feat: enable custom openai endpoints for services like opper.ai

README.md

@@ -420,6 +420,18 @@ llm:
   timeout: 5m
 ```
 
+```yaml
+# OpenAI-compatible service, e.g. a gateway
+llm:
+  provider: openai
+  model: glm-5.3-flash
+  endpoint: https://eu.gw.opper.ai/openai
+  api_keys:
+    openai: op-...
+  timeout: 5m
+  rate_limit_rps: 10
+```
+
 ```yaml
 # Local Ollama
 llm:
@@ -438,6 +450,8 @@ llm:
 
 Remote providers (`anthropic`, `openai`) get a conservative rate limit applied automatically, biased below tier-1 ceilings so bursty operations like `sdd summarize --all` don't trip 429s. Override with `rate_limit_rps` on higher tiers.
 
+The `openai` provider talks to any service that implements the OpenAI chat completions protocol: set `endpoint` to its base URL and the request goes there instead of to OpenAI. The same URL works with or without a trailing `/v1`. Since the automatic rate limit assumes OpenAI's tier-1 ceilings, set `rate_limit_rps` explicitly for a service that allows more.
+
 ### Embedding provider (vector search)
 
 Two providers supported: `openai` and `ollama`. Configuring an embedding provider activates `sdd search --query` (semantic mode) and hybrid retrieval.

go.mod

@@ -92,4 +92,4 @@ require (
 // source and honor replace; only `go install pkg@version` (which sdd does not
 // use) would break. Remove this directive once the change lands upstream and
 // bump the require above to the released teilomillet/gollm version.
-replace github.com/teilomillet/gollm => github.com/networkteam/gollm v0.0.0-20260831213159-f6f84ac83bb5
+replace github.com/teilomillet/gollm => github.com/networkteam/gollm v0.0.0-20260919121334-aa9e2d1faebc

go.sum

@@ -107,8 +107,8 @@ github.com/modelcontextprotocol/go-sdk v1.7.0 h1:yqjY2dsbKAC0LSuWZVBMrHgiG8ukXv6
 github.com/modelcontextprotocol/go-sdk v1.7.0/go.mod h1:dL7u98E/zjJTGzEq+j30jQ8K2k1mb6LeAH4inEcSGts=
 github.com/muesli/cancelreader v0.2.2 h1:3I4Kt4BQjOR54NavqnDogx/MIoWBFa0StPA8ELUXHmA=
 github.com/muesli/cancelreader v0.2.2/go.mod h1:3XuTXfFS2VjM+HTLZY9Ak0l6eUKfijIfMUZ4EgX0QYo=
-github.com/networkteam/gollm v0.0.0-20260831213159-f6f84ac83bb5 h1:I6UUByvzpT4ExPlk/7RAtI+PnuF2rKEeQMYmJ+FeBXY=
-github.com/networkteam/gollm v0.0.0-20260831213159-f6f84ac83bb5/go.mod h1:ba4KWPPrYoc7PbPkcjTD6uMi2s5yXVWVsE5uzP+mvCg=
+github.com/networkteam/gollm v0.0.0-20260919121334-aa9e2d1faebc h1:3SID2v1Qy85u5mchX+wlNUctJLa5NjRWMfj8HkeXcAo=
+github.com/networkteam/gollm v0.0.0-20260919121334-aa9e2d1faebc/go.mod h1:ba4KWPPrYoc7PbPkcjTD6uMi2s5yXVWVsE5uzP+mvCg=
 github.com/networkteam/slogutils v0.4.0 h1:tf6Gf//H/a3EUmNszWEw8fhQkscrDdKt7HEdb2wBUhM=
 github.com/networkteam/slogutils v0.4.0/go.mod h1:j6siLHGfAkUpfhNUXw43EUdZv/HBaGd/RW7ToPwiJZo=
 github.com/philippgille/chromem-go v0.7.0 h1:4jfvfyKymjKNfGxBUhHUcj1kp7B17NL/I1P+vGh1RvY=

internal/llm/gollm/gollm.go

@@ -87,6 +87,10 @@ func NewRunner(cfg model.LLMConfig) (*Runner, error) {
 		}
 	}
 
