Q17: budget every model request against the provider's context window

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src/main/database/migrations/062-provider-context-window.ts

@@ -0,0 +1,6 @@
+export const providerContextWindowSchema = `
+  alter table providers add column context_window_tokens integer not null default 128000;
+
+  insert into schema_migrations(version, name, applied_at)
+  values (62, 'provider_context_window', datetime('now'));
+`;

src/main/database/schema.ts

@@ -60,6 +60,7 @@ import { integrationKeysSchema } from "./migrations/058-integration-keys";
 import { taskHandoffSchema } from "./migrations/059-task-handoff";
 import { chatWakeTextPurgeSchema } from "./migrations/060-chat-wake-text-purge";
 import { taskRunContextSchema } from "./migrations/061-task-run-context";
+import { providerContextWindowSchema } from "./migrations/062-provider-context-window";
 
 interface Migration {
   version: number;
@@ -129,6 +130,7 @@ const MIGRATIONS: Migration[] = [
   { version: 59, name: "task_handoff", sql: taskHandoffSchema },
   { version: 60, name: "chat_wake_text_purge", sql: chatWakeTextPurgeSchema },
   { version: 61, name: "task_run_context", sql: taskRunContextSchema },
+  { version: 62, name: "provider_context_window", sql: providerContextWindowSchema },
 ];
 
 const SCHEMA_VERSION = MIGRATIONS[MIGRATIONS.length - 1].version;

src/main/services/agent-goal-tools.ts

@@ -1,3 +1,4 @@
+import { fitContextBudget } from "./context-budget";
 import { AuditService } from "./audit-service";
 import { executeWorkspaceTool, WORKSPACE_TOOLS } from "./agent-workspace-tools";
 import { optionalString, requireString } from "./agent-tool-args";
@@ -54,6 +55,7 @@ export interface GoalCompletionArgs {
   providers: Pick<ProviderService, "runToolTurn">;
   providerId: string;
   workspaceId: string | null;
+  inputBudgetTokens: number;
   call: ToolCall;
 }
 
@@ -129,6 +131,7 @@ async function auditGoalCompletion(
   const instructions = "You are the independent completion auditor. You did not do the work and owe its author nothing; your only loyalty is to what actually exists. Be strict: an unverifiable claim is a false claim.";
 
   for (let iteration = 1; iteration <= AUDIT_MAX_ITERATIONS; iteration += 1) {
+    fitContextBudget(instructions, messages, auditTools, args.inputBudgetTokens, iteration);
     const turn = await args.providers.runToolTurn({
       orgId: args.ctx.orgId,
       providerId: args.providerId,

src/main/services/agent-run-loop.ts

@@ -7,6 +7,7 @@ import { canonicalMentionHandle } from "../../shared/mention-handles";
 import type { AgentTurnMessage, ToolDefinition } from "./provider-tool-turn";
 import { isComposioTool, executeComposioTool } from "./composio/composio-tool-bridge";
 import { compactToolHistory } from "./run-context-compaction";
+import { fitContextBudget } from "./context-budget";
 
 const MAX_ITERATIONS = 12;
 
@@ -22,6 +23,8 @@ export interface AgentRunLoopInput {
   pollNewMessages?: () => string | null;
   /** Called after each model turn the provider answered. */
   onModelTurn?: () => void;
+  /** Input tokens the provider's context window leaves after the output reserve. */
+  inputBudgetTokens: number;
 }
 
 export interface AgentRunLoopResult {
@@ -56,6 +59,7 @@ export async function runAgentLoop(deps: AgentRunLoopDeps, input: AgentRunLoopIn
       messages.splice(0, messages.length, ...compaction.messages);
       console.info(`[run-loop] iteration ${iteration}: elided ${compaction.elided} earlier tool result(s)`);
     }
+    fitContextBudget(input.instructions, messages, allTools, input.inputBudgetTokens, iteration);
     const turn = await deps.providers.runToolTurn({
       orgId: input.orgId,
       providerId: input.providerId,
@@ -102,6 +106,7 @@ export async function runAgentLoop(deps: AgentRunLoopDeps, input: AgentRunLoopIn
           providers: deps.providers,
           providerId: input.providerId,
           workspaceId: input.workspaceId,
+          inputBudgetTokens: input.inputBudgetTokens,
           call,
         });
       } else {

