[hanhainebula]

Summary

Make GPU selection available through the common evaluation command, refresh the coding-agent runtimes, and give all 49 existing task packages a consistent README and version baseline. Update the evaluation guides, contribution guidance, task scaffolding, and maintainer checks to match.

GPU selection

• Add optional --gpu-device <host-index-or-GPU-UUID> for all 23 GPU tasks. Resolve the device once and use the same UUID for the agent and separate verifier.
• Generate per-run Compose configurations that replace the GPU reservation while preserving mounts, network isolation, and cleanup behavior. Omitting the option retains the default single-GPU request.
• Move query-encoder-alignment onto the common interface and document selection in the English/Chinese READMEs, evaluation guides, and GPU task READMEs.

For example, select host GPU 2 using the existing evaluation command:

bash scripts/run_task.sh --task code-embedding-finetuning --agent codex \
  --gpu-device 2 --output jobs/code-embedding-finetuning-gpu2

Harness updates

• Pin Codex to 0.164.0, Claude Code to 2.1.309, and Pi to 1.4.0 across the launcher, adapters, all 49 agent images, documentation, and tests. Centralize the launcher pins in scripts/harness_versions.py and support Pi's max thinking level.
• Upgrade the Codex runtime used by all 44 RewardKit judge integrations to 0.164.0, including verifier installation and audit-user cache paths.
• Keep Harbor at 0.22.0 and RewardKit at 0.2.0. The RewardKit wrapper uses a version-guarded override of its internal Codex pin; future RewardKit upgrades require revalidating that integration.

Task documentation

• Standardize every task README around its title and five sections: Task Metadata, Change Log, Difficulty Explanation, Verification Explanation, and References.
• Set all task versions to 1.0.0 and record 2026-10-23 | 1.0.0 | Added initial task. Remove duplicate legacy version declarations from the README prose. Preserve historical validation reports and their original versions, scores, and dates.
• Retain task-specific constraints, scoring details, usage instructions, and existing references. Correct stale README status descriptions using existing validation records.
• Update the scaffolder, authoring/validation guidance, contribution documentation, and maintainer audit checklist so future tasks follow the same format.

Validation evidence

• python scripts/check_tasks.py: 0 errors.
• python -m unittest discover -s scripts/tests -p 'test_*.py': 181 tests, 179 passed, 2 skipped because RewardKit is absent from the host environment. The RewardKit integration was separately checked in an existing verifier container.
• python .agents/skills/create-searchswe-task/scripts/test_scaffold_task.py: 5 passed. Both task skills passed skill validation.
• All 49 task README structures, versions, and initial changelog entries passed consistency checks; 562 local Markdown links/anchors were checked. git diff --check passed.
• Rendered and compared all 46 GPU phase configurations. Device selection preserved other Compose settings. During implementation, author-run minimal agent/verifier containers on GPUs 6 and 7 received the selected UUIDs; CUDA/NVML enumeration and a scalar CUDA operation succeeded. Complete successful stdout and docker inspect captures were not retained in the review evidence bundle, so these observations are distinct from the independently reproduced Compose checks.
• Each upgraded coding CLI completed a minimal call against a local mock API and exited successfully. Pi output converted successfully to Harbor's trajectory format. RewardKit with the upgraded Codex parsed a structured result and used the expected versioned cache.

Validation was intentionally limited to static checks, automated tests, and minimal runtime/protocol checks. Full task evaluations were not run. Real external model API connectivity remains unverified because credentials were unavailable.

The documentation follow-up removes two stale inline version declarations in decision-model-training and lightweight-query-retrieval. All 49 task README versions, section structures, and initial changelogs were rechecked against task.toml; the follow-up diff contains only those two removals and passes git diff --check. Historical trial text and reports remain byte-for-byte unchanged. Runtime tests were not repeated for this documentation-only correction.

Contributor checklist

• Shared infrastructure changes and all affected existing task packages are included.
• Task authors, instructions, fixed inputs, asset manifests, and original validation reports are preserved.
• Changes exclude secrets, downloaded assets, and job outputs.
• Validation evidence and its limits are recorded.

Maintainer merge checklist

• Review GPU selection across both phases and the RewardKit version guard.
• Confirm required checks against the final PR head and current main.
• Merge with a merge commit to preserve contributor history.