build: expose image build parallelism knobs to workflows

[simonrosenberg]

Summary

• add explicit build-batch-size plumbing to the shared image build helper and current build entrypoints
• expose consistent workflow inputs for SWE-Bench and SWT-Bench image build batch sizing alongside max worker controls
• ensure explicit workflow values win over BUILD_BATCH_SIZE environment defaults, with regression coverage
• fix benchmarks/swebenchmultilingual/build_images.py so --force-build is forwarded instead of being silently ignored

Testing

• uv run pre-commit run --files .github/workflows/build-swebench-images.yml .github/workflows/build-swtbench-images.yml benchmarks/utils/build_utils.py benchmarks/commit0/build_images.py benchmarks/gaia/build_images.py benchmarks/multiswebench/build_images.py benchmarks/swebench/build_images.py benchmarks/swebenchmultilingual/build_images.py benchmarks/swebenchmultimodal/build_images.py benchmarks/swegym/build_images.py benchmarks/swesmith/build_images.py benchmarks/swtbench/build_images.py tests/test_image_utils.py
• uv run pytest tests/test_image_utils.py

Split out of #507.

[all-hands-bot]

Taste Rating: 🟡 Acceptable - Solves a real operational problem with clean parameter threading, but has minor consistency issues.

Linus-Style Analysis:

This is good pragmatic work. You're exposing real knobs that matter for actual production builds, and the parameter threading is straightforward. The test proves the precedence logic works. Worth merging.

Minor warts:

• Workflow parameter naming is inconsistent (see inline comments)
• One unrelated bug fix snuck in that should be called out

See inline comments for details.

[simonrosenberg]

Addressed the review feedback:

• normalized the SWT workflow input naming so the agent-image knob is now build-batch-size, matching SWE-Bench
• renamed the eval-env-specific SWT knob to eval-env-build-batch-size to keep the two concepts distinct
• updated the PR description to explicitly call out the pre-existing swebenchmultilingual --force-build forwarding bug that this branch also fixes

[simonrosenberg]

Merged current origin/main into this branch and resolved the conflicts by keeping both sets of changes:

• retained the SDK sdist-cache support from #515 in benchmarks/utils/build_utils.py, tests/test_image_utils.py, uv.lock, and the vendored SDK pointer
• kept the explicit workflow/build-script parallelism plumbing from #516, including the normalized SWT workflow input names

Validation after conflict resolution:

• uv run pre-commit run --files .github/workflows/build-swebench-images.yml .github/workflows/build-swtbench-images.yml benchmarks/utils/build_utils.py benchmarks/commit0/build_images.py benchmarks/gaia/build_images.py benchmarks/multiswebench/build_images.py benchmarks/swebench/build_images.py benchmarks/swebenchmultilingual/build_images.py benchmarks/swebenchmultimodal/build_images.py benchmarks/swegym/build_images.py benchmarks/swesmith/build_images.py benchmarks/swtbench/build_images.py tests/test_image_utils.py
• uv run pytest tests/test_image_utils.py

[simonrosenberg]

Restored the workflow defaults to match the original main behavior while keeping the new knobs exposed:

• SWE-Bench: max-workers=12, build-batch-size=15
• SWT-Bench agent image builds: max-workers=4, build-batch-size=15
• SWT eval-env builds remain on their separate eval-env-build-batch-size default of 10

The PR still changes propagation, but no longer changes the default runtime behavior.

[all-hands-bot]

🟢 Good taste - Clean parameter threading that solves a real operational problem (CI build parallelism control). Straightforward precedence logic (explicit arg > env var) with test coverage. The force_build fix is a good catch. No fundamental issues found.

Verdict: ✅ Ship it.

Key insight: This is how you add operational controls without over-engineering - simple parameter threading with clear precedence.

[juanmichelini]

LGTM!