@opengeni/runtime 4.0.3 → 4.1.0-canary.36239117573001

package/dist/sandbox/index.js

@@ -13,6 +13,7 @@ import {
   BrowserControlUnsupportedError,
   CAPABILITY_DESCRIPTORS,
   CODEMODE_TOKEN_TTL_SECONDS,
+  CODE_SEARCH_RG_MAX_BYTES,
   ChannelAConflictError,
   ChannelAFileSystemRouteChangedError,
   ChannelANotFoundError,
@@ -258,6 +259,7 @@ import {
   timeoutAgentError,
   timeoutControlResponse,
   uploadRecordingArtifact,
+  validateCodeSearchRipgrepArgs,
   validateComputerControlFrameEvidence,
   verifyRestoredWorkspace,
   verifySandboxExecReadiness,
@@ -269,7 +271,7 @@ import {
   withRunCredentialsClient,
   withRunCredentialsSession,
   withoutSandboxProviderIdentity
-} from "../chunk-GBVN7WSK.js";
+} from "../chunk-ORTF5RRC.js";
 import "../chunk-PRHMSM6C.js";
 import "../chunk-4A4XTKKX.js";
 import "../chunk-JAYHA6BF.js";
@@ -290,6 +292,7 @@ export {
   BrowserControlUnsupportedError,
   CAPABILITY_DESCRIPTORS,
   CODEMODE_TOKEN_TTL_SECONDS,
+  CODE_SEARCH_RG_MAX_BYTES,
   ChannelAConflictError,
   ChannelAFileSystemRouteChangedError,
   ChannelANotFoundError,
@@ -535,6 +538,7 @@ export {
   timeoutAgentError,
   timeoutControlResponse,
   uploadRecordingArtifact,
+  validateCodeSearchRipgrepArgs,
   validateComputerControlFrameEvidence,
   verifyRestoredWorkspace,
   verifySandboxExecReadiness,

package/dist/tool-result-spill.d.ts

@@ -1,5 +1,5 @@
 import type { AttemptToolDefinition, AttemptToolExecutionContext } from "@opengeni/codemode";
-import { type AttemptToolResult as AttemptToolResultValue, type ToolResultSpilledReceipt as ToolResultSpilledReceiptValue } from "@opengeni/contracts";
+import { type AttemptToolIdentity, type AttemptToolResult as AttemptToolResultValue, type ToolResultSpilledReceipt as ToolResultSpilledReceiptValue } from "@opengeni/contracts";
 export type SpillOversizedModelToolResult = (input: {
     operationId: string;
     result: AttemptToolResultValue;
@@ -7,6 +7,14 @@ export type SpillOversizedModelToolResult = (input: {
 }) => Promise<AttemptToolResultValue>;
 export declare function modelToolResultOverflowError(): AttemptToolResultValue;
 export declare function spilledModelToolResult(receipt: ToolResultSpilledReceiptValue): AttemptToolResultValue;
-export declare function projectAttemptToolResultForCaller(result: AttemptToolResultValue, context: AttemptToolExecutionContext, spill?: SpillOversizedModelToolResult): Promise<AttemptToolResultValue>;
-export declare function wrapAttemptToolExecute(execute: AttemptToolDefinition["execute"], spill?: SpillOversizedModelToolResult): AttemptToolDefinition["execute"];
+/**
+ * The single per-caller seam over one executor result. Codemode receives the
+ * exact result. The model receives its model-visible projection (compact
+ * Knowledge discovery output for the exact tool identity) when that fits in
+ * 1 MiB; otherwise the exact result is spilled to a file, as for any tool.
+ * MCP-backed tools are already bounded on their exact result by the MCP
+ * transport cap before this seam runs.
+ */
+export declare function projectAttemptToolResultForCaller(result: AttemptToolResultValue, context: AttemptToolExecutionContext, spill?: SpillOversizedModelToolResult, identity?: AttemptToolIdentity): Promise<AttemptToolResultValue>;
+export declare function wrapAttemptToolExecute(execute: AttemptToolDefinition["execute"], spill?: SpillOversizedModelToolResult, identity?: AttemptToolIdentity): AttemptToolDefinition["execute"];
 export declare function wrapAttemptToolDefinitions(definitions: readonly AttemptToolDefinition[], spill?: SpillOversizedModelToolResult): AttemptToolDefinition[];

