[wizard1203]

Migrates Zhenheng Tang's OpenAI BrowseComp adapter from harbor#1521, source d4ed8a1b806cf7bc9b8a7497ba7d4095135fc39f, to the independent package at src/browsecomp/. This preserves the original contribution, run configurations, historical metrics and evidence links, following the maintainer's migration request.

This adapter covers BrowseComp's 1,266 questions. BrowseComp-Plus is outside its scope. No new model, oracle, Docker/Daytona benchmark, or parity experiments were run for this migration.

Migration changes

• Import the original adapter in a separate commit, followed by standalone packaging and documentation updates.
• Package the existing instruction as task-template/instruction.md, expose output_dir, and add --split parity for the historical first 30 task IDs. Default generation retains its existing task IDs and behavior.
• Preserve the six existing task-template assets, including the verifier and oracle scripts, byte-for-byte. Add eight focused offline tests using synthetic encrypted rows and ensure the installed wheel includes every template.
• Align README/metadata with adapter contract v1, retain author and compute-support credits, and distinguish unpublished dataset status and historical reports from migration validation.

Historical evidence and review continuity

The original experiment date, three-run arrays, solver labels and metric summaries in parity_experiment.json are retained:

Historical comparison	Three 30-task run scores (%)	Mean ± sample SEM (%)
Original Claude-Code/Haiku	26.67, 20.00, 26.67	24.4 ± 2.2
Harbor Claude-Code/Haiku	23.33, 23.33, 20.00	22.2 ± 1.1

The separate archived oracle job records 1,266/1,266 tasks, mean reward 1.0 and zero errors. It predates later source revisions and is preserved as historical structural/verifier evidence, not a new validation of this head or model parity.

Preserved links:

• Original #1521 review history, including the May 14 bot review.
• HF parity PR #244, currently a public draft; existing reports/artifacts remain in its PR ref and were not copied into this code repository.
• Historical dataset PR #226, still open; this migration does not claim registry publication.
• Original-side fork.

The README documents the limits of these reports: original arguments specify a gpt-5-mini judge whereas Harbor configurations select Haiku; Harbor jobs contain 13/7/6 recorded errors, with scores based on 30 scheduled tasks; retry selection and original-side canonical samples/source are incomplete. The linked original fork currently lacks the documented custom runner and Dockerfile. Numerical range overlap is preserved without claiming that these protocol/reproducibility questions have been resolved.

Related BrowseComp work

harbor#2355 has already migrated to adapters#44. That contribution emphasizes a pinned simple-evals one-shot protocol and separately reports Codex browsing. This PR independently preserves #1521's Claude-Code browsing implementation and evidence, with cross-links so maintainers can coordinate the overlapping package scope and eventual consolidation. It does not attribute #44's implementation or results to this contribution.

Dataset publication handling discussed in harbor-datasets#246 remains a separate maintainer decision. No generated benchmark questions/answers or raw runtime configurations are added here.

Migration validation

Local Python 3.12.12 checks:

• uv run pytest tests/unit/: 8 passed, synthetic inputs only, with network access blocked by the test fixture.
• uv run ruff check --fix ., uv run ruff format ., and uv run ty check: passed from the adapter package directory. Historical task-template assets are excluded from rewriting to preserve their bytes.
• Compared old and migrated generators on three synthetic questions: all 27 generated files byte-identical; all six pre-existing template assets unchanged.
• uv build --wheel: passed; all seven template assets included. Installed the wheel in a separate temporary environment with no index access and generated a complete synthetic task through its console entry point.
• python3 scripts/validate_adapter.py browsecomp: 31 passed, 0 errors, 1 warning for unknown historical parity costs (null).
• Both reference YAMLs parse; documented external Harbor command flags were checked against local Harbor 0.1.36. Historical date/metrics/raw arrays and source TOML parsing were checked separately.

These are offline packaging/transformation checks, separate from the historical benchmark reports above.