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
+  const compactBefore = toolTurnIndexes.length > keepRecent
-    ? toolTurnIndexes[toolTurnIndexes.length - KEEP_RECENT_TOOL_TURNS]
+    ? 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
+   * until input.maxTokens; whatever does not fit is named in the compiled
-   * compiled prompt instead of silently overflowing the model's context.
+   * 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 {

src/renderer/src/i18n/zh.ts

@@ -28,6 +28,9 @@ export const zh: Record<string, string> = {
   "Advanced": "高级",
   "Integrations": "集成密钥",
   "Read by {names}": "已读：{names}",
+  "Context window (tokens)": "上下文窗口（token）",
+  "{label} context window": "{label} 上下文窗口",
+  "The model's context window. Runs keep prompts, skills, history, and tool results inside it.": "模型的上下文窗口。运行时提示词、技能、历史和工具结果都要放进这个范围。",
   "Keys are stored on this device, encrypted by the operating system, and shared by every company.": "密钥保存在本机，由操作系统加密，所有公司共用。",
   "Composio API key": "Composio API 密钥",
   "Powers the Connectors page and agent connector tools. Create one at platform.composio.dev.": "用于连接器页面和 Agent 的连接器工具。在 platform.composio.dev 创建。",

src/renderer/src/pages/settings/ProvidersPanel.tsx

@@ -26,7 +26,7 @@ interface ProvidersPanelProps {
   actions: SnapshotActions;
 }
 
-const POLICY_GRID = "minmax(150px, 1.2fr) minmax(118px, 0.8fr) minmax(140px, 1fr) 112px 124px 92px";
+const POLICY_GRID = "minmax(150px, 1.2fr) minmax(118px, 0.8fr) minmax(140px, 1fr) 112px 124px 124px 92px";
 type WorkspaceScope = "organization" | "workspace";
 
 export function ProvidersPanel({ orgId, employeeId, providers, authProfiles, catalog, agents, providerRateLimits, actions }: ProvidersPanelProps) {
@@ -138,9 +138,9 @@ function ProviderPolicyTable({
       <div className="card" style={{ padding: 0, overflowX: "auto", marginBottom: 28 }}>
         <div
           className="table-row head"
-          style={{ gridTemplateColumns: POLICY_GRID, minWidth: 780 }}
+          style={{ gridTemplateColumns: POLICY_GRID, minWidth: 900 }}
         >
-          <div>{t("Provider")}</div><div>{t("Scope")}</div><div>{t("Data policy")}</div><div>{t("Background")}</div><div>{t("Monthly limit")}</div><div>{t("Usage")}</div>
+          <div>{t("Provider")}</div><div>{t("Scope")}</div><div>{t("Data policy")}</div><div>{t("Background")}</div><div>{t("Monthly limit")}</div><div>{t("Context window (tokens)")}</div><div>{t("Usage")}</div>
         </div>
         {providers.map((provider, index) => {
           const rateLimit = requireProviderRateLimit(provider, providerRateLimits);
@@ -176,6 +176,7 @@ function ProviderPolicyRow({
   const [dataPolicy, setDataPolicy] = useState(provider.dataPolicy);
   const [backgroundEnabled, setBackgroundEnabled] = useState(provider.backgroundEnabled);
   const [monthlyRequestLimit, setMonthlyRequestLimit] = useState(String(provider.monthlyRequestLimit));
+  const [contextWindowTokens, setContextWindowTokens] = useState(String(provider.contextWindowTokens));
   const [workspaceScope, setWorkspaceScope] = useState<WorkspaceScope>(provider.workspaceScope === "workspace" ? "workspace" : "organization");
 
   const savePolicy = () => {
@@ -186,6 +187,7 @@ function ProviderPolicyRow({
       backgroundEnabled,
       monthlyRequestLimit: Number(monthlyRequestLimit),
       workspaceScope,
+      contextWindowTokens: Number(contextWindowTokens),
     }));
   };
 
