feat: add native DeepSeek provider by zhujian7 · Pull Request #11 · genai-io/gen-code

zhujian7 · 2026-05-15T02:15:56Z

Summary

Add native DeepSeek API provider (api.deepseek.com) using OpenAI-compatible SDK
Support deepseek-v4-flash and deepseek-v4-pro models with reasoning_effort extraction
Include model catalog with token limits and USD pricing, with cost tracking wired into model.go

Test plan

Build passes
All new unit tests pass (model list error propagation, stream request, cost estimation, thinking support)
Existing provider tests unaffected
Manual end-to-end test with valid DEEPSEEK_API_KEY

🤖 Generated with Claude Code

Add native DeepSeek API provider (api.deepseek.com) using the OpenAI-compatible SDK pattern. Supports deepseek-chat (V3) and deepseek-reasoner (R1) models with reasoning_content extraction for the reasoner model. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> Signed-off-by: zhujian <jiazhu@redhat.com>

zhujian7 · 2026-05-15T02:22:43Z

/hold

- Fix llm.IsReady to use secret.Resolve instead of os.Getenv, so provider readiness checks are consistent with the UI layer (RenderEnvVarStatus and providerIsEnvReady). This fixes the API key input prompt not appearing. - Update DeepSeek catalog to V4 models (deepseek-v4-flash, deepseek-v4-pro) with 1M context and current pricing; keep legacy chat/reasoner as aliases. - Default model changed to deepseek-v4-flash. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> Signed-off-by: zhujian <jiazhu@redhat.com>

zhujian7 · 2026-05-15T02:39:05Z

/hold cancel

yanmxa

Six inline comments on the items I think need addressing before merge. Full review context in the PR conversation.

yanmxa · 2026-05-15T09:24:16Z

+var catalog = []modelCatalogEntry{
+	{
+		info: llm.ModelInfo{
+			ID:               "deepseek-v4-flash",


PR description says V3/R1, but the catalog ships deepseek-v4-flash / deepseek-v4-pro as primary entries with 1M input / 384K output limits and a specific deprecation date for the legacy names (line 38). These IDs aren't in DeepSeek's published lineup — please cite the source for the V4 family + pricing, or drop the V4 entries and keep only deepseek-chat / deepseek-reasoner.

DeepSeek V4 launched April 2025 and is the current production lineup. deepseek-v4-flash and deepseek-v4-pro are the canonical model IDs per https://api-docs.deepseek.com/quick_start/pricing. The legacy names (deepseek-chat, deepseek-reasoner) will be deprecated July 24, 2026 — they currently map to V4 Flash on the backend. Updated the PR description accordingly.

the PR description updated, we still use the deepseek-v4-flash / deepseek-v4-pro as "The model names deepseek-chat and deepseek-reasoner will be deprecated on 2026/07/24"

and I removed the deepseek-chat / deepseek-reasoner, because when I call the deepseek list models api: these 2 are removed:

╰─$ curl -s https://api.deepseek.com/models -H "Authorization: Bearer $DEEPSEEK_API_KEY" {"object":"list","data":[{"id":"deepseek-v4-flash","object":"model","owned_by":"deepseek"},{"id":"deepseek-v4-pro","object":"model","owned_by":"deepseek"}]}%

yanmxa · 2026-05-15T09:24:16Z

+	return llm.ModelInfo{}, false
+}
+
+func EstimateCost(modelID string, usage llm.Usage) (llm.Money, bool) {


EstimateCost is dead code — internal/app/model.go only switches on llm.MinMax for cost roll-up, nothing calls this. Either add a case llm.DeepSeek: branch in model.go, or remove EstimateCost + the per-entry pricing data + TestDeepSeekEstimateCost from this PR.

yanmxa · 2026-05-15T09:24:16Z

+// V4 models (deepseek-v4-*) and legacy deepseek-reasoner all support thinking.
+func supportsThinking(model string) bool {
+	lower := strings.ToLower(model)
+	return strings.Contains(lower, "reasoner") || strings.Contains(lower, "v4")


Allow-list defaults unknown IDs to non-thinking. BigModel uses the opposite shape (bigmodel/client.go:58-68): switch on known non-thinking IDs and default true. As written, every future DeepSeek model silently loses the thinking-effort UI until someone updates this string match. Recommend flipping to a deny-list.

yanmxa · 2026-05-15T09:24:16Z

+func (c *Client) ListModels(ctx context.Context) ([]llm.ModelInfo, error) {
+	page, err := c.client.Models.List(ctx)
+	if err != nil {
+		return StaticModels(), nil


Silently swallowing the API error and returning the static catalog masks real failures (bad/expired key, network) — user sees a populated model list, then gets a confusing 401 mid-stream. Moonshot/BigModel propagate the error; please match. (Also lets you drop the static fallback entirely, consistent with the dynamic-list pattern.)

yanmxa · 2026-05-15T09:24:16Z

+func NewAPIKeyClient(ctx context.Context) (llm.Provider, error) {
+	baseURL := os.Getenv("DEEPSEEK_BASE_URL")
+	if baseURL == "" {
+		baseURL = "https://api.deepseek.com"


Missing /v1. The openai-go SDK appends /chat/completions, so this resolves to https://api.deepseek.com/chat/completions — DeepSeek's documented endpoint is /v1/chat/completions. Moonshot/BigModel both include /v1 in the default. Suggest https://api.deepseek.com/v1.

Per the DeepSeek API docs, both /chat/completions and /v1/chat/completions are valid endpoints. The base URL https://api.deepseek.com is the documented default. Additionally, the Models.List endpoint is at /models — adding /v1 to the base URL would break it to /v1/models. Given both paths work for chat and the models endpoint requires the non-v1 path, leaving the base URL as-is is safer.

https://api-docs.deepseek.com/guides/anthropic_api

yanmxa · 2026-05-15T09:24:16Z

 	case "bigmodel":
 		return "glm-5.1"
+	case "deepseek":
+		return "deepseek-v4-flash"


Default depends on deepseek-v4-flash actually existing at the API. If the V4 IDs aren't real (see catalog.go comment), every fresh user picks a model the API will reject. Recommend deepseek-chat as a safer default until V4 ships.

DeepSeek V4 is the current production lineup (launched April 2025). deepseek-v4-flash is the recommended default — it is the canonical model ID per the API docs. The legacy deepseek-chat name will be deprecated July 24, 2026.

- Wire up DeepSeek cost estimation in model.go - Remove legacy deepseek-chat/deepseek-reasoner (API only returns V4) - Propagate ListModels error instead of silently falling back - Simplify supportsThinking (all V4 models support reasoning_effort) Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> Signed-off-by: zhujian <jiazhu@redhat.com>

zhujian7 · 2026-05-15T14:05:55Z

@yanmxa PTAL

yanmxa reviewed May 15, 2026

View reviewed changes

yanmxa merged commit b7dca93 into genai-io:main May 15, 2026
1 check passed

Conversation

zhujian7 commented May 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Uh oh!

zhujian7 commented May 15, 2026

Uh oh!

zhujian7 commented May 15, 2026

Uh oh!

yanmxa left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhujian7 commented May 15, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

zhujian7 commented May 15, 2026 •

edited

Loading