feat: add model-aware context budget calculation for status bar warnings

web/src/chat/modelConfig.ts

@@ -0,0 +1,28 @@
+import type { ModelMode } from '@/types/api'
+
+/**
+ * Context windows vary by model/provider and may change over time.
+ *
+ * The UI only needs this to compute a conservative "context remaining" warning.
+ * We intentionally keep a headroom budget to avoid false confidence near the limit
+ * (system prompts, tool overhead, and other hidden tokens can consume extra space).
+ *
+ * If/when the server provides an explicit per-session context limit, prefer that
+ * and use this only as a fallback.
+ */
+const CONTEXT_HEADROOM_TOKENS = 10_000
+
+const MODEL_CONTEXT_WINDOWS: Record<NonNullable<ModelMode>, number> = {
+    // Claude Code modes used in this app; currently treated as ~200k context.
+    default: 200_000,
+    sonnet: 200_000,
+    opus: 200_000
+}
+
+export function getContextBudgetTokens(modelMode: ModelMode): number | null {
+    const mode: NonNullable<ModelMode> = modelMode ?? 'default'
+    const windowTokens = MODEL_CONTEXT_WINDOWS[mode]
+    if (!windowTokens) return null
+    return Math.max(1, windowTokens - CONTEXT_HEADROOM_TOKENS)
+}
+