@@ -199,7 +201,7 @@ function ProviderPolicyRow({
       className="table-row"
       style={{
         gridTemplateColumns: POLICY_GRID,
-        minWidth: 780,
+        minWidth: 900,
         borderBottom: last ? "none" : "1px solid var(--border-soft)",
       }}
     >
@@ -242,6 +244,16 @@ function ProviderPolicyRow({
         value={monthlyRequestLimit}
         onChange={(event) => setMonthlyRequestLimit(event.target.value)}
       />
+      <input
+        className="input"
+        aria-label={t("{label} context window", { label: provider.label })}
+        title={t("The model's context window. Runs keep prompts, skills, history, and tool results inside it.")}
+        type="number"
+        min={24000}
+        step={1000}
+        value={contextWindowTokens}
+        onChange={(event) => setContextWindowTokens(event.target.value)}
+      />
       <div>
         <div className="num mono">{rateLimit.requestCount} / {rateLimit.monthlyRequestLimit}</div>
         <button type="button" className="button small" aria-label={t("Save {label} policy", { label: provider.label })} onClick={savePolicy}>

src/shared/commands/provider.ts

@@ -61,6 +61,7 @@ export interface UpdateProviderPolicyInput {
   backgroundEnabled: boolean;
   monthlyRequestLimit: number;
   workspaceScope: "organization" | "workspace";
+  contextWindowTokens: number;
 }
 
 export interface CreateProviderResponseInput {

src/shared/commands/skills.ts

@@ -19,6 +19,8 @@ export interface ResolveAgentSkillsInput {
   orgId: string;
   agentId: string;
   runId: string;
+  /** Token budget for the compiled skills. */
+  maxTokens: number;
 }
 
 export interface RunSkillTestsInput {

src/shared/domain.ts

@@ -284,6 +284,8 @@ export interface ProviderRecord {
   backgroundEnabled: boolean;
   monthlyRequestLimit: number;
   workspaceScope: string;
+  /** Model context window the run budget is computed against. */
+  contextWindowTokens: number;
   createdAt: string;
 }
 

tests/integration/collaboration-flow.test.ts

@@ -98,7 +98,7 @@ describe("collaboration integration flow", () => {
       revisionId = skills.importDirectory({ orgId, directoryPath: createSkillDirectory(), versionSemver: "1.0.0" });
       skillId = (connection.db.prepare("select skill_id from skill_revisions where id = ?").get(revisionId) as { skill_id: string }).skill_id;
       skills.bind({ orgId, agentId, skillId, revisionId });
-      skills.resolveRuntimeSnapshot({ orgId, agentId, runId: `run_${randomUUID().replaceAll("-", "")}` });
+      skills.resolveRuntimeSnapshot({ orgId, agentId, runId: `run_${randomUUID().replaceAll("-", "")}`, maxTokens: 44_800 });
       suiteId = (connection.db.prepare("select id from eval_suites where org_id = ? and key = ?")
         .get(orgId, EVAL_SUITE_KEY.INTEGRATION) as { id: string }).id;
       evalRunId = sdlc.runEvalSuite({ orgId, suiteId });

tests/unit/agent-goal-tools.test.ts

@@ -65,7 +65,7 @@ function call(args: Record<string, unknown>): ToolCall {
   return { id: `call_${randomUUID()}`, name: "complete_goal", arguments: args };
 }
 
-function argsFor(h: Harness, agentId: string, auditor: Auditor, toolCall: ToolCall): GoalCompletionArgs {
+function argsFor(h: Harness, agentId: string, auditor: Auditor, toolCall: ToolCall, inputBudgetTokens = 112_000): GoalCompletionArgs {
   const ctx: AgentToolContext = {
     orgId: h.orgId,
     agent: { id: agentId, modelProviderId: h.providerId, modelId: "gpt-5.4" },
@@ -73,7 +73,7 @@ function argsFor(h: Harness, agentId: string, auditor: Auditor, toolCall: ToolCa
     taskGoalId: h.goalId, taskId: null, sourceThreadId: null,
     workspace: { id: "wsp_audit", localPath: workspaceRoot, writeAccess: true },
   };
-  return { deps: h.deps, ctx, providers: { runToolTurn: auditor }, providerId: h.providerId, workspaceId: null, call: toolCall };
+  return { deps: h.deps, ctx, providers: { runToolTurn: auditor }, providerId: h.providerId, workspaceId: null, inputBudgetTokens, call: toolCall };
 }
 
