feat(swe-bench): add an iterative predictor with ～70% success rate

[Jerryguan777]

Implemented an iterative agent that solves problems by executing bash commands step-by-step, observing results, and generating patches. Achieved 70% success rate (7/10) in initial evaluation.

• Add IterativeAgent
• Add config_iterative.yml
• Add git tools
• Add SweBenchPredictorIterativeConfig
• Register iterative predictor and git tool
• Update README.md

How Has This Been Tested?

export ANTHROPIC_API_KEY=sk-xxxxxx

nat eval --config_file examples/evaluation_and_profiling/swe_bench/configs/config_iterative.yml

Running 10 instances...
10 ran successfully, 0 failed: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████| 10/10 [04:30<00:00, 27.07s/it]
All instances run.
Cleaning cached images...
Removed 0 images.
Total instances: 10
Instances submitted: 10
Instances completed: 10
Instances incomplete: 0
Instances resolved: 7
Instances unresolved: 3
Instances with empty patches: 0
Instances with errors: 0
Unstopped containers: 0
Unremoved images: 2
Report written to nv_predictor.nat_iterative_1.json
2026-01-15 16:11:45 - INFO     - nat.eval.swe_bench_evaluator.evaluate:194 - Completed swe_bench run nat_iterative_1
2026-01-15 16:11:45 - INFO     - nat.eval.swe_bench_evaluator.evaluate:200 - SWE_bench report and logs written to .tmp/nat/examples/evaluation_and_profiling/swe_bench/iterative/swe_bench_reports directory
2026-01-15 16:11:46 - INFO     - nat.eval.evaluate:275 - Profiler is not enabled. Skipping profiling.
2026-01-15 16:11:46 - INFO     - nat.eval.evaluate:366 - Original config file copied to .tmp/nat/examples/evaluation_and_profiling/swe_bench/iterative/config_original.yml
2026-01-15 16:11:46 - INFO     - nat.eval.evaluate:384 - Effective config (with overrides) saved to .tmp/nat/examples/evaluation_and_profiling/swe_bench/iterative/config_effective.yml
2026-01-15 16:11:46 - INFO     - nat.eval.evaluate:427 - Configuration metadata saved to .tmp/nat/examples/evaluation_and_profiling/swe_bench/iterative/config_metadata.json
2026-01-15 16:11:46 - INFO     - nat.eval.evaluate:449 - Workflow output written to .tmp/nat/examples/evaluation_and_profiling/swe_bench/iterative/workflow_output.json
2026-01-15 16:11:46 - INFO     - nat.eval.evaluate:460 - Evaluation results written to .tmp/nat/examples/evaluation_and_profiling/swe_bench/iterative/swe_bench_output.json

=== EVALUATION SUMMARY ===
Workflow Status: COMPLETED (workflow_output.json)
Total Runtime: 480.70s

Per evaluator results:
| Evaluator   |   Avg Score | Output File           |
|-------------|-------------|-----------------------|
| swe_bench   |         0.7 | swe_bench_output.json |

Description

Closes #1397

By Submitting this PR I confirm:

• I am familiar with the Contributing Guidelines.
• We require that all contributors "sign-off" on their commits. This certifies that the contribution is your original work, or you have rights to submit it under the same license, or a compatible license.
  • Any contribution which contains commits that are not Signed-Off will not be accepted.
• When the PR is ready for review, new or existing tests cover these changes.
• When the PR is ready for review, the documentation is up to date with these changes.

Summary by CodeRabbit

• New Features

  • Added an iterative LLM-driven predictor for SWE-bench with multi-step planning, configurable step limits/timeouts, per-step observability, robust command validation, and an integrated git-repo workspace tool for isolated setup, checkout, and cleanup.

• Documentation

  • README and example workflow updated with benchmark context, evaluation guidance, metrics framing, harder benchmarks, and concrete onboarding steps for adding new predictors.

• Tests

  • Extensive test suite validating iterative flows, safety checks, error recovery, timeouts, workspace isolation, and integrations.