Support mini-swe-agent as agent harness

[wasiahmad]

Summary by CodeRabbit

• New Features

  • Added mini-SWE-agent as a supported agent framework for SWE-bench evaluation, including per-datapoint execution, containerized run support, result generation in SWE-bench format, and integration into the processing flow.
  • Added three mini-SWE-agent prompt/config templates (standard, backticks, XML) to drive interactive, single-command iterative agent runs.

• Documentation

  • Updated SWE-bench docs to reflect support and defaults for three agent frameworks: SWE-agent, mini-SWE-agent, and OpenHands.

[greptile-apps]

_{2 files reviewed, 3 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{No files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{2 files reviewed, 6 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Additional Comments (4)

nemo_skills/inference/eval/swebench.py
extra_body type mismatch

SweBenchInferenceConfig.extra_body is declared as a plain dict, but later code treats it as an OmegaConf object (OmegaConf.to_container(self.cfg.inference.extra_body, ...)). If extra_body remains a regular dict (as annotated/defaulted here), this call will error at runtime. Either declare extra_body with the same OmegaConf-compatible type used elsewhere in the codebase (e.g., DictConfig), or convert/merge it without calling OmegaConf.to_container when it’s already a dict.

---

nemo_skills/inference/eval/swebench.py
Malformed CLI arguments

In _run_swe_agent, there’s no && after uv pip install -e . so the next fragment (uv pip install rich==...) will be appended as additional arguments to the python invocation instead of a separate shell command, causing the container command to fail.

                "uv pip install -e . && "

---

nemo_skills/inference/eval/swebench.py
Unescaped instance_id in shell

data_point['instance_id'] is interpolated into the container command for --output trajectories/{instance_id}.traj.json without shell quoting. If an instance id contains spaces/quotes/metacharacters, the command will break or execute unintended tokens. Please wrap it with shlex.quote(...) (similar to problem_statement) or otherwise ensure it’s safely escaped before embedding in the shell string.

---

nemo_skills/inference/eval/swebench.py
Temp config path traversal

tmp_config_filename = f"configs/config_{data_point['instance_id']}.yaml" is used to build host_tmp_path under self.output_dir. If instance_id contains / or .., this will create nested paths or allow writing/removing files outside output_dir when creating/cleaning up the temp config. Sanitize the filename (e.g., replace path separators and strip ..) or use a hash of instance_id for the on-disk name.

[coderabbitai]

Actionable comments posted: 3

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

docs/evaluation/code.md (1)

280-281: ⚠️ Potential issue | 🟡 Minor

Add note clarifying that mini-SWE-agent is not supported for multilingual evaluation.

The swe-bench-multilingual section should explicitly document that mini-SWE-agent does not have multilingual support yet. While the current documentation correctly omits mini-SWE-agent from the multilingual example, this should be made explicit for clarity, especially since mini-SWE-agent is mentioned as a supported alternative in the regular SWE-bench section (lines 160-161).

Consider adding a note like: "Currently, only OpenHands and SWE-agent support multilingual evaluation. Mini-SWE-agent support for multilingual datasets is not yet available."

🤖 Fix all issues with AI agents

In `@nemo_skills/inference/eval/swebench.py`:
- Line 635: Replace the silent fallback to {} and fail-fast when expected key is
missing: change the access of trajectory_dict.get("info", {}) to direct
dictionary access trajectory_dict["info"] inside the function where
trajectory_info is assigned (variable name trajectory_info in swebench.py) so a
missing "info" raises a KeyError; if you need clearer context, wrap the access
in a try/except and re-raise a more specific error mentioning the trajectory id
or source (use the same trajectory_dict/trajectory_info variables) rather than
silently using an empty dict.

In `@nemo_skills/prompt/config/eval/swe-bench/mini-swe-agent/swebench.yaml`:
- Around line 40-45: The numbered rules in the prompt skip item 2 (they read 1,
3, 4, 5); update the ordered list in swebench.yaml so the sequence is
consecutive by changing "3. The system executes the command(s) in a subshell" to
"2. The system executes the command(s) in a subshell" and then renumber the
following lines accordingly ("4."→"3.", "5."→"4.") so the list reads 1–4 in
proper order.
- Line 18: The XML-like tag casing is inconsistent: the opening tag
`<IMPORTANT>` is uppercase but the closing tag `</important>` is lowercase;
update the closing tag to exactly match the opener (change `</important>` to
`</IMPORTANT>`) or make both tags consistently lowercase (e.g.,
`<important>`/`</important>`), ensuring the pair is identical and preserving the
surrounding sentence and punctuation in the `swebench.yaml` snippet.