src/main/services/context-budget.ts

@@ -0,0 +1,39 @@
+import { compactToolHistory } from "./run-context-compaction";
+import type { AgentTurnMessage, ToolDefinition } from "./provider-tool-turn";
+
+/** Tokens kept free for the model's answer. */
+export const OUTPUT_RESERVE_TOKENS = 16_000;
+
+/** Share of the input budget the compiled skills may take. */
+export const SKILL_BUDGET_SHARE = 0.4;
+
+const CJK = /[\u3000-\u9fff\uac00-\ud7af\uf900-\ufaff\uff00-\uffef]/g;
+
+/** Conservative: CJK counts one token per character, other text one per three. */
+export function estimateTokens(text: string): number {
+  const cjk = text.match(CJK)?.length ?? 0;
+  return cjk + Math.ceil((text.length - cjk) / 3);
+}
+
+export function inputBudgetTokens(contextWindowTokens: number): number {
+  return contextWindowTokens - OUTPUT_RESERVE_TOKENS;
+}
+
+export function skillBudgetTokens(contextWindowTokens: number): number {
+  return Math.floor(inputBudgetTokens(contextWindowTokens) * SKILL_BUDGET_SHARE);
+}
+
+export /** Shrinks history to fit the window, or fails with the numbers. */
+function fitContextBudget(instructions: string, messages: AgentTurnMessage[], tools: ToolDefinition[], budget: number, iteration: number): void {
+  const fixed = estimateTokens(instructions) + estimateTokens(JSON.stringify(tools));
+  let total = fixed + estimateTokens(JSON.stringify(messages));
+  if (total > budget) {
+    messages.splice(0, messages.length, ...compactToolHistory(messages, 1).messages);
+    const compacted = fixed + estimateTokens(JSON.stringify(messages));
+    console.warn(`[run-loop] iteration ${iteration}: ~${total} tokens over the ${budget} budget; aggressive compaction brought it to ~${compacted}`);
+    total = compacted;
+  }
+  if (total > budget) {
+    throw new Error(`This run needs ~${total} input tokens but the provider's context window leaves ${budget}. Raise the provider's context window in Settings > Advanced if the model supports more, or split the task.`);
+  }
+}

src/main/services/provider-service.ts

@@ -36,6 +36,9 @@ export type { ProviderResponseCreateInput };
 
 const AUTH_SESSION_TTL_MS = 15 * 60 * 1000;
 
+/** Room for instructions, the output reserve, and some history. */
+const MIN_CONTEXT_WINDOW_TOKENS = 24_000;
+
 export class ProviderService {
   private readonly audit: AuditService;
   private readonly authBroker: AuthBrokerService;
@@ -254,6 +257,10 @@ export class ProviderService {
   updatePolicy(input: UpdateProviderPolicyInput): void {
     const monthlyRequestLimit = Math.trunc(input.monthlyRequestLimit);
     if (monthlyRequestLimit < 1) throw new Error("Monthly request limit must be at least 1.");
+    const contextWindowTokens = Math.trunc(input.contextWindowTokens);
+    if (!(contextWindowTokens >= MIN_CONTEXT_WINDOW_TOKENS)) {
+      throw new Error(`Context window must be at least ${MIN_CONTEXT_WINDOW_TOKENS} tokens.`);
+    }
     if (!Object.values(PROVIDER_DATA_POLICY).includes(input.dataPolicy)) throw new Error("Provider data policy is invalid.");
     if (!WORKSPACE_SCOPES.has(input.workspaceScope)) throw new Error("Provider workspace scope is invalid.");
     if (input.dataPolicy === PROVIDER_DATA_POLICY.ZERO_DATA_RETENTION && input.backgroundEnabled) {
@@ -265,9 +272,9 @@ export class ProviderService {
 
