From 2ac655b69e2c03e7279ce0bbc03cca6fd21669bb Mon Sep 17 00:00:00 2001
From: shivammittal274 <56757235+shivammittal274@users.noreply.github.com>
Date: Wed, 7 Jan 2026 17:13:00 +0530
Subject: [PATCH] fix: context window size (#171)

---
 .../options/ai-settings/NewProviderDialog.tsx    |  6 ++++--
 .../sidepanel/index/useChatSession.ts            |  2 ++
 .../agent/lib/schedules/getChatServerResponse.ts |  2 ++
 apps/server/src/agent/agent/gemini-agent.ts      | 16 +++++++++++-----
 packages/shared/src/constants/limits.ts          |  5 ++++-
 5 files changed, 23 insertions(+), 8 deletions(-)

diff --git a/apps/agent/entrypoints/options/ai-settings/NewProviderDialog.tsx b/apps/agent/entrypoints/options/ai-settings/NewProviderDialog.tsx
index c294b510..c981988a 100644
--- a/apps/agent/entrypoints/options/ai-settings/NewProviderDialog.tsx
+++ b/apps/agent/entrypoints/options/ai-settings/NewProviderDialog.tsx
@@ -245,8 +245,10 @@ export const NewProviderDialog: FC<NewProviderDialogProps> = ({
     setIsCustomModel(false)
   }
 
-  // Auto-fill context window when model changes
+  // Auto-fill context window when model changes (only for new providers)
   useEffect(() => {
+    if (initialValues?.id) return
+
     if (watchedModelId && watchedModelId !== 'custom') {
       const contextLength = getModelContextLength(
         watchedType as ProviderType,
@@ -256,7 +258,7 @@ export const NewProviderDialog: FC<NewProviderDialogProps> = ({
         form.setValue('contextWindow', contextLength)
       }
     }
-  }, [watchedModelId, watchedType, form])
+  }, [watchedModelId, watchedType, form, initialValues?.id])
 
   // Handle model selection (including custom option)
   const handleModelChange = (value: string) => {
diff --git a/apps/agent/entrypoints/sidepanel/index/useChatSession.ts b/apps/agent/entrypoints/sidepanel/index/useChatSession.ts
index 95a9c168..1decc7cd 100644
--- a/apps/agent/entrypoints/sidepanel/index/useChatSession.ts
+++ b/apps/agent/entrypoints/sidepanel/index/useChatSession.ts
@@ -261,6 +261,8 @@ export const useChatSession = () => {
             conversationId: conversationIdRef.current,
             model: provider?.modelId ?? 'default',
             mode: currentMode,
+            contextWindowSize: provider?.contextWindow,
+            temperature: provider?.temperature,
             // Azure-specific
             resourceName: provider?.resourceName,
             // Bedrock-specific
diff --git a/apps/agent/lib/schedules/getChatServerResponse.ts b/apps/agent/lib/schedules/getChatServerResponse.ts
index d080b269..8b71a8ad 100644
--- a/apps/agent/lib/schedules/getChatServerResponse.ts
+++ b/apps/agent/lib/schedules/getChatServerResponse.ts
@@ -78,6 +78,8 @@ export async function getChatServerResponse(
       conversationId,
       model: provider?.modelId ?? 'default',
       mode: request.mode ?? 'agent',
+      contextWindowSize: provider?.contextWindow,
+      temperature: provider?.temperature,
       resourceName: provider?.resourceName,
       accessKeyId: provider?.accessKeyId,
       secretAccessKey: provider?.secretAccessKey,
diff --git a/apps/server/src/agent/agent/gemini-agent.ts b/apps/server/src/agent/agent/gemini-agent.ts
index d5323be3..49ffd5e7 100644
--- a/apps/server/src/agent/agent/gemini-agent.ts
+++ b/apps/server/src/agent/agent/gemini-agent.ts
@@ -49,16 +49,22 @@ export class GeminiAgent {
     // Calculate compression threshold based on context window size
     const contextWindow =
       config.contextWindowSize ?? AGENT_LIMITS.DEFAULT_CONTEXT_WINDOW
+
+    // Hybrid compression: ensure minimum headroom while capping ratio for large contexts
+    const headroomBasedRatio =
+      (contextWindow - AGENT_LIMITS.COMPRESSION_MIN_HEADROOM) / contextWindow
+    const compressionRatio = Math.min(
+      AGENT_LIMITS.COMPRESSION_MAX_RATIO,
+      Math.max(AGENT_LIMITS.COMPRESSION_MIN_RATIO, headroomBasedRatio),
+    )
     const compressionThreshold =
-      (AGENT_LIMITS.DEFAULT_COMPRESSION_RATIO * contextWindow) /
-      AGENT_LIMITS.DEFAULT_CONTEXT_WINDOW
+      (compressionRatio * contextWindow) / AGENT_LIMITS.DEFAULT_CONTEXT_WINDOW
 
     logger.info('Compression config', {
       contextWindow,
+      compressionRatio,
       compressionThreshold,
-      compressesAtTokens: Math.floor(
-        AGENT_LIMITS.DEFAULT_COMPRESSION_RATIO * contextWindow,
-      ),
+      compressesAtTokens: Math.floor(compressionRatio * contextWindow),
     })
 
     logger.debug('MCP servers config', {
diff --git a/packages/shared/src/constants/limits.ts b/packages/shared/src/constants/limits.ts
index 343dd3e5..c721e4fe 100644
--- a/packages/shared/src/constants/limits.ts
+++ b/packages/shared/src/constants/limits.ts
@@ -15,7 +15,10 @@ export const RATE_LIMITS = {
 export const AGENT_LIMITS = {
   MAX_TURNS: 100,
   DEFAULT_CONTEXT_WINDOW: 1_000_000,
-  DEFAULT_COMPRESSION_RATIO: 0.75,
+  // Compression settings - hybrid approach with minimum headroom
+  COMPRESSION_MIN_HEADROOM: 10_000, // Always leave at least 10K tokens for tool responses
+  COMPRESSION_MAX_RATIO: 0.75, // Never wait longer than 75% for large models
+  COMPRESSION_MIN_RATIO: 0.4, // Never compress too early (before 40%)
 } as const
 
 export const PAGINATION = {