Sync harbor-adapter-creator and harbor-task-creator with upstream changes (2026-04-19)

[benediktstroebl]

Syncs two skills with upstream changes merged into harbor-framework/harbor in the last 48 hours, primarily from commit 9ad34d5 (Split adapter tutorial to human/ai, update registry handling) and related adapter PRs.

Skills updated

harbor-adapter-creator

• Directory structure: Updated from flat layout to src layout. Adapter code now lives under src/<adapter_name>/ with adapter.py, main.py, and task-template/ (was template/). Added pyproject.toml, .python-version, __init__.py to the structure diagram.
• Run command: Updated from python3 adapters/mybenchmark/run_adapter.py to uv run python -m <adapter_name>.main, matching the new package-based layout.
• task.toml: Added required [task] section with name field. The upstream fix (commit 9ad34d5) explicitly documents that name must be under [task], not at the TOML top level — tasks without it cannot be registered. Added naming requirements guidance.
• parity_experiment.json: Renamed original_trials/harbor_trials to original_runs/harbor_runs throughout. Added new required fields: adapted_benchmark_size, parity_benchmark_size, number_of_runs. Updated the field reference table and example. The validator on main now enforces *_runs naming.
• adapter_metadata.json: Added added_agents and parity_unmatching_agents fields to harbor_adapter entries. Updated parity_costs format to include $ prefix.
• Reference adapters: Added scenario table covering cooperbench (multi-agent workflow via messaging/sidecars) and featurebench (GPU tasks, comprehensive Docker + Modal example), plus adebench, evoeval, bixbench, financeagent, medagentbench.
• GPU tasks: Added dedicated section pointing to featurebench as the canonical GPU adapter example.
• Gotchas: Added entries for missing [task] section and unstable task names.

harbor-task-creator

• task.toml: Added [task] section to the minimal example with name, description, and keywords fields. This matches the upstream fix for the task.toml format where name must be under [task], not at the top level.
• Keywords guidance: Added "Always populate keywords" note — pick 3–8 lowercase tokens covering domain, verifier style, and hardware. Leaving it empty makes the task invisible to registry search.
• Step 9: Update README.md: Added new section documenting what to put in README before publishing (agent description, environment, verifier, layout, run commands).
• Common Pitfalls: Added rows for keywords = [], missing [task] section, and stub README.

Skills not updated

harbor-cli: No confirmed CLI command changes in the 48-hour window that would affect the reference. Other merged PRs (llmsr-bench adapter, ScienceAgentBench adapter, RExBench Modal support) add new adapters but do not change CLI commands or flags.