PR #70963: feat(deepseek): add deepseek-v4-pro and deepseek-v4-flash to model catalog [AI-assisted]

• Repository: openclaw/openclaw
• Author: 294898950a-glitch
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/70963

Description (problem / solution / changelog)

Summary

Adds the two current-generation DeepSeek models to the bundled DeepSeek provider catalog:

• deepseek-v4-pro — reasoning model, 1M context window, 384K max output
• deepseek-v4-flash — reasoning model, 1M context window, 384K max output

DeepSeek's /models endpoint now returns only these two IDs. The legacy deepseek-chat and deepseek-reasoner entries are kept (backwards compatibility — legacy aliases still resolve on the upstream API), but without this change users cannot select V4 through openclaw (it falls back to openai/deepseek-v4-pro, which is not what users want).

Changes

Pricing source: https://api-docs.deepseek.com/quick_start/pricing

Files changed:

• extensions/deepseek/models.ts — add V4 pricing constants and catalog entries
• extensions/deepseek/index.test.ts — extend static-catalog assertion with V4 ids and attribute checks

Validation

Live API

/models endpoint confirms the two IDs:

{"object":"list","data":[
  {"id":"deepseek-v4-flash","object":"model","owned_by":"deepseek"},
  {"id":"deepseek-v4-pro","object":"model","owned_by":"deepseek"}
]}

deepseek-v4-pro — live request/response (returns reasoning_content, confirming reasoning: true):

Request:  {"model":"deepseek-v4-pro","messages":[{"role":"user","content":"What is 2+2? Answer in one word."}],"max_tokens":50}
Response: {"role":"assistant","content":"","reasoning_content":"We are asked: \"What is 2+2?\" ... the answer is \"four\"..."}
usage:    {"prompt_tokens":16,"completion_tokens":50,"completion_tokens_details":{"reasoning_tokens":50}}
system_fingerprint: "fp_9954b31ca7_prod0820_fp8_kvcache_20260402"

deepseek-v4-flash — live request/response:

Request:  {"model":"deepseek-v4-flash","messages":[{"role":"user","content":"What is 2+2? Answer in one word."}],"max_tokens":50}
Response: {"role":"assistant","content":"Four","reasoning_content":"We need to answer \"What is 2+2?\" ... The answer is obviously four..."}
usage:    {"prompt_tokens":16,"completion_tokens":33,"completion_tokens_details":{"reasoning_tokens":31}}
system_fingerprint: "fp_058df29938_prod0820_fp8_kvcache_20260402"

Local checks

$ pnpm test extensions/deepseek
Test Files  1 passed (1)
Tests  3 passed (3)

$ pnpm format:check extensions/deepseek
All matched files use the correct format.

$ node_modules/.bin/oxlint extensions/deepseek --tsconfig tsconfig.oxlint.extensions.json
Found 0 warnings and 0 errors.

$ pnpm tsgo:extensions / pnpm tsgo:extensions:test
both pass (37s / 123s)

pnpm check:changed full run also passes the typecheck and conflict-marker gates; the tree-wide oxlint step OOMs on my WSL2 dev box (7.6GB RAM) but is unrelated to this diff.

Known follow-ups (not part of this PR)

• Multi-turn sessions: these V4 models enforce DeepSeek's new "thinking mode" contract, which requires reasoning_content to be replayed in subsequent assistant turns. openclaw's current openai-completions replay path drops that field, so multi-turn sessions fail with 400 The \reasoning_content` in the thinking mode must be passed back to the API`. That is tracked in #70931 and needs a separate PR (likely in the transcript → completions message transform for reasoning-capable openai-completions providers). Single-turn requests work against the newly-catalogued models.
• tieredPricing: DeepSeek's pricing page does not currently show context-length tiers for V4; leaving tieredPricing unset. If V4 gains tiered input rates this should be revisited.

AI-assisted disclosure

• Mark as AI-assisted: yes, Claude Opus 4.7
• Degree of testing: fully tested — local unit tests pass, live API requests/responses captured above, pricing verified against official docs
• Prompts / session logs: happy to share on request
• Confirm understanding of code: yes
• `codex review --base origin/main` — not available in my environment; will address any review-bot comments on this PR

Test plan

• pnpm test extensions/deepseek — passes
• pnpm format:check extensions/deepseek — passes
• oxlint extensions/deepseek — 0 errors
• pnpm tsgo:extensions — passes
• pnpm tsgo:extensions:test — passes
• Live /models listing + per-model chat completion requests — captured above

Changed files

• extensions/deepseek/index.test.ts (modified, +10/-0)
• extensions/deepseek/models.ts (modified, +38/-0)