package/dist/workspace-tool-gateway.js

@@ -1,7 +1,7 @@
 import {
   prepareAgentTools
-} from "./chunk-LL2YQO5H.js";
-import "./chunk-GBVN7WSK.js";
+} from "./chunk-YAA7WDNO.js";
+import "./chunk-ORTF5RRC.js";
 import "./chunk-PRHMSM6C.js";
 import "./chunk-4A4XTKKX.js";
 import "./chunk-JAYHA6BF.js";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@opengeni/runtime",
-  "version": "4.0.3",
+  "version": "4.1.0-canary.36239117573001",
   "license": "Apache-2.0",
   "repository": {
     "type": "git",
@@ -63,16 +63,16 @@
     "@openai/agents": "0.14.3",
     "@openai/agents-core": "0.14.3",
     "@openai/agents-extensions": "0.14.3",
-    "@opengeni/agent-proto": "^0.6.0",
-    "@opengeni/capabilities": "^0.3.4",
-    "@opengeni/codemode": "^0.6.2",
-    "@opengeni/codex": "^0.2.26",
-    "@opengeni/config": "^2.1.1",
-    "@opengeni/contracts": "^5.1.1",
-    "@opengeni/network": "^0.3.1",
-    "@opengeni/sdk": "^7.1.1",
-    "@opengeni/tool-gateway": "^0.1.13",
-    "@opengeni/xai-subscription": "^0.1.4",
+    "@opengeni/agent-proto": "^0.6.0-canary.36239117573001",
+    "@opengeni/capabilities": "^0.3.4-canary.36239117573001",
+    "@opengeni/codemode": "^0.6.3-canary.36239117573001",
+    "@opengeni/codex": "^0.2.27-canary.36239117573001",
+    "@opengeni/config": "^3.0.0-canary.36239117573001",
+    "@opengeni/contracts": "^5.2.0-canary.36239117573001",
+    "@opengeni/network": "^0.3.1-canary.36239117573001",
+    "@opengeni/sdk": "^7.2.0-canary.36239117573001",
+    "@opengeni/tool-gateway": "^0.1.14-canary.36239117573001",
+    "@opengeni/xai-subscription": "^0.1.4-canary.36239117573001",
     "debug": "4.4.3",
     "modal": "0.9.0",
     "openai": "6.47.0",

package/src/codex-tool-search.ts

@@ -185,6 +185,41 @@ export function bm25RankTools(tools: Tool[], query: string, limit: number): Tool
   return hits.slice(0, limit).map((s) => s.tool); // no hits ⇒ [] (codex-rs parity; see doc)
 }
 
