[jithinAB]

Adds two hosted DeepSeek rows to the shipped catalogue, so an agent can pin model: deepseek-flash or deepseek-v4-pro, a needs: line can rank them, and cost < lines can be held against them. Bud Meta Harness binds DeepSeek by default; until now it had to write these rows into each workspace catalogue.

Every fact is read from DeepSeek's own API documentation on 2026-11-11 (pricing page and thinking-mode guide), cited in the rows:

• Price: DeepSeek publishes peak and off-peak rates (off-peak is half) plus cache-hit input rates. A row holds one price and a cost < line must never pass on a figure a run can exceed, so cost: records the peak cache-miss rates; off-peak and cache-hit rates are written beside it.
• Context window: 1,048,576 tokens for both.
• Modality: image input on deepseek-flash only.
• B5: both think unless told (thinks-unless-told: yes); temperature and both penalties have no effect while thinking, and top-p too on v4-pro. On flash, top-p takes effect only while thinking, which ignores-while-thinking cannot express, so it is left out instead of mis-stated.
• Aliases: the legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names, which the API routes to flash.
• tool-calling: single and reasoning: unknown, by the same rule the Gemini and Grok rows use.

Evidence (local):

• cargo build --release -p pact-cli --locked
• A tree pinning deepseek-v4-pro loads. Adding settings: {temperature: 0.2} is refused by loader/a-setting-ignored-while-it-thinks. The legacy deepseek-v4-flash pin resolves, and an unknown name is refused.
• resolve.price_of gives 1.5 USD (flash) and 5.28 USD (v4-pro) per 1M+1M tokens; window_of gives 1,048,576.
• pytest on the catalogue, pin, portability and model-choosing tests: 13 passed, 42 skipped.

This is BMH plan card UP-1 (PACT v97 plan §8.4). budmh pact catalog adopt writes the same rows into workspace catalogues until a re-pin brings this in.

🤖 Generated with Claude Code