 function goalStatus(h: Harness): string {
@@ -265,4 +265,18 @@ describe("company_status", () => {
     expect(output).toContain("Ask the CEO");
     h.connection.close();
   });
+
+  test("an audit that outgrows the context budget fails as an error instead of sending an oversized request", async () => {
+    const h = createHarness();
+    let calls = 0;
+    const auditor = async (): Promise<AuditorTurn> => {
+      calls += 1;
+      return { text: "", toolCalls: [], rawResponseJson: "{}", requestId: "r" };
+    };
+    const output = await completeGoalTool(argsFor(h, h.ceoId, auditor, call({ evidence: "证".repeat(5_000) }), 3_000));
+    expect(output).toMatch(/^ERROR: This run needs ~\d+ input tokens but the provider's context window leaves 3000/);
+    expect(calls).toBe(0);
+    expect(goalStatus(h)).toBe("active");
+    h.connection.close();
+  });
 });

tests/unit/agent-run-loop.test.ts

@@ -92,7 +92,7 @@ describe("agent run loop (real tool dispatch)", () => {
         tools: { db: connection.db, organization, chat: new ChatService(connection.db) },
         onToolResults: () => { toolResultNotifications += 1; },
       },
-      { orgId, providerId, workspaceId, instructions: "You are the CEO.", initialUserTurn: "Start the company.", toolContext },
+      { orgId, providerId, inputBudgetTokens: 112_000, workspaceId, instructions: "You are the CEO.", initialUserTurn: "Start the company.", toolContext },
     );
 
     expect(result.iterations).toBe(4);
@@ -147,7 +147,7 @@ describe("agent run loop (real tool dispatch)", () => {
     ];
     const result = await runAgentLoop(
       { providers: fakeProvider(script, connection.db), tools: { db: connection.db, organization, chat: new ChatService(connection.db) } },
-      { orgId, providerId, workspaceId: null, instructions: "i", initialUserTurn: "u", toolContext: ctx },
+      { orgId, providerId, inputBudgetTokens: 112_000, workspaceId: null, instructions: "i", initialUserTurn: "u", toolContext: ctx },
     );
     // The tool returns an error string to the model; nothing is written outside the root.
     expect(existsSync(join(workspaceRoot, "..", "escape.txt"))).toBe(false);
@@ -184,7 +184,7 @@ describe("agent run loop (real tool dispatch)", () => {
 
     await runAgentLoop(
       { providers: fakeProvider(script, connection.db), tools: { db: connection.db, organization, chat: new ChatService(connection.db) } },
-      { orgId, providerId, workspaceId: null, instructions: "i", initialUserTurn: "u", toolContext: ctx },
+      { orgId, providerId, inputBudgetTokens: 112_000, workspaceId: null, instructions: "i", initialUserTurn: "u", toolContext: ctx },
     );
 
     const alex = connection.db.prepare("select id from agents where org_id = ? and name = 'Alex Rivera'").get(orgId) as { id: string };
@@ -230,7 +230,7 @@ describe("agent run loop (real tool dispatch)", () => {
     };
     await expect(runAgentLoop(
       { providers: fakeProvider([{ text: "", toolCalls: [] }], connection.db), tools: { db: connection.db, organization, chat: new ChatService(connection.db) } },
-      { orgId, providerId, workspaceId: null, instructions: "i", initialUserTurn: "u", toolContext: ctx },
+      { orgId, providerId, inputBudgetTokens: 112_000, workspaceId: null, instructions: "i", initialUserTurn: "u", toolContext: ctx },
     )).rejects.toThrow(/empty turn/);
     connection.close();
   });

tests/unit/builtin-skills.test.ts

@@ -6,7 +6,10 @@ import { describe, expect, test } from "vitest";
 import { DatabaseConnection } from "../../src/main/database/connection";
 import { OrganizationService } from "../../src/main/services/organization-service";
 import { ChatService } from "../../src/main/services/chat-service";
-import { SkillService, MAX_COMPILED_PROMPT_CHARS } from "../../src/main/services/skill-service";
+import { SkillService } from "../../src/main/services/skill-service";
+import { skillBudgetTokens } from "../../src/main/services/context-budget";
+
+const SKILL_TOKENS = skillBudgetTokens(128_000);
 import { importStagedSkills, parseFrontmatter, stageBuiltinSkillsLibrary } from "../../src/main/services/builtin-skills";
 import { dispatchAgentTool, type AgentToolContext } from "../../src/main/services/agent-tools";
 
