini executable ignored environment config in custom config files (https://github.com/SWE-agent/mini-swe-agent/pull/700), thanks to @IKACE for the report[!note] v1.x releases are now based on the
v1branch. Thev2branch will soon become themainbranch.
This release adds important fixes.
This release adds important fixes.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.17.2...v1.17.3
Enh: Keep batch progress main bar always on top by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/634
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.17.1...v1.17.2
Fix: Make test_whitelist_actions_bypass_confirmation more robust by @Chesars in https://github.com/SWE-agent/mini-swe-agent/pull/620
Minor fixes.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.17.0...v1.17.1
Enh: Include timestamps & parsed actions in traj output by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/618
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.16.0...v1.17.0
This release adds more options to support models, in particular to use the responses API from OpenAI that is necessary to use Codex models, among othe
This release adds more options to support models, in particular to use the responses API from OpenAI that is necessary to use Codex models, among others.
Welcome on board! ❤️
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.15.0...v1.16.0
This version introduces a switch to ignore errors about cost tracking: Set cost_tracking: "ignore_errors" in your agent config or export MSWEA_COST_TR
This version introduces a switch to ignore errors about cost tracking: Set cost_tracking: "ignore_errors" in your agent config or export MSWEA_COST_TRACKING="ignore_errors" globally to ignore errors about missing cost information for local or free models. This is an easier alternative to setting up custom model registries etc. Thank you to @Chesars for initial work on this.
See https://github.com/SWE-agent/mini-swe-agent/pull/593, https://github.com/SWE-agent/mini-swe-agent/pull/594, https://github.com/SWE-agent/mini-swe-agent/pull/596
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.14.4...v1.15.0
Fix: Type error with portkey cost tracking workaround by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/580
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.14.3...v1.14.4
Small release containing fix for mini --cost-limit 0 bug spotted by @Chesars (thank you for your continued contributions)
Small release containing fix for mini --cost-limit 0 bug spotted by @Chesars (thank you for your continued contributions)
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.14.2...v1.14.3
Portkey returns inconsistent token counts for anthropic models as soon as tokens are cached that can even lead to negative calculated costs. This is a
Portkey returns inconsistent token counts for anthropic models as soon as tokens are cached that can even lead to negative calculated costs. This is a workaround. This only affected you if you were using --model-class portkey or similar.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.14.1...v1.14.2
Quick but important fixup to previous release (the set_cache_control method had been accidentally overwritten with a nonsensical implementation).
Quick but important fixup to previous release (the set_cache_control method had been accidentally overwritten with a nonsensical implementation).
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.14.0...v1.14.1
Anthropic key rotation is deprecated (as are some minor internal methods)
This release introduces changes to how we handle Anthropic models, in particular with respect to caching.
These are mostly internal changes, it is unlikely that you need to change anything on your end.
The advantage is that you can now use Anthropic models also with other "backends" like portkey rather than having to use the litellm class.
Here's the breakdown of the changes:
litellmset_cache_control option is explicitly set to NoneFull Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.13.4...v1.14.0
This patch release mostly fixes a bug in the command palette in visual (textual) mode reported & fixed by @Chesars
This patch release mostly fixes a bug in the command palette in visual (textual) mode reported & fixed by @Chesars
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.13.3...v1.13.4
This fixes a bug reported by @lamanchy affecting Windows/WSL where a system property (processor type) referenced in the default prompt template isn't
This fixes a bug reported by @lamanchy affecting Windows/WSL where a system property (processor type) referenced in the default prompt template isn't present.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.13.2...v1.13.3
Fix un-incremented version of last release.
Fix un-incremented version of last release.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.13.1...v1.13.2
Feat: Add portkey support by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/534
Adding portkey support.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.12.0...v1.13.0
@riazarbi contributed an improvement to the action parser after observing that Qwen3-Coder 30B frequently adds extra whitespace before the linebreak a
@riazarbi contributed an improvement to the action parser after observing that Qwen3-Coder 30B frequently adds extra whitespace before the linebreak after <code>```bash</code> in codeblocks, causing mini to not recognize the action.
