fix: restore full context percentage limits

Refs #522

Author: Tarquinen

lib/hooks.ts

@@ -13,7 +13,6 @@ import {
     stripHallucinationsFromString,
     stripStaleMetadata,
     syncCompressionBlocks,
-    computeInputBudget,
 } from "./messages"
 import { renderSystemPrompt, type PromptStore } from "./prompts"
 import { buildProtectedToolsExtension } from "./prompts/extensions/system"
@@ -54,17 +53,11 @@ export function createSystemPromptHandler(
     prompts: PromptStore,
 ) {
     return async (
-        input: {
-            sessionID?: string
-            model: { limit: { context: number; input?: number; output?: number } }
-        },
+        input: { sessionID?: string; model: { limit: { context: number } } },
         output: { system: string[] },
     ) => {
         if (input.model?.limit?.context) {
-            const inputBudget = computeInputBudget(input.model.limit)
-            if (inputBudget !== undefined) {
-                state.modelContextLimit = inputBudget
-            }
+            state.modelContextLimit = input.model.limit.context
             logger.debug("Cached model context limit", { limit: state.modelContextLimit })
         }
 

lib/messages/index.ts

@@ -1,7 +1,6 @@
 export { prune } from "./prune"
 export { syncCompressionBlocks } from "./sync"
 export { injectCompressNudges } from "./inject/inject"
-export { computeInputBudget } from "./inject/utils"
 export { injectMessageIds } from "./inject/inject"
 export { injectExtendedSubAgentResults } from "./inject/subagent-results"
 export { stripStaleMetadata } from "./reasoning-strip"

lib/messages/inject/utils.ts

@@ -34,20 +34,6 @@ export interface LastNonIgnoredMessage {
     index: number
 }
 
-interface ModelLimit {
-    context: number
-    input?: number
-    output?: number
-}
-
-export function computeInputBudget(limit: ModelLimit): number | undefined {
-    if (!limit.context) {
-        return undefined
-    }
-
-    return limit.input ?? Math.max(0, limit.context - (limit.output ?? 0))
-}
-
 export function getNudgeFrequency(config: PluginConfig): number {
     return Math.max(1, Math.floor(config.compress.nudgeFrequency || 1))
 }

tests/hooks-permission.test.ts

@@ -5,6 +5,7 @@ import {
     createChatMessageTransformHandler,
     createCommandExecuteHandler,
     createEventHandler,
+    createSystemPromptHandler,
     createTextCompleteHandler,
 } from "../lib/hooks"
 import { Logger } from "../lib/logger"
@@ -86,6 +87,36 @@ function buildMessage(id: string, role: "user" | "assistant", text: string): Wit
     }
 }
 
+test("system prompt handler caches full model context for percentage thresholds", async () => {
+    const state = createSessionState()
+    const handler = createSystemPromptHandler(
+        state,
+        new Logger(false),
+        buildConfig("deny"),
+        {
+            reload() {},
+            getRuntimePrompts() {
+                return {} as any
+            },
+        } as any,
+    )
+
+    await handler(
+        {
+            sessionID: "session-1",
+            model: {
+                limit: {
+                    context: 200000,
+                    output: 131072,
+                },
+            },
+        } as any,
+        { system: ["base system"] },
+    )
+
+    assert.equal(state.modelContextLimit, 200000)
+})
+
 test("chat message transform strips hallucinated tags even when compress is denied", async () => {
     const state = createSessionState()
     const logger = new Logger(false)