@saadnvd1/agent-os 0.3.472 → 0.3.474

package/lib/agents/launch.ts

@@ -52,6 +52,14 @@ export function envArgs(env: Record<string, string>): string[] {
   return Object.entries(env).flatMap(([k, v]) => ["-e", `${k}=${v}`]);
 }
 
+// Tool search is off on a routed provider, so every MCP server the user or
+// the project has would put all its schemas in every prompt. A terminal
+// agent reaches AgentOS through the aos CLI, so it needs none. Shell text
+// for the agent's command line; empty on Claude.
+export function routedMcpFlags(provider?: string | null): string {
+  return provider ? "--strict-mcp-config " : "";
+}
+
 // Written just before the agent's tmux session is created: a session with
 // no prompt file is not one launchClaude started.
 export const promptFileFor = (sessionId: string) =>
@@ -97,7 +105,8 @@ export async function launchClaude(opts: {
     model: opts.model,
     model_provider: opts.provider ?? null,
   });
-  const agent = `export PATH=${shellQuote(AOS_BIN_DIR)}:"$HOME/.local/bin:$PATH"; ${routing.exportKey} ${provider.command} ${flags} ${CLAUDE_STATUS_SETTINGS_FLAG} --append-system-prompt-file ${shellQuote(briefFile)} "$(cat ${shellQuote(promptFile)})"; exec "\${SHELL:-/bin/sh}" -l`;
+  const mcpFlags = routedMcpFlags(opts.provider);
+  const agent = `export PATH=${shellQuote(AOS_BIN_DIR)}:"$HOME/.local/bin:$PATH"; ${routing.exportKey} ${provider.command} ${flags} ${mcpFlags}${CLAUDE_STATUS_SETTINGS_FLAG} --append-system-prompt-file ${shellQuote(briefFile)} "$(cat ${shellQuote(promptFile)})"; exec "\${SHELL:-/bin/sh}" -l`;
   await execFileAsync(
     "tmux",
     [

package/lib/chat/driver.ts

@@ -30,6 +30,9 @@ export interface ChatStartOptions {
   // Tools served in-process for this conversation, and tools it may use
   // without asking.
   mcpServers?: Record<string, McpServerConfig>;
+  // Load no MCP servers but those above: the user's own would otherwise
+  // put every tool schema in each prompt.
+  strictMcpConfig?: boolean;
   allowedTools?: string[];
   disallowedTools?: string[];
   // A fixed permission mode in place of the access setting, which then

package/lib/chat/drivers/claude.ts

@@ -188,6 +188,7 @@ export const claudeDriver: ChatDriver = {
         allowDangerouslySkipPermissions: true,
         canUseTool: approvals.canUseTool,
         mcpServers: options.mcpServers,
+        strictMcpConfig: options.strictMcpConfig,
         allowedTools: options.allowedTools,
         disallowedTools: options.disallowedTools,
         hooks: {

package/lib/chat/refresh.ts

@@ -11,6 +11,7 @@
  */
 
 import { db } from "../db";
+import { effectiveModels } from "../models/config";
 import { MAX_REFRESH_TOKENS, MIN_REFRESH_TOKENS } from "./refresh-limits";
 
 export const ORCHESTRATOR_KEY = "orchestrator.refresh_tokens";
@@ -87,6 +88,25 @@ export function setRefreshSettings(input: Partial<RefreshSettings>): void {
 // as 150 and clamps anything under 100k up to it).
 export const AUTO_COMPACT_ENV = "CLAUDE_CODE_AUTO_COMPACT_WINDOW";
 const AUTO_COMPACT_MIN = 100_000;
+// Claude Code turns tool search off for a host that isn't Anthropic's, so
+// every MCP and deferred tool schema rides in each prompt: measured on
+// DeepSeek, the first call after a compaction is ~100k tokens against ~53k
+// on Claude. A window measured from zero left it 15-30k of room, so it
+// compacted every couple of minutes. Providers get that extra on top.
+export const PROVIDER_BASE_TOKENS = 60_000;
+
+// The compact threshold for a session on this provider (null: Anthropic).
+export function compactThreshold(
+  tokens: number,
+  provider: string | null | undefined
+): number {
+  if (tokens <= 0 || !provider) return tokens;
+  const window = effectiveModels().providers.find(
+    (p) => p.id === provider
+  )?.contextWindow;
+  const wanted = tokens + PROVIDER_BASE_TOKENS;
+  return window ? Math.min(wanted, window) : wanted;
+}
 