+	if cfg.Provider == "openai" && cfg.Endpoint != "" {
+		opts = append(opts, upstream.SetOpenAIEndpoint(cfg.Endpoint))
+	}
+
 	// Enable Anthropic prompt caching — sends the anthropic-beta header so
 	// cache_control blocks on system prompts are honored server-side.
 	useCache := cfg.Provider == "anthropic"

internal/llm/gollm/wire_test.go

@@ -78,6 +78,61 @@ func TestOllamaRequestWire(t *testing.T) {
 	}
 }
 
+// The openai provider is the one whose target is configurable, so where a
+// request actually lands is a property worth asserting on the bytes: a base URL
+// only serves an OpenAI-compatible gateway if the path is appended to it.
+func TestOpenAIRequestWire(t *testing.T) {
+	var mu sync.Mutex
+	var body map[string]any
+	var path string
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		raw, _ := io.ReadAll(r.Body)
+		mu.Lock()
+		path = r.URL.Path
+		_ = json.Unmarshal(raw, &body)
+		mu.Unlock()
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"SERVED"}}],"usage":{"prompt_tokens":33,"completion_tokens":44}}`))
+	}))
+	defer srv.Close()
+
+	runner, err := gollmrunner.NewRunner(model.LLMConfig{
+		Provider: "openai", Model: "m", Endpoint: srv.URL,
+		APIKeys: map[string]string{"openai": "op-notarealkeyaaaaaaaaaa"},
+	})
+	if err != nil {
+		t.Fatalf("NewRunner: %v", err)
+	}
+	res, err := runner.Run(context.Background(), llm.Request{
+		SystemPrompt: "SYSTEM-BLOCK-MARKER",
+		UserPrompt:   "USER-BLOCK-MARKER",
+	})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+
+	mu.Lock()
+	defer mu.Unlock()
+	t.Logf("request keys: %v", keysOf(body))
+
+	if path != "/v1/chat/completions" {
+		t.Errorf("request path = %q, want /v1/chat/completions", path)
+	}
+	// A key OpenAI itself would reject must still reach a gateway that issued it.
+	sent, _ := json.Marshal(body["messages"])
+	for _, marker := range []string{"SYSTEM-BLOCK-MARKER", "USER-BLOCK-MARKER"} {
+		if !strings.Contains(string(sent), marker) {
+			t.Errorf("%s missing from the messages sent: %s", marker, sent)
+		}
+	}
+	if res.Text != "SERVED" {
+		t.Errorf("text = %q, want SERVED", res.Text)
+	}
+	if res.Usage.InputTokens != 33 || res.Usage.OutputTokens != 44 {
+		t.Errorf("usage not parsed: %+v", res.Usage)
+	}
+}
+
 func keysOf(m map[string]any) []string {
 	out := make([]string, 0, len(m))
 	for k := range m {

internal/model/config.go

@@ -243,6 +243,10 @@ type LLMConfig struct {
 	// Concurrency bounds the worker pool for batch operations. Zero means
 	// "use DefaultLLMConcurrency".
 	Concurrency int `yaml:"concurrency,omitempty"`
+	// Endpoint overrides the OpenAI-compatible base URL for the openai
+	// provider. Empty defaults to the OpenAI public endpoint. Useful for a
+	// gateway or a self-hosted service that implements the same protocol.
+	Endpoint string `yaml:"endpoint,omitempty"`
 	// OllamaEndpoint overrides the default Ollama URL for the gollm adapter.
 	OllamaEndpoint string `yaml:"ollama_endpoint,omitempty"`
 	// APIKeys maps provider name to API key. Never belongs in the
@@ -449,6 +453,9 @@ func mergeLLMConfig(base, overlay LLMConfig) LLMConfig {
 	if overlay.Concurrency != 0 {
 		out.Concurrency = overlay.Concurrency
 	}
+	if overlay.Endpoint != "" {
+		out.Endpoint = overlay.Endpoint
+	}
 	if overlay.OllamaEndpoint != "" {
 		out.OllamaEndpoint = overlay.OllamaEndpoint
 	}