     this.db.prepare(`
       update providers
-      set data_policy = ?, background_enabled = ?, monthly_request_limit = ?, workspace_scope = ?
+      set data_policy = ?, background_enabled = ?, monthly_request_limit = ?, workspace_scope = ?, context_window_tokens = ?
       where id = ? and org_id = ?
-    `).run(input.dataPolicy, input.backgroundEnabled ? 1 : 0, monthlyRequestLimit, input.workspaceScope, input.providerId, input.orgId);
+    `).run(input.dataPolicy, input.backgroundEnabled ? 1 : 0, monthlyRequestLimit, input.workspaceScope, contextWindowTokens, input.providerId, input.orgId);
     this.audit.record({
       orgId: input.orgId,
       entityType: "provider",
@@ -281,6 +288,7 @@ export class ProviderService {
         backgroundEnabled: input.backgroundEnabled,
         monthlyRequestLimit,
         workspaceScope: input.workspaceScope,
+        contextWindowTokens,
       },
     });
   }

src/main/services/run-context-compaction.ts

@@ -5,10 +5,10 @@ const ELIDE_ABOVE_CHARS = 600;
 const PREVIEW_CHARS = 300;
 
 /** Older tool results shrink to a preview; the latest turns stay verbatim. */
-export function compactToolHistory(messages: AgentTurnMessage[]): { messages: AgentTurnMessage[]; elided: number } {
+export function compactToolHistory(messages: AgentTurnMessage[], keepRecent = KEEP_RECENT_TOOL_TURNS): { messages: AgentTurnMessage[]; elided: number } {
   const toolTurnIndexes = messages.flatMap((message, index) => (message.role === "tool" ? [index] : []));
-  const compactBefore = toolTurnIndexes.length > KEEP_RECENT_TOOL_TURNS
-    ? toolTurnIndexes[toolTurnIndexes.length - KEEP_RECENT_TOOL_TURNS]
+  const compactBefore = toolTurnIndexes.length > keepRecent
+    ? toolTurnIndexes[toolTurnIndexes.length - keepRecent]
     : 0;
   let elided = 0;
   const compacted = messages.map((message, index) => {

src/main/services/skill-service.ts

@@ -1,3 +1,4 @@
+import { estimateTokens } from "./context-budget";
 import { existsSync, readFileSync } from "node:fs";
 import { createHash } from "node:crypto";
 import { normalize } from "node:path";
@@ -34,13 +35,6 @@ export function forcedSkillSlugCandidates(language: string): string[] {
   return language === "zh" ? [HUMANIZER_SLUG_ZH, HUMANIZER_SLUG_EN] : [HUMANIZER_SLUG_EN, HUMANIZER_SLUG_ZH];
 }
 
-/**
- * Compiled skill prompts stop packing past this size so one over-granted
- * agent cannot overflow the model's context window. Skills pack whole, forced
- * skills first, and anything omitted is named in the prompt.
- */
-export const MAX_COMPILED_PROMPT_CHARS = 200_000;
-
 export class SkillService {
   private readonly audit: AuditService;
   private readonly operations: OperationsService;
@@ -243,8 +237,8 @@ export class SkillService {
    * Freezes the exact skill set this run carries: forced house-style skills
    * first (identical for everyone, CEO included), then the agent's enabled
    * bindings in priority order, deduplicated per skill. Skills pack whole
-   * until MAX_COMPILED_PROMPT_CHARS; whatever does not fit is named in the
-   * compiled prompt instead of silently overflowing the model's context.
+   * until input.maxTokens; whatever does not fit is named in the compiled
+   * prompt instead of silently overflowing the model's context.
    */
   resolveRuntimeSnapshot(input: ResolveAgentSkillsInput): string {
     // Idempotent per run: a run interrupted after resolving (crash, quit) is
@@ -280,7 +274,7 @@ export class SkillService {
     const tools = new Set<string>();
     const revisionIds: string[] = [];
     const omitted: string[] = [];
-    let compiledLength = 0;
+    let compiledTokens = 0;
     for (const row of rows) {
       if (!existsSync(row.object_key)) throw new Error(`Skill artifact is missing: ${row.object_key}`);
       const manifest = JSON.parse(row.manifest_json) as SkillManifest;
@@ -289,14 +283,15 @@ export class SkillService {
         `Version: ${row.version_semver}`,
         readFileSync(row.object_key, "utf8").trim(),
       ].join("\n\n");
-      if (compiledParts.length > 0 && compiledLength + part.length > MAX_COMPILED_PROMPT_CHARS) {
+      const partTokens = estimateTokens(part);
+      if (compiledParts.length > 0 && compiledTokens + partTokens > input.maxTokens) {
         omitted.push(manifest.name);
         continue;
       }
       for (const tool of manifest.requiredTools) tools.add(tool);
       revisionIds.push(row.revision_id);
       compiledParts.push(part);
-      compiledLength += part.length;
+      compiledTokens += partTokens;
     }
     if (omitted.length > 0) {
       compiledParts.push(

