diff --git a/packages/ai-runtime/src/context-compression.test.ts b/packages/ai-runtime/src/context-compression.test.ts
new file mode 100644
index 00000000..fedc0c33
--- /dev/null
+++ b/packages/ai-runtime/src/context-compression.test.ts
@@ -0,0 +1,12 @@
+import assert from 'node:assert/strict'
+import { describe, it } from 'node:test'
+
+import { resolveRuntimeContextWindow } from './context-compression'
+
+describe('resolveRuntimeContextWindow', () => {
+  it('uses only exact provider matches and a conservative fallback for custom endpoints', () => {
+    assert.equal(resolveRuntimeContextWindow('deepseek', 'deepseek-v4-flash'), 1_000_000)
+    assert.equal(resolveRuntimeContextWindow('openai-compatible', 'gpt-5'), 8_192)
+    assert.equal(resolveRuntimeContextWindow('openai-compatible', 'custom-model'), 8_192)
+  })
+})
diff --git a/packages/ai-runtime/src/context-compression.ts b/packages/ai-runtime/src/context-compression.ts
index 5d830176..ae228fef 100644
--- a/packages/ai-runtime/src/context-compression.ts
+++ b/packages/ai-runtime/src/context-compression.ts
@@ -1,9 +1,12 @@
-import { BUILTIN_MODELS, getBuiltinModelById } from '@openchatlab/core'
+import { getBuiltinModelById } from '@openchatlab/core'
 import { generateText } from 'ai'
 
 import type { RunAgentInput, RuntimeContextSummary, RuntimeMessage } from './types'
 
-const DEFAULT_CONTEXT_WINDOW = 32_768
+// OpenAI-compatible endpoints do not expose standard context-window metadata.
+// Underestimating an unknown model compresses earlier; overestimating it makes
+// the provider reject an otherwise recoverable conversation.
+const DEFAULT_CONTEXT_WINDOW = 8_192
 const COMPRESSION_THRESHOLD_PERCENT = 75
 const BUFFER_SIZE_PERCENT = 20
 const MAX_SUMMARY_TOKENS = 4_096
@@ -18,9 +21,7 @@ Merge the previous summary when present. Do not copy long passages verbatim.
 {history}`
 
 export function resolveRuntimeContextWindow(providerId: string, modelId: string): number {
-  const exact = getBuiltinModelById(providerId, modelId)
-  const knownById = exact ?? BUILTIN_MODELS.find((model) => model.id === modelId)
-  return knownById?.contextWindow ?? DEFAULT_CONTEXT_WINDOW
+  return getBuiltinModelById(providerId, modelId)?.contextWindow ?? DEFAULT_CONTEXT_WINDOW
 }
 
 export async function prepareContextMessages(