feat: Complete DeepSeek API adapter implementation

[dixoxib]

Description

Enhanced DeepSeek API adapter from a minimal 73‑line implementation to a complete, production‑ready adapter with all features expected from a modern LLM provider integration following official docs.

Key improvements:

• Added streaming & non‑streaming chat completions with reasoning support
• Implemented tool calling compatibility and thinking tool-call chains
• Implemented model listing via DeepSeek's /models endpoint
• Enhanced FIM beta endpoint
• Added type safety with comprehensive TypeScript types
• Improved converting with detailed warnings
• Added prefix completion support (beta chat endpoint)
• Added workarounds for DeepSeek Reasoner edge cases
• Improved documentation, configuration and onboarding
• minimal invasive to codebase

Files changed: 35 files, including core implementation, tests, documentation, and configuration updates, small fixes.

AI Code Review

• Team members only: AI review runs automatically when PR is opened or marked ready for review
• Team members can also trigger a review by commenting @continue-review

Checklist

• I've read the contributing guide
• The relevant docs have been updated or created
• The relevant tests have been updated or created

Screen recording or screenshot

Key demonstrations:

1. Onboarding
2. DeepSeek Reasoner showing reasoning_content in responses
3. Thinking-Tool-calling-Chains
4. Non-streaming summarize functionality

sv.mp4

Tests

Added/updated tests:

• packages/openai-adapters/src/test/deepseek-converters.test.ts – Comprehensive unit tests for converter functions
• core/llm/llms/DeepSeek.unit.test.ts – Unit tests for DeepSeek provider class
• core/llm/llms/DeepSeek.vitest.ts – Integration tests
• packages/openai-adapters/src/test/DeepSeekApi.test.ts – API adapter test
• core/llm/llms/DeepSeek.tools.test.ts – Thinking tool chain test

The enhancement addresses limitations in the current minimal implementation and enables full DeepSeek functionality including Agent mode with tool calling for Reasoner model.

---

I have read the CLA Document and I hereby sign the CLA

---

Continue Tasks: 🔄 7 running — View all

---

Summary by cubic

Complete DeepSeek provider with chat, reasoning, tool calling, and FIM (beta) support. Adds onboarding/docs, streaming usage and cost, parallel generation, strict token reservation, exact apiBase handling, and assistant‑only streaming/accumulation.

• New Features

  • DeepSeek provider: streaming/non‑streaming chat, reasoning_content, thinking→tool call pairing, tool choice, and model autodetect with OS‑models edit template.
  • FIM via deepseek-fim-beta only; supportsFim is true for the FIM beta (chat/reasoner do not report FIM).
  • Token/cost handling: reserve full maxTokens for DeepSeek; improved pruning; streamed usage stats; DeepSeek pricing with cache hit/miss.
  • Stability/setup: only assistant chunks are streamed/accumulated; apiBase preserved exactly and defaulted; added llm‑info entries, onboarding/UI and VS Code updates; adapter/converter types and tests.

• Migration

  • Use provider: deepseek with models: deepseek-chat, deepseek-reasoner, deepseek-fim-beta (for FIM).
  • apiBase defaults to https://api.deepseek.com (preserved exactly); set maxTokens (chat: 8k, reasoner: 32k).
  • Replace legacy Deepseek class or deepseek-coder with the new DeepSeek class and models; provider ID remains deepseek.

^{Written for commit e85216e. Summary will update on new commits.}

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

1 out of 2 committers have signed the CLA.
✅ (dixoxib)[https://github.com/dixoxib]
❌ @ci Bot
CI Bot seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}

[cubic-dev-ai]

6 issues found across 35 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="core/config/yaml/models.ts">

<violation number="1" location="core/config/yaml/models.ts:55">
P1: Forcing `contextLength` to `DEFAULT_CONTEXT_LENGTH` in YAML model conversion can override provider runtime autodetection (e.g., Ollama), causing incorrect context windows for unknown/autodetected models.</violation>
</file>

<file name="packages/openai-adapters/src/types.ts">

<violation number="1" location="packages/openai-adapters/src/types.ts:63">
P1: `deepseek` was added to `OpenAIConfigSchema` even though `DeepseekConfigSchema` already handles that discriminator in `LLMConfigSchema`, creating a duplicate discriminator value in the same `z.discriminatedUnion`.</violation>
</file>

<file name="packages/openai-adapters/src/apis/DeepSeek.ts">

<violation number="1" location="packages/openai-adapters/src/apis/DeepSeek.ts:317">
P2: FIM completions are beta-only and require the beta base URL. Building the endpoint as `/completions` against the default base (`https://api.deepseek.com/`) no longer targets the beta API, so FIM will fail unless callers override apiBase manually.</violation>
</file>

<file name="packages/openai-adapters/src/util/deepseek-converters.ts">

<violation number="1" location="packages/openai-adapters/src/util/deepseek-converters.ts:327">
P2: `top_logprobs` is included even when `logprobs` is false/undefined, but DeepSeek requires `logprobs=true` whenever `top_logprobs` is used. This can generate invalid requests that the API rejects.</violation>
</file>

<file name="core/llm/llms/DeepSeek.ts">

<violation number="1" location="core/llm/llms/DeepSeek.ts:73">
P2: Non-streaming chat ignores top-level options.model, unlike streamChat/streamFim, causing inconsistent model selection.</violation>

<violation number="2" location="core/llm/llms/DeepSeek.ts:217">
P2: supportsPrefill() reports true unconditionally even though DeepSeek prefill is beta-only. This misreports capability on non‑beta API bases and can enable beta-only prompt paths against non‑beta endpoints.</violation>
</file>

---

Since this is your first cubic review, here's how it works:

• cubic automatically reviews your code and comments on bugs and improvements
• Teach cubic by replying to its comments. cubic learns from your replies and gets better over time
• Add one-off context when rerunning by tagging @cubic-dev-ai with guidance or docs links (including llms.txt)
• Ask questions if you need clarification on any suggestion

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[dixoxib]

I have read the CLA Document and I hereby sign the CLA

[cubic-dev-ai]

2 issues found across 2 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="packages/openai-adapters/src/apis/DeepSeek.ts">

<violation number="1" location="packages/openai-adapters/src/apis/DeepSeek.ts:50">
P2: apiBase trailing-slash normalization was removed; relative URL construction can drop path segments for custom bases like https://host/v1, misrouting requests.</violation>

<violation number="2" location="packages/openai-adapters/src/apis/DeepSeek.ts:316">
P2: FIM is a beta feature that requires the beta base URL; building the endpoint as `new URL("completions", this.apiBase)` sends FIM requests to the non‑beta `/completions` path when using the default base URL, which breaks FIM. Restore beta-aware routing or force the beta base URL for FIM calls.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}