feat: preserve pages and compress context

src/shared/assistant-contracts.ts

@@ -147,6 +147,14 @@ export const conversationMessageSchema = z
     createdAt: z.number().int().nonnegative(),
     state: z.enum(['streaming', 'complete', 'error']),
     status: z.string().max(4_000).optional(),
+    contextCompression: z
+      .object({
+        state: z.enum(['compressing', 'completed', 'failed']),
+        estimatedBeforeTokens: z.number().int().nonnegative(),
+        estimatedAfterTokens: z.number().int().nonnegative().optional()
+      })
+      .strict()
+      .optional(),
     tools: z.array(conversationToolActivitySchema).max(100).optional(),
     sources: z.array(z.string().max(8_192)).max(100).optional(),
     sourceReferences: z

src/shared/context-window.ts

@@ -0,0 +1,63 @@
+export const minimumModelContextWindowTokens = 32_000
+export const maximumModelContextWindowTokens = 10_000_000
+export const contextOutputAndSafetyTokens = 12_000
+export const estimatedContextRequestOverheadTokens = 4_000
+
+export type ContextWindowMessage = {
+  role: 'user' | 'assistant'
+  content: string
+}
+
+export function estimateTextTokens(value: string): number {
+  let asciiCharacters = 0
+  let nonAsciiCharacters = 0
+  for (const character of value) {
+    if (character.codePointAt(0)! <= 0x7f) {
+      asciiCharacters += 1
+    } else {
+      nonAsciiCharacters += 1
+    }
+  }
+  return Math.max(
+    1,
+    Math.ceil(asciiCharacters / 4 + nonAsciiCharacters)
+  )
+}
+
+export function estimateMessagesTokens(
+  messages: readonly ContextWindowMessage[]
+): number {
+  return messages.reduce(
+    (total, message) => total + estimateTextTokens(message.content) + 4,
+    0
+  )
+}
+
+export function estimateContextInputTokens(input: {
+  history: readonly ContextWindowMessage[]
+  prompt: string
+  summaryTokens?: number
+}): number {
+  return (
+    estimateMessagesTokens(input.history) +
+    estimateTextTokens(input.prompt) +
+    (input.summaryTokens ?? 0) +
+    estimatedContextRequestOverheadTokens
+  )
+}
+
+export function getEffectiveContextTriggerTokens(input: {
+  triggerTokens: number
+  contextWindowTokens?: number
+}): number {
+  if (input.contextWindowTokens === undefined) {
+    return input.triggerTokens
+  }
+  return Math.min(
+    input.triggerTokens,
+    Math.max(
+      input.contextWindowTokens,
+      minimumModelContextWindowTokens
+    ) - contextOutputAndSafetyTokens
+  )
+}

src/shared/contracts.ts

@@ -1,4 +1,8 @@
 import { z } from 'zod'
+import {
+  maximumModelContextWindowTokens,
+  minimumModelContextWindowTokens
+} from './context-window'
 import type {
   BrowserProfileCreateInput,
   BrowserProfileRenameInput,
@@ -416,6 +420,11 @@ const modelApiKeyUpdateSchema = z.discriminatedUnion('action', [
   z.object({ action: z.literal('clear') }).strict()
 ])
 
+export {
+  maximumModelContextWindowTokens,
+  minimumModelContextWindowTokens
+} from './context-window'
+
 const modelProfileInputSchema = z
   .object({
     id: modelProfileIdSchema,
@@ -433,8 +442,8 @@ const modelProfileInputSchema = z
     contextWindowTokens: z
       .number()
       .int()
-      .min(8_000)
-      .max(10_000_000)
+      .min(minimumModelContextWindowTokens)
+      .max(maximumModelContextWindowTokens)
       .optional(),
     imageGenerationQuality: imageGenerationQualitySchema,
     apiKey: modelApiKeyUpdateSchema
@@ -907,6 +916,29 @@ export type AgentEvent =
       type: 'reasoning'
       delta: string
     }
+  | {
+      requestId: string
+      type: 'context-metrics'
+      estimatedInputTokens: number
+      effectiveTriggerTokens: number
+      contextWindowTokens?: number
+      compressionEnabled: boolean
+      recentRawTokens: number
+      coveredMessageCount: number
+      summaryTokens: number
+    }
+  | {
+      requestId: string
+      type: 'context-compression'
+      state: 'started' | 'completed'
+      estimatedBeforeTokens: number
+      estimatedAfterTokens?: number
+      effectiveTriggerTokens: number
+      contextWindowTokens?: number
+      recentRawTokens: number
+      coveredMessageCount: number
+      summaryTokens?: number
+    }
   | {
       requestId: string
       type: 'tool'