fix: give DeepSeek V4 models a 1M last-resort context window (#226)

Flash and v4.1 ids were falling through to the V3 128K default, so trim
and the ctx gauge used 131k against a 1M-class model.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: jkyberneees

cmd/odek/serve_api_test.go

@@ -457,6 +457,25 @@ func TestHandleModelList_ReturnsOnlyConfiguredModel(t *testing.T) {
 	if models[0]["description"] == "" {
 		t.Error("description should be non-empty for a known model")
 	}
+	if ctx, ok := models[0]["max_context"].(float64); !ok || ctx != 1_000_000 {
+		t.Errorf("max_context = %v, want 1000000 (deepseek-v4 1M window)", models[0]["max_context"])
+	}
+}
+
+func TestHandleModelList_DeepSeekV41UsesV4Window(t *testing.T) {
+	handler := handleModelList("deepseek-v4.1-flash-expires-on-0910", nil)
+	w := httptest.NewRecorder()
+	handler(w, httptest.NewRequest(http.MethodGet, "/api/models", nil))
+	var models []modelEntry
+	if err := json.NewDecoder(w.Body).Decode(&models); err != nil {
+		t.Fatal(err)
+	}
+	if len(models) != 1 {
+		t.Fatalf("len = %d, want 1", len(models))
+	}
+	if models[0].MaxContext != 1_000_000 {
+		t.Errorf("v4.1 last-resort ctx = %d, want 1000000", models[0].MaxContext)
+	}
 }
 
 func TestHandleModelList_EmptyConfigModel_ReturnsEmptyList(t *testing.T) {

docs/MIGRATION.md

@@ -67,7 +67,7 @@ DeepSeek-only leftover: when `provider` is `deepseek`, `ODEK_API_KEY` → `DEEPS
 
 - Thinking: set `--thinking medium` (or `low`/`high`/`disabled`; `enabled` still works as an alias for `medium`).
 - Timeout: default **300s** for every model. Raise with `llm.request_timeout_seconds`.
-- Context window: `llm.context_window` → `ListModels` → last-resort table for shipped ids (`deepseek-v4-flash` 128K, `deepseek-v4-pro` 1M, GLM/Kimi/OpenAI prefixes) → else 0 (no trim).
+- Context window: `llm.context_window` → last-resort table for shipped ids (`deepseek-v4*` 1M, `deepseek-` 128K, GLM/Kimi/OpenAI prefixes) → `ListModels` → else 0 (no trim).
 
 `ProfileLabel` now returns the model id.
 

docs/PROVIDERS.md

@@ -71,8 +71,8 @@ window. **No auto-thinking and no auto-timeout.**
 
 | Prefix | Tokens |
 |---|---|
-| `deepseek-v4-pro` | 1M |
-| `deepseek-v4-flash` / `deepseek-` | 128K |
+| `deepseek-v4` (pro, flash, v4.1-*) | 1M |
+| `deepseek-` (v3 chat / reasoner) | 128K |
 | `glm-5.3` / `glm-5.2` | 1M |
 | `glm-5-turbo` | 200K |
 | `glm-` | 128K |

internal/llmclient/client.go

@@ -647,9 +647,8 @@ func LastResortContext(model string) int {
 		{"kimi-", 262_144},
 		{"k3-256k", 262_144},
 		{"k3", 1_000_000},
-		{"deepseek-v4-pro", 1_000_000},
-		{"deepseek-v4-flash", 131_072},
-		{"deepseek-", 131_072},
+		{"deepseek-v4", 1_000_000}, // pro, flash, v4.1-* (official 1M window)
+		{"deepseek-", 131_072},     // v3 chat / reasoner
 		// OpenAI — api.openai.com ListModels omits context_length.
 		{"gpt-6", 1_050_000},
 		{"gpt-5.6", 1_050_000},

internal/llmclient/client_test.go

@@ -78,8 +78,12 @@ func TestLastResortContext(t *testing.T) {
 		model string
 		want  int
 	}{
-		{"deepseek-v4-flash", 131_072},
+		{"deepseek-v4-flash", 1_000_000},
 		{"deepseek-v4-pro", 1_000_000},
+		{"deepseek-v4.1", 1_000_000},
+		{"deepseek-v4.1-flash-expires-on-0910", 1_000_000},
+		{"deepseek-chat", 131_072},
+		{"deepseek-reasoner", 131_072},
 		{"gpt-6-astra", 1_050_000},
 		{"gpt-6", 1_050_000},
 		{"gpt-5.6-luna", 1_050_000},