[minixalpha]

Add the official DeepSeek Harness sdk-minimal agent to the standard Harbor workflow and evaluate it on the same 34 development tasks as Pi. Pin the offline SDK/runtime distribution, preserve native model behavior and task limits, and confirm process completion before verification. Add reviewed dependency profiles, a PyPI wheel cache verified against current original metadata, and a keyless SDK service-lifecycle probe.

The original run produced 22 passes, 6 failures, and 6 unscored errors. A separately authorized six-task supplement produced 4 passes, 2 failures, and no unscored errors. Preserve both records: the original 28 accepted attempts plus six new attempts provide mixed coverage of 26 passes and 8 failures, not an independent benchmark or best-of-two score.

KV/PyPI failures retain the SDK shutdown caveat; Pi's pipeline pass retains reference-solution exposure. Different dates, endpoints, generation budgets and preparation prevent causal harness attribution.

Validation: 301 Harbor tests, including 16 wheel-cache contracts; native uv 0.9.5 transport smoke; confirmed completion boundaries; unchanged original task artifacts and 2,506 frozen files; 590 supplemental files frozen before trajectory review. Both complete report pairs and the Harbor READMEs are synchronized. Add the workflow under 0.10.x [Unreleased]; root READMEs, dev notes and research notes are unchanged.

• Original report: English · 简体中文
• Supplement: English · 简体中文
• Supplement registration and evidence