Add TMax v1 taskset

[rasdani]

Summary

• Add built-in tmax-v1 under verifiers.v1.tasksets as a narrow Harbor specialization pinned to tmax/TMax-15K-Harbor@latest.
• Require an explicit [taskset.image_map] entry for each selected TMax task and set the mapped Prime image on the loaded Harbor task.
• Add configs/tmax-v1-smoke.toml with the five pushed public Prime smoke images and document Harbor as the runnable source.

Validation

• uv run ruff check verifiers/v1/tasksets/tmax_v1/__init__.py verifiers/v1/tasksets/tmax_v1/taskset.py verifiers/v1/tasksets/__init__.py
• uv run python -m py_compile verifiers/v1/tasksets/tmax_v1/__init__.py verifiers/v1/tasksets/tmax_v1/taskset.py verifiers/v1/tasksets/__init__.py
• uv lock --check
• uv run --python 3.13 ty check verifiers
• git diff --check HEAD~1..HEAD
• uv run eval @ configs/tmax-v1-smoke.toml --taskset.tasks task_000004_aeddda76 -m openai/gpt-5.5 -n 1 -r 1 -c 1 --rich false -o outputs/smokes/tmax-v1-20260623T211516Z
  • results.jsonl: tmax/task_000004_aeddda76, solved=1.0, completed, no errors

No unit tests were added, per the requested scope.

---

Note

Low Risk
Additive built-in taskset and docs; runtime behavior is inherited from harbor-v1 with explicit image injection and clear errors for missing mappings.

Overview
Adds a built-in tmax-v1 taskset that specializes harbor-v1 for the TMax Harbor dataset tmax/TMax-15K-Harbor@latest. TMaxConfig pins that dataset, sets ignore_dockerfile so Dockerfile-only tasks are not rejected, and requires [taskset.image_map] entries keyed by Harbor task directory id. TMaxTaskset.load_tasks loads Harbor tasks then injects the mapped Prime image on each row, failing fast if a selected task has no mapping.

Package exports and docs (GUIDE.md, README.md) describe Harbor as the execution source and the image-map workflow. configs/tmax-v1-smoke.toml wires five smoke task ids to public Prime images with bash harness on prime runtime.

^{Reviewed by Cursor Bugbot for commit 6201f27. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add TMax v1 taskset backed by Harbor with per-task Prime image mapping

• Adds TMaxTaskset and TMaxConfig in taskset.py, a HarborTaskset subclass that loads tasks from tmax/TMax-15K-Harbor@latest and injects a Prime image per task from a required image_map.
• Missing image_map entries cause a hard failure at load time, so all selected tasks must have a mapped image before the taskset can run.
• Adds a smoke config in configs/tmax-v1-smoke.toml with 5 tasks, 1 rollout, and max_concurrent=1.
• Updates verifiers/v1/tasksets/__init__.py to export TMaxConfig and TMaxTaskset, and adds documentation in the v1 README and GUIDE.

^{Macroscope summarized 6201f27.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR adds a new tmax-v1 taskset capability with new Python classes and configuration. While the implementation cleanly extends existing Harbor abstractions, introducing new user-facing features warrants human review.

^{You can customize Macroscope's approvability policy. Learn more.}