The new regular expression tolerates whitespace before the linebreak.
env.execute by @SumanthRH in https://github.com/SWE-agent/mini-swe-agent/pull/517Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.11.1...v1.12.0
Fix for serialization errors introduced in v1.11.0
Fix for serialization errors introduced in v1.11.0
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.11.0...v1.11.1
Feat: Support for SWE-bench startup command by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/490
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.10.0...v1.11.0
Also thanks to @ai-jz for detailed experiments on running mini with the OpenAI responses API, see here.
Biggest addition is the bubblewrap environment (lightweight sandboxing) contributed by @pcmoritz, and the code for our blog post on using random model
Biggest addition is the bubblewrap environment (lightweight sandboxing) contributed by @pcmoritz, and the code for our blog post on using random models.
Other than that, the biggest fix is avoiding an error when saving with mini -v.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.9.1...v1.10.0
Fix: InteractiveAgent and TextualAgent missing kwargs support by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/437
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.9.0...v1.9.1
Small improvements & fixes, mostly related to template variables.
Small improvements & fixes, mostly related to template variables.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.8.1...v1.9.0
Thanks for @pcmoritz for spotting this
Thanks for @pcmoritz for spotting this
@pcmoritz added proper singularity support to the SWE-bench evaluation. You can access it with --environment-class singularity
@pcmoritz added proper singularity support to the SWE-bench evaluation. You can access it with --environment-class singularity
Main change is in the python bindings of the agent with textual interface. TextualAgent is ow invoked exactly identically to DefaultAgent and IntaractiveAgent
The fix by @ai-jz is important: Previously, we were incorrectly stripping away trailing newlines in the final output of the agent.
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.7.0...v1.8.0
The agent now finishes its run and submits output by running echo COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT (previously we used MINI_SWE_AGENT_FINAL_OUTPU
The agent now finishes its run and submits output by running echo COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT (previously we used MINI_SWE_AGENT_FINAL_OUTPUT).
This change was made because some models got confused by the directive.
This version also introduces slight changes to the swebench.yaml config that is used for the SWE-bench (bash-only) leaderboard
Reference PR: Change: exit sequence; SWE-bench prompts in https://github.com/SWE-agent/mini-swe-agent/pull/
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.6.0...v1.7.0
This release makes mini play nice with gpt-5 and friends. Check out our guide here.
This release makes mini play nice with gpt-5 and friends. Check out our guide here.
The most important changes in this version are some tweaks to the default config files.
mini_no_temp.yaml config by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/372Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.5.1...v1.6.0
Fix MSWEA_VISUAL_MODE_DEFAULT behavior by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/364
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.5.0...v1.5.1
Change(textual): Quit app at end by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/352
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.4.2...v1.5.0
Fix: Show shortcut for submitting multi-line input by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/344
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.4.1...v1.4.2
Fix: loading custom model registry by @carlosejimenez
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.4.0...v1.4.1
Our leaderboard is live at https://www.swebench.com/: *SWE-bench (bash only)* benchmarks LMs against how well they do with mini on SWE-bench Verified.
Our leaderboard is live at https://www.swebench.com/: SWE-bench (bash only) benchmarks LMs against how well they do with mini on SWE-bench Verified.
mini -v (enter commands manually like in a shell)mini -v (allow to give agent a new task when it wants to finish)Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.3.0...v1.4.0
Better support for local models with custom litellm registry by @carlosejimenez in https://github.com/SWE-agent/mini-swe-agent/pull/306
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.2.0...v1.3.0
Feat: Add --exit-immediately flag to mini by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/294
--exit-immediately flag to mini by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/294Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.1.2...v1.2.0
Fix: output file for run_textual by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/272
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.1.1...v1.1.2
Rename to mini-swe-agent by @klieret in https://github.com/SWE-agent/mini-swe-agent/pull/263
Full Changelog: https://github.com/SWE-agent/mini-swe-agent/compare/v1.0.2...v1.1.1
Nothing published for this version
Your coding agent can read these notes before it upgrades. Set up the MCP server →