@@ -141,7 +144,7 @@ describe("built-in skills", () => {
     // A third, oversized flat skill to prove the compiled-prompt cap.
     writeFileSync(
       join(h.sourceRoot, "engineering", "engineering-epic-novelist.md"),
-      `---\nname: Epic Novelist\ndescription: too big to pack\n---\n\n${"x".repeat(MAX_COMPILED_PROMPT_CHARS + 10)}\n`,
+      `---\nname: Epic Novelist\ndescription: too big to pack\n---\n\n${"x".repeat(SKILL_TOKENS * 3 + 30)}\n`,
       "utf8",
     );
     const staged = stageBuiltinSkillsLibrary(h.sourceRoot, h.libraryRoot);
@@ -175,17 +178,17 @@ describe("built-in skills", () => {
     expect(hired).toContain("Unknown skills skipped: no-such-skill");
 
     const danaId = (h.connection.db.prepare("select id from agents where org_id = ? and name = 'Dana Lee'").get(h.orgId) as { id: string }).id;
-    const danaSnapshot = h.connection.transaction(() => h.skills.resolveRuntimeSnapshot({ orgId: h.orgId, agentId: danaId, runId: "run_dana" }));
+    const danaSnapshot = h.connection.transaction(() => h.skills.resolveRuntimeSnapshot({ orgId: h.orgId, agentId: danaId, runId: "run_dana", maxTokens: SKILL_TOKENS }));
     const danaPrompt = (h.connection.db.prepare("select compiled_prompt from skill_runtime_snapshots where id = ?").get(danaSnapshot) as { compiled_prompt: string }).compiled_prompt;
     expect(danaPrompt.indexOf("# Skill: humanizer")).toBe(0);
     expect(danaPrompt).toContain("# Skill: Backend Architect");
     expect(danaPrompt).toContain("Skills omitted for size");
     expect(danaPrompt).toContain("Epic Novelist");
-    expect(danaPrompt.length).toBeLessThan(MAX_COMPILED_PROMPT_CHARS + 5_000);
+    expect(danaPrompt.length).toBeLessThan(SKILL_TOKENS * 3 + 5_000);
 
     // The CEO carries humanizer too - forced by resolution, not by binding.
     expect(h.skills.hasRuntimeSkills(h.orgId, ceoId)).toBe(true);
-    const ceoSnapshot = h.connection.transaction(() => h.skills.resolveRuntimeSnapshot({ orgId: h.orgId, agentId: ceoId, runId: "run_ceo" }));
+    const ceoSnapshot = h.connection.transaction(() => h.skills.resolveRuntimeSnapshot({ orgId: h.orgId, agentId: ceoId, runId: "run_ceo", maxTokens: SKILL_TOKENS }));
     const ceoPrompt = (h.connection.db.prepare("select compiled_prompt from skill_runtime_snapshots where id = ?").get(ceoSnapshot) as { compiled_prompt: string }).compiled_prompt;
     expect(ceoPrompt).toContain("# Skill: humanizer");
     expect(ceoPrompt).not.toContain("# Skill: Backend Architect");
@@ -210,7 +213,7 @@ describe("built-in skills", () => {
       });
     });
     const promptFor = (runId: string) => {
-      const snapshotId = h.connection.transaction(() => h.skills.resolveRuntimeSnapshot({ orgId: h.orgId, agentId, runId }));
+      const snapshotId = h.connection.transaction(() => h.skills.resolveRuntimeSnapshot({ orgId: h.orgId, agentId, runId, maxTokens: SKILL_TOKENS }));
       return (h.connection.db.prepare("select compiled_prompt from skill_runtime_snapshots where id = ?").get(snapshotId) as { compiled_prompt: string }).compiled_prompt;
     };
 

