chore(context): Apply cruft removal and strict hyperparameter types

packages/cli/src/config/settingsSchema.ts

@@ -2482,6 +2482,35 @@ const SETTINGS_SCHEMA = {
         default: {},
         showInDialog: false,
         properties: {
+          maxPressureStrategy: {
+            type: 'string',
+            label: 'Max Pressure Strategy',
+            category: 'Context Management',
+            requiresRestart: true,
+            default: 'truncate',
+            description: 'Action to take when hitting the synchronous token ceiling.',
+            showInDialog: false,
+            enum: ['truncate', 'compress', 'rollingSummarizer'],
+          },
+          gcTarget: {
+            type: 'string',
+            label: 'GC Target',
+            category: 'Context Management',
+            requiresRestart: true,
+            default: 'incremental',
+            description: 'The target floor for synchronous context shedding.',
+            showInDialog: false,
+            enum: ['incremental', 'freeNTokens', 'max'],
+          },
+          freeTokensTarget: {
+            type: 'number',
+            label: 'Free Tokens Target',
+            category: 'Context Management',
+            requiresRestart: true,
+            default: undefined,
+            description: 'Amount of tokens to free when gcTarget is freeNTokens.',
+            showInDialog: false,
+          },
           maxTokens: {
             type: 'number',
             label: 'Max Tokens',

packages/core/src/config/config.ts

@@ -1133,10 +1133,8 @@ export class Config implements McpContext, AgentLoopContext {
         maxTokens: params.contextManagement?.budget?.maxTokens ?? 150000,
         retainedTokens:
           params.contextManagement?.budget?.retainedTokens ?? 40000,
-        protectedEpisodes:
-          params.contextManagement?.budget?.protectedEpisodes ?? 1,
-        protectSystemEpisode:
-          params.contextManagement?.budget?.protectSystemEpisode ?? true,
+        gcTarget: params.contextManagement?.budget?.gcTarget ?? 'incremental',
+        freeTokensTarget: params.contextManagement?.budget?.freeTokensTarget ?? undefined,
         maxPressureStrategy: params.contextManagement?.budget?.maxPressureStrategy ?? 'truncate',
       },
       strategies: {
@@ -1154,9 +1152,7 @@ export class Config implements McpContext, AgentLoopContext {
           nodeThresholdTokens:
             params.contextManagement?.strategies?.semanticCompression
               ?.nodeThresholdTokens ?? 5000,
-          compressionModel:
-            params.contextManagement?.strategies?.semanticCompression
-              ?.compressionModel ?? 'chat-compression-2.5-flash-lite',
+          
         },
       },
     };

packages/core/src/context/processors/semanticCompressionProcessor.ts

@@ -34,7 +34,7 @@ export class SemanticCompressionProcessor implements ContextProcessor {
       this.config.getContextManagementConfig().strategies.semanticCompression;
     // We estimate 4 chars per token for truncation logic
     const thresholdChars = semanticConfig.nodeThresholdTokens * 4;
-    this.modelToUse = semanticConfig.compressionModel;
+    this.modelToUse = 'gemini-2.5-flash';
 
     let currentDeficit = state.deficitTokens;
     const newEpisodes = [...episodes];

packages/core/src/context/profiles.ts

@@ -11,8 +11,7 @@ export const GENERALIST_PROFILE: ContextManagementConfig = {
     maxPressureStrategy: 'truncate',
     maxTokens: 150_000,
     retainedTokens: 65_000,
-    protectedEpisodes: 1,
-    protectSystemEpisode: true,
+    gcTarget: 'incremental',
   },
   strategies: {
     // Brutal fallback truncation threshold
@@ -22,7 +21,7 @@ export const GENERALIST_PROFILE: ContextManagementConfig = {
     // Intelligently summarize large text blocks before they hit the truncation guillotine
     semanticCompression: {
       nodeThresholdTokens: 3000,
-      compressionModel: 'chat-compression-2.5-flash-lite',
+      
     },
   },
 };
@@ -33,15 +32,14 @@ export const POWER_USER_PROFILE: ContextManagementConfig = {
     maxPressureStrategy: 'truncate',
     maxTokens: 150_000, // The absolute ceiling
     retainedTokens: 65_000, // The "bloom filter" backbuffer floor
-    protectedEpisodes: 1,
-    protectSystemEpisode: true,
+    gcTarget: 'incremental',
   },
   strategies: {
     historySquashing: { maxTokensPerNode: 4000 },
     toolMasking: { stringLengthThresholdTokens: 8000 },
     semanticCompression: {
       nodeThresholdTokens: 3000,
-      compressionModel: 'chat-compression-2.5-flash-lite',
+      
     },
   },
 };
@@ -53,15 +51,14 @@ export const STRESS_TEST_PROFILE: ContextManagementConfig = {
     maxPressureStrategy: 'truncate',
     maxTokens: 12_000,
     retainedTokens: 6_000,
-    protectedEpisodes: 1,
-    protectSystemEpisode: true,
+    gcTarget: 'incremental',
   },
   strategies: {
     historySquashing: { maxTokensPerNode: 2000 },
     toolMasking: { stringLengthThresholdTokens: 2000 },
     semanticCompression: {
       nodeThresholdTokens: 1000,
-      compressionModel: 'gemini-2.5-flash',
+      
     },
   },
 };

packages/core/src/context/types.ts

@@ -23,7 +23,9 @@ export interface ContextManagementConfig {
      * - 'truncate': Drop oldest episodes until under limit (Instant, data loss)
      * - 'compress': Block request, perform N-to-1 Snapshot generation, then proceed (Slow, no data loss)
      */
-    maxPressureStrategy: 'truncate' | 'compress';
+    maxPressureStrategy: 'truncate' | 'compress' | 'rollingSummarizer';
+    gcTarget: 'incremental' | 'freeNTokens' | 'max';
+    freeTokensTarget?: number;
   };
 
   /** Specific hyperparameters for degrading the context when over budget */