From 88523e099cf09e76830d848e29ddcdbe61cbf876 Mon Sep 17 00:00:00 2001
From: yin1245 <2573101189@qq.com>
Date: Sun, 29 Mar 2026 12:31:46 +0800
Subject: [PATCH] fix: use per-model context window size for accurate
 input_tokens estimation

Different Claude models have different context window sizes:
- Opus 4.5/4.6: 1M tokens
- Sonnet 4.5/4.6: 200K tokens
- Haiku 4.5: 200K tokens

The current implementation uses a single hardcoded TOTAL_CONTEXT_TOKENS
constant for all models, which causes input_tokens to be overestimated
for small-context models or underestimated for large-context models.

This change adds a MODEL_CONTEXT_TOKENS mapping and a helper function
getContextTokensForModel() that selects the correct context window size
based on the model being used, falling back to KIRO_CONSTANTS.TOTAL_CONTEXT_TOKENS
for unknown models.
---
 src/providers/claude/claude-kiro.js | 18 +++++++++++++++++-
 1 file changed, 17 insertions(+), 1 deletion(-)

diff --git a/src/providers/claude/claude-kiro.js b/src/providers/claude/claude-kiro.js
index 303533d..7c185e3 100644
--- a/src/providers/claude/claude-kiro.js
+++ b/src/providers/claude/claude-kiro.js
@@ -47,6 +47,21 @@ const KIRO_CONSTANTS = {
     TOTAL_CONTEXT_TOKENS: 200000, // Claude Sonnet 4.5 actual context is 200K
 };
 
+// Per-model context window sizes for accurate token estimation
+const MODEL_CONTEXT_TOKENS = {
+    "claude-opus-4-6": 1000000,
+    "claude-opus-4-5": 1000000,
+    "claude-opus-4-5-20251101": 1000000,
+    "claude-sonnet-4-6": 200000,
+    "claude-sonnet-4-5": 200000,
+    "claude-sonnet-4-5-20250929": 200000,
+    "claude-haiku-4-5": 200000,
+    "claude-haiku-4-5-20251001": 200000,
+};
+
+function getContextTokensForModel(model) {
+    return MODEL_CONTEXT_TOKENS[model] || KIRO_CONSTANTS.TOTAL_CONTEXT_TOKENS;
+}
 // 从 provider-models.js 获取支持的模型列表
 const KIRO_MODELS = getProviderModels(MODEL_PROVIDER.KIRO_API);
 
@@ -2631,7 +2646,8 @@ async saveCredentialsToFile(filePath, newData) {
             // 总 token = TOTAL_CONTEXT_TOKENS * contextUsagePercentage / 100
             // input token = 总 token - output token
             if (contextUsagePercentage !== null && contextUsagePercentage > 0) {
-                const totalTokens = Math.round(KIRO_CONSTANTS.TOTAL_CONTEXT_TOKENS * contextUsagePercentage / 100);
+                const contextTokens = getContextTokensForModel(finalModel);
+                const totalTokens = Math.round(contextTokens * contextUsagePercentage / 100);
                 inputTokens = Math.max(0, totalTokens - outputTokens);
                 logger.info(`[Kiro] Token calculation from contextUsagePercentage: total=${totalTokens}, output=${outputTokens}, input=${inputTokens}`);
             } else {