tests/unit/context-budget.test.ts

@@ -0,0 +1,97 @@
+import { randomUUID } from "node:crypto";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { describe, expect, test } from "vitest";
+import { estimateTokens, inputBudgetTokens, skillBudgetTokens } from "../../src/main/services/context-budget";
+import { runAgentLoop } from "../../src/main/services/agent-run-loop";
+import { DatabaseConnection } from "../../src/main/database/connection";
+import { OrganizationService } from "../../src/main/services/organization-service";
+import { ChatService } from "../../src/main/services/chat-service";
+import { IntegrationKeyStore } from "../../src/main/services/integration-key-store";
+import { ProviderService } from "../../src/main/services/provider-service";
+import type { AgentTurnMessage, AgentTurnResult } from "../../src/main/services/provider-tool-turn";
+import { AGENT_TOOLS } from "../../src/main/services/agent-tools";
+import { WORKSPACE_TOOLS } from "../../src/main/services/agent-workspace-tools";
+import { WEB_TOOLS } from "../../src/main/services/agent-web-tools";
+
+const FIXED_TOKENS = estimateTokens("i") + estimateTokens(JSON.stringify([...AGENT_TOOLS, ...WORKSPACE_TOOLS, ...WEB_TOOLS]));
+
+describe("token estimates", () => {
+  test("count CJK per character and other text per three characters", () => {
+    expect(estimateTokens("abcdef")).toBe(2);
+    expect(estimateTokens("abcdefg")).toBe(3);
+    expect(estimateTokens("技能规则")).toBe(4);
+    expect(estimateTokens("规则 rule")).toBe(4);
+    expect(inputBudgetTokens(128_000)).toBe(112_000);
+    expect(skillBudgetTokens(128_000)).toBe(44_800);
+  });
+});
+
+async function loop(budget: number, sent: AgentTurnMessage[][]) {
+  const connection = new DatabaseConnection(join(tmpdir(), `ac-budget-${randomUUID()}.db`));
+  const organization = new OrganizationService(connection.db);
+  let orgId = "";
+  let agentId = "";
+  connection.transaction(() => {
+    orgId = organization.createOrganization({ organizationName: "B", employeeName: "Y", mission: "m", goal: "g" });
+    agentId = organization.createAgent({ orgId, name: "Ana", role: "Analyst", department: "Operations", modelProviderId: null, modelId: null, budgetCents: 1000, managerAgentId: null });
+  });
+  const chat = new ChatService(connection.db);
+  const general = connection.db.prepare("select id from chat_spaces where org_id = ? and name = 'general'").get(orgId) as { id: string };
+  chat.sendMessage({ orgId, spaceId: general.id, threadId: null, senderType: "agent", senderId: agentId, bodyMd: "y".repeat(9000), attachmentFileIds: [] });
+  let turn = 0;
+  try {
+    return await runAgentLoop(
+      {
+        providers: {
+          integrationKeys: new IntegrationKeyStore(connection.db),
+          runToolTurn: async (input): Promise<AgentTurnResult & { requestId: string }> => {
+            sent.push(input.messages.map((message) => ({ ...message })));
+            turn += 1;
+            return turn <= 2
+              ? { text: "", toolCalls: [{ id: `c${turn}`, name: "read_messages", arguments: { channel: "general" } }], rawResponseJson: "{}", requestId: `r${turn}` }
+              : { text: "done", toolCalls: [], rawResponseJson: "{}", requestId: `r${turn}` };
+          },
+        },
+        tools: { db: connection.db, organization, chat },
+      },
+      {
+        orgId, providerId: "p", inputBudgetTokens: budget, workspaceId: null, instructions: "i", initialUserTurn: "go",
+        toolContext: { orgId, agent: { id: agentId, modelProviderId: null, modelId: null }, taskProjectId: null, taskGoalId: null, taskId: null, sourceThreadId: null, workspace: null },
+      },
+    );
+  } finally {
+    connection.close();
+  }
+}
+
+describe("run context budget", () => {
+  test("over budget, older tool results are elided before the next call", async () => {
+    const sent: AgentTurnMessage[][] = [];
+    const result = await loop(FIXED_TOKENS + 5_000, sent);
+    expect(result.finalText).toBe("done");
+    const outputs = sent[2].flatMap((message) => (message.role === "tool" ? message.results.map((item) => item.output) : []));
+    expect(outputs[0]).toMatch(/^\[Earlier result of read_messages elided/);
+    expect(outputs[1]).toContain("y".repeat(9000));
+  });
+
+  test("a request that cannot fit even after compaction fails loudly with the numbers", async () => {
+    await expect(loop(2_000, [])).rejects.toThrow(/This run needs ~\d+ input tokens but the provider's context window leaves 2000/);
+  });
+
+  test("the provider context window has a floor", () => {
+    const connection = new DatabaseConnection(join(tmpdir(), `ac-window-${randomUUID()}.db`));
+    const organization = new OrganizationService(connection.db);
+    let orgId = "";
+    connection.transaction(() => {
+      orgId = organization.createOrganization({ organizationName: "W", employeeName: "Y", mission: "m", goal: "g" });
+    });
+    const providerId = (connection.db.prepare("select id from providers where org_id = ? and provider_key = 'openai'").get(orgId) as { id: string }).id;
+    const providers = new ProviderService(connection.db);
+    const policy = { orgId, providerId, dataPolicy: "standard" as const, backgroundEnabled: false, monthlyRequestLimit: 100, workspaceScope: "organization" as const };
+    expect(() => providers.updatePolicy({ ...policy, contextWindowTokens: 8_000 })).toThrow("Context window must be at least 24000 tokens.");
+    providers.updatePolicy({ ...policy, contextWindowTokens: 200_000 });
+    expect(connection.db.prepare("select context_window_tokens from providers where id = ?").get(providerId)).toEqual({ context_window_tokens: 200_000 });
+    connection.close();
+  });
+});

