fix(agent): correct media token arithmetic and tool call double-counting

Two estimation bugs fixed:

1. Media tokens were added to the chars accumulator before the chars*2/5
   conversion, resulting in 256*2/5=102 tokens per item instead of 256.
   Fix: add media tokens directly to the final token count, bypassing
   the character-based heuristic.

2. estimateMessageTokens counted both tc.Name and tc.Function.Name for
   tool calls, but providers only send one (OpenAI-compat uses
   function.name, Anthropic uses tc.Name). Fix: count tc.Function.Name
   when Function is present, fall back to tc.Name only otherwise.

Also fix i18n hint text: "auto-detect" was misleading — the backend
uses a 4x max_tokens heuristic, not actual model detection.

pkg/agent/context_budget.go

@@ -95,10 +95,14 @@ func estimateMessageTokens(msg providers.Message) int {
 	}
 
 	for _, tc := range msg.ToolCalls {
-		// Count tool call metadata: ID, type, function name
-		chars += len(tc.ID) + len(tc.Type) + len(tc.Name)
+		chars += len(tc.ID) + len(tc.Type)
 		if tc.Function != nil {
+			// Count function name + arguments (the wire format for most providers).
+			// tc.Name mirrors tc.Function.Name — count only once to avoid double-counting.
 			chars += len(tc.Function.Name) + len(tc.Function.Arguments)
+		} else {
+			// Fallback: some provider formats use top-level Name without Function.
+			chars += len(tc.Name)
 		}
 	}
 
@@ -106,17 +110,20 @@ func estimateMessageTokens(msg providers.Message) int {
 		chars += len(msg.ToolCallID)
 	}
 
-	// Media items (images, files) are serialized by provider adapters into
-	// multipart or image_url payloads. Use a fixed per-item estimate since
-	// actual token cost depends on resolution and provider tokenization.
-	const mediaTokensPerItem = 256
-	chars += len(msg.Media) * mediaTokensPerItem
-
 	// Per-message overhead for role label, JSON structure, separators.
 	const messageOverhead = 12
 	chars += messageOverhead
 
-	return chars * 2 / 5
+	tokens := chars * 2 / 5
+
+	// Media items (images, files) are serialized by provider adapters into
+	// multipart or image_url payloads. Add a fixed per-item token estimate
+	// directly (not through the chars heuristic) since actual cost depends
+	// on resolution and provider-specific image tokenization.
+	const mediaTokensPerItem = 256
+	tokens += len(msg.Media) * mediaTokensPerItem
+
+	return tokens
 }
 
 // estimateToolDefsTokens estimates the total token cost of tool definitions

pkg/agent/context_budget_test.go

@@ -503,6 +503,13 @@ func TestEstimateMessageTokens_MediaItems(t *testing.T) {
 		t.Errorf("message with Media (%d tokens) should exceed plain message (%d tokens)",
 			mediaTokens, plainTokens)
 	}
+
+	// Each media item should add exactly 256 tokens (not run through chars*2/5).
+	expectedDelta := 256 * 2
+	actualDelta := mediaTokens - plainTokens
+	if actualDelta != expectedDelta {
+		t.Errorf("2 media items should add %d tokens, got delta %d", expectedDelta, actualDelta)
+	}
 }
 
 // --- estimateToolDefsTokens tests ---

web/frontend/src/i18n/locales/en.json

@@ -397,7 +397,7 @@
       "max_tokens": "Max Tokens",
       "max_tokens_hint": "Upper token limit per model response.",
       "context_window": "Context Window",
-      "context_window_hint": "Model input context capacity in tokens. Leave empty to auto-detect (default: 4x max tokens).",
+      "context_window_hint": "Model input context capacity in tokens. Leave empty to use the default (4x max tokens).",
       "max_tool_iterations": "Max Tool Iterations",
       "max_tool_iterations_hint": "Maximum tool-call loops in a single task.",
       "summarize_threshold": "Summarize Message Threshold",

web/frontend/src/i18n/locales/zh.json

@@ -397,7 +397,7 @@
       "max_tokens": "最大 Token 数",
       "max_tokens_hint": "单次模型响应允许的最大 Token 数。",
       "context_window": "上下文窗口",
-      "context_window_hint": "模型输入上下文容量（Token 数）。留空则自动推算（默认为最大 Token 数的 4 倍）。",
+      "context_window_hint": "模型输入上下文容量（Token 数）。留空使用默认值（最大 Token 数的 4 倍）。",
       "max_tool_iterations": "最大工具迭代次数",
       "max_tool_iterations_hint": "单个任务中允许的工具调用循环上限。",
       "summarize_threshold": "触发摘要的消息阈值",