[reward] refactor: extract make_eval_script / parse_eval_output as public helpers

[aoshen02]

Summary

Extract two standalone functions from SWEBenchRewardSpec so external callers (vime/slime coding_agent_rl sandbox) can build eval scripts and parse results without instantiating the full reward spec class:

• make_eval_script(metadata, workdir) → str: builds a self-contained bash eval script
• parse_eval_output(metadata, output) → (solved, report): parses test output using swebench grading

SWEBenchRewardSpec.compute_reward and _get_eval_report now delegate to these functions, removing ~20 lines of duplicated logic.

Motivation

vime (#250) and slime (THUDM/slime#2079) both need SWE-bench eval in their sandbox.evaluate() path. With public helpers:

from uni_agent.reward.swe_bench import make_eval_script, parse_eval_output

script = make_eval_script(metadata, workdir)
# ... write to sandbox and execute ...
solved, report = parse_eval_output(metadata, output)

Test plan

• Existing tests/ pass (no behavioral change — same logic, just extracted)
• from uni_agent.reward.swe_bench import make_eval_script, parse_eval_output works

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request refactors uni_agent/reward/swe_bench.py by extracting script generation and output parsing logic into standalone, module-level functions (make_eval_script and parse_eval_output), simplifying compute_reward and _get_eval_report. The review feedback suggests making parse_eval_output more robust by defensively handling empty or None values for eval_output and checking if the test lists (FAIL_TO_PASS and PASS_TO_PASS) are already parsed objects rather than raw JSON strings.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

To make this public helper more robust for external callers (e.g., when integrated with Hugging Face datasets or when the evaluation fails/times out), we should:

1. Defensively handle cases where eval_output is None or empty to avoid TypeError during substring checks.
2. Handle cases where FAIL_TO_PASS and PASS_TO_PASS are already parsed as lists/dicts rather than JSON-serialized strings.

Suggested change

	if START_TEST_OUTPUT not in eval_output or END_TEST_OUTPUT not in eval_output:
	return False, report
	test_content = eval_output.split(START_TEST_OUTPUT)[1].split(END_TEST_OUTPUT)[0]
	status_map = MAP_REPO_TO_PARSER[repo](test_content, None)
	report["found_eval_status"] = True
	eval_ref = {
	"instance_id": metadata["instance_id"],
	"FAIL_TO_PASS": json.loads(metadata.get("FAIL_TO_PASS", "[]")),
	"PASS_TO_PASS": json.loads(metadata.get("PASS_TO_PASS", "[]")),
	}
	if not eval_output or START_TEST_OUTPUT not in eval_output or END_TEST_OUTPUT not in eval_output:
	return False, report
	test_content = eval_output.split(START_TEST_OUTPUT)[1].split(END_TEST_OUTPUT)[0]
	status_map = MAP_REPO_TO_PARSER[repo](test_content, None)
	report["found_eval_status"] = True
	fail_to_pass = metadata.get("FAIL_TO_PASS", "[]")
	pass_to_pass = metadata.get("PASS_TO_PASS", "[]")
	eval_ref = {
	"instance_id": metadata["instance_id"],
	"FAIL_TO_PASS": json.loads(fail_to_pass) if isinstance(fail_to_pass, str) else fail_to_pass,
	"PASS_TO_PASS": json.loads(pass_to_pass) if isinstance(pass_to_pass, str) else pass_to_pass,
	}