tests/unit/domain-services.test.ts

@@ -24,7 +24,7 @@ describe("domain services", () => {
     const connection = createConnection();
 
     const rows = connection.db.prepare("select version from schema_migrations order by version asc").all() as Array<{ version: number }>;
-    expect(rows.map((row) => row.version)).toEqual([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61]);
+    expect(rows.map((row) => row.version)).toEqual([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62]);
     connection.close();
   });
 
@@ -186,7 +186,7 @@ describe("domain services", () => {
       skills.bind({ orgId, agentId, skillId: skill.id, revisionId });
       rmSync(skillDir, { recursive: true, force: true });
       skills.runTests({ orgId, revisionId });
-      skills.resolveRuntimeSnapshot({ orgId, agentId, runId: "run-review-1" });
+      skills.resolveRuntimeSnapshot({ orgId, agentId, runId: "run-review-1", maxTokens: 44_800 });
     });
 
     // Catalogue metadata ships in the snapshot; skill internals stay in the
@@ -247,10 +247,10 @@ describe("domain services", () => {
         .get(orgA, "review-routes") as { id: string }).id;
       expect(() => skills.bind({ orgId: orgA, agentId: agentB, skillId, revisionId })).toThrow("Agent was not found.");
       skills.bind({ orgId: orgA, agentId: agentA, skillId, revisionId });
-      const firstResolve = skills.resolveRuntimeSnapshot({ orgId: orgA, agentId: agentA, runId: "run-review-1" });
+      const firstResolve = skills.resolveRuntimeSnapshot({ orgId: orgA, agentId: agentA, runId: "run-review-1", maxTokens: 44_800 });
       // Idempotent per run: a recovered run reuses its frozen snapshot
       // instead of failing on the uniqueness guard.
-      expect(skills.resolveRuntimeSnapshot({ orgId: orgA, agentId: agentA, runId: "run-review-1" })).toBe(firstResolve);
+      expect(skills.resolveRuntimeSnapshot({ orgId: orgA, agentId: agentA, runId: "run-review-1", maxTokens: 44_800 })).toBe(firstResolve);
       expect(connection.db.prepare("select count(*) as n from skill_runtime_snapshots where run_id = 'run-review-1'").get())
         .toEqual({ n: 1 });
     });