src/main/services/task-execution-service.ts

@@ -26,6 +26,7 @@ import {
 import { renderTaskHandoff } from "./agent-task-handoff";
 import { createRunContextRecorder, createRunMessagePoller, loadChatWakeContext } from "./chat-wake-context";
 import { runningWorkspaceWriter } from "./workspace-write-lock";
+import { inputBudgetTokens, skillBudgetTokens } from "./context-budget";
 import { loadComposioTools } from "./composio/composio-tool-bridge";
 import { connectorDisplayName } from "./composio/connector-brands";
 import { ConnectorGrantsService } from "./composio/connector-grants";
@@ -155,7 +156,7 @@ export class TaskExecutionService {
     let providerRequestId: string | null = null;
     let skillRuntimeSnapshot: SkillRuntimeSnapshotRow | null = null;
     try {
-      skillRuntimeSnapshot = this.resolveSkillRuntimeSnapshot(orgId, agent.id, runId);
+      skillRuntimeSnapshot = this.resolveSkillRuntimeSnapshot(orgId, agent.id, runId, skillBudgetTokens(provider.context_window_tokens));
       // Watermark for reply delivery and mid-run message polling.
       const runStartedAt = new Date().toISOString();
       const chatContext = loadChatWakeContext(this.db, orgId, agent.id, isolation.secretScopeJson);
@@ -208,6 +209,7 @@ export class TaskExecutionService {
           toolContext,
           externalTools,
           onModelTurn: contextRecorder.commit,
+          inputBudgetTokens: inputBudgetTokens(provider.context_window_tokens),
           pollNewMessages: createRunMessagePoller(this.db, { orgId, onInjected: contextRecorder.addMidRun, agentId: agent.id, sourceThreadId: chatContext?.threadId ?? null, runStartedAt }),
         },
       );
@@ -312,7 +314,7 @@ export class TaskExecutionService {
   private requireAgentProvider(orgId: string, agent: TaskExecutionAgentRow): TaskExecutionProviderRow {
     if (!agent.model_provider_id) throw new Error("Task assignee requires a model provider.");
     const provider = this.db.prepare(`
-      select id, label, provider_key, model
+      select id, label, provider_key, model, context_window_tokens
       from providers
       where id = ? and org_id = ?
     `).get(agent.model_provider_id, orgId) as TaskExecutionProviderRow | undefined;
@@ -614,9 +616,9 @@ export class TaskExecutionService {
     `).run(TASK_RUN_STATE.FAILED, providerRequestId, new Date().toISOString(), errorMessage, runId, orgId);
   }
 
-  private resolveSkillRuntimeSnapshot(orgId: string, agentId: string, runId: string): SkillRuntimeSnapshotRow | null {
+  private resolveSkillRuntimeSnapshot(orgId: string, agentId: string, runId: string, maxTokens: number): SkillRuntimeSnapshotRow | null {
     if (!this.skills.hasRuntimeSkills(orgId, agentId)) return null;
-    const snapshotId = this.skills.resolveRuntimeSnapshot({ orgId, agentId, runId });
+    const snapshotId = this.skills.resolveRuntimeSnapshot({ orgId, agentId, runId, maxTokens });
     const snapshot = this.db.prepare(`
       select id, compiled_prompt, tools_manifest, source_revision_ids
       from skill_runtime_snapshots

src/main/services/task-prompts.ts

@@ -34,6 +34,7 @@ export interface TaskExecutionProviderRow {
   label: string;
   provider_key: string;
   model: string | null;
+  context_window_tokens: number;
 }
 
 export interface ProjectRow {