🧹 Nitpick comments (2)

docs/evaluation/code.md (1)

160-161: Consider adding mini-SWE-agent-specific expected results or a note about expected performance.

The sample run section only shows expected results for OpenHands (issues_resolved: 48.4). While the instructions say to replace openhands with mini_swe_agent, users would benefit from knowing what resolve rate to expect with mini-swe-agent, even approximately. Based on learnings, documentation for benchmarks should include expected results for tested models.

nemo_skills/inference/eval/swebench.py (1)

558-652: <!-- [scratchpad_start]
[task review_file_5/5 review_method_run_mini_swe_agent]
Deep review of _run_mini_swe_agent method.
[observations]

• Past review comments addressed:
  1. Config path: now uses "eval/swe-bench/mini-swe-agent/swebench" ✓
  2. Python -m: now uses python -m minisweagent.run.mini (dotted module path) ✓
  3. Search path: now uses f"{data_point['instance_id']}.traj.json" matching output ✓
• The completion_kwargs building logic (lines 563-572) is duplicated from _run_swe_agent (lines 497-506)
• The method structure is clean: build kwargs → load config → modify config → write temp config → run command → parse output → cleanup

[analysis]

• Code duplication of completion_kwargs building across _run_swe_agent, _run_mini_swe_agent, and _run_openhands
• This is a refactor opportunity but not essential
• The method correctly handles cleanup in the finally block
• The YAML config manipulation is straightforward

[actions]

• Suggest optional refactor for completion_kwargs duplication
• Approve the overall structure
  [scratchpad_end] -->
  Overall implementation looks solid — past review issues are resolved.

The three issues flagged in previous reviews are all addressed:

1. Default config path now correctly points to "eval/swe-bench/mini-swe-agent/swebench" (line 574)
2. Module invocation uses proper dotted path python -m minisweagent.run.mini (line 613)
3. Search path {instance_id}.traj.json matches the --output argument (lines 617, 624)

The completion_kwargs construction logic (lines 563–572) is duplicated across _run_swe_agent, _run_mini_swe_agent, and _run_openhands. Consider extracting it into a shared helper to reduce duplication.

[greptile-apps]

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[coderabbitai]

Actionable comments posted: 1

🤖 Fix all issues with AI agents

In `@nemo_skills/prompt/config/eval/swe-bench/mini-swe-agent/swebench.yaml`:
- Around line 21-24: In swebench.yaml update the contradictory prompt
requirement on the response rules: replace the phrase "Provide exactly ONE bash
command to execute" with "Provide AT LEAST ONE bash command to execute" so it
matches the other occurrences and the format_error_template semantics; ensure
the change is made in the response instruction block near the top of the file so
all references (including format_error_template and lines that currently state
"AT LEAST ONE") are consistent.

🧹 Nitpick comments (2)

nemo_skills/prompt/config/eval/swe-bench/mini-swe-agent/swebench_backticks.yaml (1)

175-176: Nit: trailing dot on cost_limit value.

cost_limit: 3. is valid YAML (parses as 3.0) but reads oddly. Consider 3.0 for clarity. This is consistent with the same pattern in the sibling swebench.yaml, so it's a minor cosmetic point across all three configs.

nemo_skills/prompt/config/eval/swe-bench/mini-swe-agent/swebench.yaml (1)

113-114: Nit: trailing dot on cost_limit value.

Same as the sibling config — cost_limit: 3. is valid YAML but 3.0 would be more conventional.

[Kipok]

no issues from my side, but will leave this to @ludwig-n for final approval. @wasiahmad please fix DCO

[greptile-apps]

_{3 files reviewed, 4 comments}

_{Edit Code Review Agent Settings | Greptile}

[ludwig-n]

Score for Qwen3-Coder-480B-A35B: ~52.5% (avg of 3 runs). Officially reported score is 55.4%. Good enough to merge in my opinion.

With this framework, patches don't always follow the git diff format, so it causes a higher percentage of "patch can't apply" errors. However, this seems to be a "feature" of mini-swe-agent, as it asks the LLM to create the patch file, rather than running git diff manually after the agent is done like SWE-agent/OpenHands. So it's not an issue with our implementation.

[greptile-apps]

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[wasiahmad]

Evaluated Minimax-M2.1 and Minimax-M2.5 on swe-bench-verified using mini-swe-agent and got the following scores.

• Minimax-M2.1 => Pass@1: 70.6% (avg-of-3: 71.0, 71.0, 69.8)
• Minimax-M2.5 => Pass@1: 75.9% (avg-of-3: 76.6, 75.0, 76.0)

No official reported score of these models with mini-swe-agent.
Their official score for swe-bench-verified is 74.0 and 80.2, respectively.