tests/unit/operations-service.test.ts

@@ -85,7 +85,7 @@ describe("operations service", () => {
         dataPolicy: PROVIDER_DATA_POLICY.STANDARD,
         backgroundEnabled: true,
         monthlyRequestLimit: 10,
-        workspaceScope: "organization",
+        workspaceScope: "organization", contextWindowTokens: 128_000,
       });
       operations.setFeatureFlag({ orgId, flagId, enabled: false });
     });

tests/unit/provider-background-service.test.ts

@@ -52,7 +52,7 @@ describe("provider background service", () => {
         dataPolicy: PROVIDER_DATA_POLICY.STANDARD,
         backgroundEnabled: true,
         monthlyRequestLimit: 10,
-        workspaceScope: "organization",
+        workspaceScope: "organization", contextWindowTokens: 128_000,
       });
     });
 
@@ -115,7 +115,7 @@ describe("provider background service", () => {
         dataPolicy: PROVIDER_DATA_POLICY.STANDARD,
         backgroundEnabled: true,
         monthlyRequestLimit: 10,
-        workspaceScope: "organization",
+        workspaceScope: "organization", contextWindowTokens: 128_000,
       });
     });
 
@@ -152,7 +152,7 @@ describe("provider background service", () => {
         dataPolicy: PROVIDER_DATA_POLICY.STANDARD,
         backgroundEnabled: true,
         monthlyRequestLimit: 10,
-        workspaceScope: "organization",
+        workspaceScope: "organization", contextWindowTokens: 128_000,
       });
     });
 

tests/unit/provider-service.test.ts

@@ -448,7 +448,7 @@ describe("provider service", () => {
         dataPolicy: PROVIDER_DATA_POLICY.STANDARD,
         backgroundEnabled: false,
         monthlyRequestLimit: 1,
-        workspaceScope: "organization",
+        workspaceScope: "organization", contextWindowTokens: 128_000,
       });
     });
     await providers.createResponse({
@@ -524,7 +524,7 @@ describe("provider service", () => {
         dataPolicy: PROVIDER_DATA_POLICY.STANDARD,
         backgroundEnabled: false,
         monthlyRequestLimit: 10,
-        workspaceScope: "workspace",
+        workspaceScope: "workspace", contextWindowTokens: 128_000,
       });
     });
 
@@ -582,7 +582,7 @@ describe("provider service", () => {
         dataPolicy: PROVIDER_DATA_POLICY.ZERO_DATA_RETENTION,
         backgroundEnabled: false,
         monthlyRequestLimit: 10,
-        workspaceScope: "organization",
+        workspaceScope: "organization", contextWindowTokens: 128_000,
       });
     });
 

tests/unit/run-context-compaction.test.ts

@@ -72,7 +72,7 @@ describe("tool history compaction", () => {
         tools: { db: connection.db, organization, chat },
       },
       {
-        orgId, providerId: "p", workspaceId: null, instructions: "i", initialUserTurn: "go",
+        orgId, providerId: "p", inputBudgetTokens: 112_000, workspaceId: null, instructions: "i", initialUserTurn: "go",
         toolContext: { orgId, agent: { id: agentId, modelProviderId: null, modelId: null }, taskProjectId: null, taskGoalId: null, taskId: null, sourceThreadId: null, workspace: null },
       },
     );

tests/unit/task-execution-service.test.ts

@@ -408,7 +408,7 @@ describe("task execution service", () => {
         dataPolicy: PROVIDER_DATA_POLICY.STANDARD,
         backgroundEnabled: false,
         monthlyRequestLimit: 5,
-        workspaceScope: "organization",
+        workspaceScope: "organization", contextWindowTokens: 128_000,
       });
       const agentId = organization.createAgent({
         orgId,
@@ -452,7 +452,7 @@ describe("task execution service", () => {
       dataPolicy: PROVIDER_DATA_POLICY.STANDARD,
       backgroundEnabled: false,
       monthlyRequestLimit: 50,
-      workspaceScope: "organization",
+      workspaceScope: "organization", contextWindowTokens: 128_000,
     }));
     const enqueued = connection.transaction(() => execution.enqueueHeartbeatRuns());
     expect(enqueued).toHaveLength(1);