 // What an agent is launched with so it compacts itself at the threshold.
 // Not an orchestrator: it starts over from its brief at its own threshold
@@ -95,11 +115,17 @@ const AUTO_COMPACT_MIN = 100_000;
 // the user's own environment still applies.
 export function autoCompactEnv(sessionId: string): Record<string, string> {
   const row = db
-    .prepare(`SELECT role FROM sessions WHERE id = ?`)
-    .get(sessionId) as { role: string | null } | undefined;
+    .prepare(`SELECT role, model_provider FROM sessions WHERE id = ?`)
+    .get(sessionId) as
+    | { role: string | null; model_provider: string | null }
+    | undefined;
   const tokens = refreshSettings().compactTokens;
   if (row?.role === "orchestrator" || tokens <= 0) return {};
-  return { [AUTO_COMPACT_ENV]: String(Math.max(AUTO_COMPACT_MIN, tokens)) };
+  const window = compactThreshold(
+    Math.max(AUTO_COMPACT_MIN, tokens),
+    row?.model_provider
+  );
+  return { [AUTO_COMPACT_ENV]: String(window) };
 }
 
 const k = (tokens: number) => `${Math.round(tokens / 1000)}k`;

package/lib/chat/worker/host.ts

@@ -40,7 +40,7 @@ import {
   visualsTools,
 } from "../../artifacts/tools";
 import type { RoleExtras } from "./extras";
-import { refreshSettings, refreshedNote } from "../refresh";
+import { compactThreshold, refreshSettings, refreshedNote } from "../refresh";
 import { compactionDue, staleLabel, type Staleness } from "../stale";
 import type { WorkerCommand, WorkerEvent } from "./protocol";
 
@@ -203,6 +203,10 @@ export class ChatHost {
       mcpServers: visuals
         ? { ...extras.mcpServers, [VISUALS_SERVER]: visuals }
         : extras.mcpServers,
+      // Claude Code turns tool search off on a non-Anthropic host, so each
+      // MCP server the user has adds all its schemas to every prompt
+      // (~45k tokens): a routed agent gets only the servers AgentOS gives it.
+      strictMcpConfig: !!session.model_provider,
       allowedTools: extras.allowedTools,
       disallowedTools: extras.disallowedTools,
       permissionMode: extras.permissionMode,
@@ -743,7 +747,10 @@ export class ChatHost {
     }
     return compactionDue(context, {
       state: this.state,
-      thresholdTokens: refreshSettings().compactTokens,
+      thresholdTokens: compactThreshold(
+        refreshSettings().compactTokens,
+        this.session.model_provider
+      ),
     });
   }
 

package/lib/models/config.ts

@@ -100,6 +100,7 @@ const providerSchema = z.strictObject({
       z.string().max(200)
     )
     .optional(),
+  contextWindow: z.number().int().min(100_000).max(10_000_000).optional(),
   enabled: z.boolean(),
 });
 

package/lib/models/providers.ts

@@ -37,6 +37,9 @@ export interface ModelProvider {
   tiers?: ModelTiers;
   // Plain settings the endpoint wants (a longer timeout, say). Never a key.
   env?: Record<string, string>;
+  // The context window of its models, in tokens: Claude Code assumes a
+  // size from the model's name, and the compact window is capped at this.
+  contextWindow?: number;
   enabled: boolean;
 }
 
@@ -77,6 +80,9 @@ export const PROVIDER_PRESETS: ModelProvider[] = [
       sonnet: "deepseek-flash",
       haiku: "deepseek-flash",
     },
+    // 1M for deepseek-flash (api-docs.deepseek.com/quick_start/pricing);
+    // deepseek-v4-pro's isn't stated there, and is assumed to match.
+    contextWindow: 1_000_000,
     enabled: false,
   },
   {
@@ -88,6 +94,8 @@ export const PROVIDER_PRESETS: ModelProvider[] = [
     models: ["glm-5.3", "glm-5.3-flash"],
     tiers: { opus: "glm-5.3", sonnet: "glm-5.3", haiku: "glm-5.3-flash" },
     env: { API_TIMEOUT_MS: "3000000" },
+    // docs.z.ai/guides/llm/glm-5.3
+    contextWindow: 1_000_000,
     enabled: false,
   },
 ];

package/lib/models/resolve.ts

@@ -218,6 +218,9 @@ function routingEnv(p: ModelProvider, model: string): Record<string, string> {
       : "ANTHROPIC_API_KEY";
   return {
     ...p.env,
+    ...(p.contextWindow
+      ? { CLAUDE_CODE_MAX_CONTEXT_TOKENS: String(p.contextWindow) }
+      : {}),
     ANTHROPIC_BASE_URL: p.baseUrl!,
     [other]: "",
     ANTHROPIC_DEFAULT_OPUS_MODEL: p.tiers?.opus ?? model,

package/lib/orchestrator/tool-schemas.ts

@@ -27,10 +27,15 @@ const modelRef = z
 const modelOverride = z
   .union([
     modelRef,
-    z.partialRecord(
-      z.enum(TASK_ROLES as [ModelRole, ...ModelRole[]]),
-      modelRef
-    ),
+    // Named optional keys, not z.partialRecord: its "propertyNames" makes
+    // Claude Code drop every tool on this server, not just this one.
+    z
+      .object(
+        Object.fromEntries(
+          TASK_ROLES.map((role) => [role, modelRef.optional()])
+        ) as Record<ModelRole, z.ZodOptional<typeof modelRef>>
+      )
+      .strict(),
   ])
   .optional()
   .describe(

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@saadnvd1/agent-os",
-  "version": "0.3.472",
+  "version": "0.3.474",
   "description": "Self-hosted web UI for managing Claude Code sessions",
   "bin": {
     "agent-os": "scripts/agent-os"