+/**
+ * Tools a free-text query names exactly, in query order. A query token that
+ * contains "_" names a tool when it equals the tool's full model name or the
+ * part after its last "__" server separator (`goal_complete` and
+ * `opengeni__goal_complete` both name `opengeni__goal_complete`). Single words
+ * such as "read" are never treated as names. Agents often search by exact name,
+ * and BM25 alone ranked a neighbouring tool first for about a fifth of those
+ * queries (`goal_complete` returned `goal_set` first).
+ */
+export function toolsNamedInQuery<T extends Tool & { name: string }>(
+  tools: T[],
+  query: string,
+): T[] {
+  const byName = new Map<string, T[]>();
+  for (const tool of tools) {
+    const full = tool.name.toLowerCase();
+    const separator = full.lastIndexOf(MCP_TOOL_NAME_SEPARATOR);
+    const short = separator >= 0 ? full.slice(separator + MCP_TOOL_NAME_SEPARATOR.length) : full;
+    for (const key of new Set([full, short])) {
+      const named = byName.get(key) ?? [];
+      named.push(tool);
+      byName.set(key, named);
+    }
+  }
+  const named: T[] = [];
+  for (const raw of query.toLowerCase().split(/[^a-z0-9_]+/)) {
+    const token = raw.replace(/^_+|_+$/g, "");
+    if (!token.includes("_")) continue;
+    for (const tool of byName.get(token) ?? []) {
+      if (!named.includes(tool)) named.push(tool);
+    }
+  }
+  return named;
+}
+
 /** Parse ranked or exact-name discovery arguments (string or object). */
 function parseSearchArgs(raw: unknown): { query: string; limit: number; names?: string[] } {
   let obj: Record<string, unknown> = {};
@@ -231,12 +266,25 @@ export function searchToolPool(availableTools: Tool[], rawArguments: unknown): T
       tool.type === "function" && typeof (tool as { name?: unknown }).name === "string",
   );
   if (searchable.length === 0) return [];
-  const { query, limit, names } = parseSearchArgs(rawArguments);
-  // Exact disclosure is a lookup in this already-authorized pool, not a fuzzy
-  // search or another executor registry. Keep the SDK's original references.
-  const ranked = names
-    ? searchable.filter((tool) => names.includes(tool.name))
-    : bm25RankTools(searchable, query, searchable.length);
+  const { query, limit: requestedLimit, names } = parseSearchArgs(rawArguments);
+  let ranked: Tool[];
+  let limit = requestedLimit;
+  if (names) {
+    // Exact disclosure is a lookup in this already-authorized pool, not a fuzzy
+    // search or another executor registry. Keep the SDK's original references.
+    ranked = searchable.filter((tool) => names.includes(tool.name));
+  } else {
+    // Every tool the query names comes first, even beyond the requested limit;
+    // BM25 fills any remaining places.
+    const named = toolsNamedInQuery(searchable, query);
+    ranked = [
+      ...named,
+      ...bm25RankTools(searchable, query, searchable.length).filter(
+        (tool) => !named.includes(tool as Tool & { name: string }),
+      ),
+    ];
+    limit = Math.min(MAX_SEARCH_LIMIT, Math.max(requestedLimit, named.length));
+  }
   const bounded: Tool[] = [];
   let disclosedBytes = 0;
   for (const tool of ranked) {

package/src/context-compaction.ts

@@ -43,6 +43,10 @@ export type CompactionItem = Record<string, unknown>;
 export const COMPACTION_SUMMARY_MARKER = "opengeni_context_summary";
 
 export const SUMMARY_BUFFER_TOKENS = 20_000;
+/** Leave input room on models whose entire context is smaller than the normal reserve. */
+export function compactionSummaryOutputTokens(contextWindowTokens: number): number {
+  return Math.min(SUMMARY_BUFFER_TOKENS, Math.max(1, Math.floor(contextWindowTokens / 4)));
+}
 // A single cumulative budget for all retained real user messages, matching
 // Codex core's build_compacted_history_with_limit (not a per-message allowance).
 export const COMPACT_USER_MESSAGE_MAX_TOKENS = 20_000;
@@ -1243,7 +1247,7 @@ export class EmptyCompactionSummaryError extends Error {
   constructor(diagnostics: Record<string, unknown> = {}) {
     const compact = JSON.stringify(diagnostics).slice(0, 2_000);
     super(
-      `Compaction summarizer returned no assistant text; active history was preserved${compact ? ` (${compact})` : ""}`,
+      `Compaction could not produce a usable checkpoint; active history was preserved${compact ? ` (${compact})` : ""}`,
     );
     this.name = "EmptyCompactionSummaryError";
     this.diagnostics = diagnostics;
@@ -1520,10 +1524,18 @@ export function omitOpaqueArtifactsFromPortableCompactionHistory(
     }
     if (type === "reasoning" && hasOpaqueProviderArtifact(item)) {
       const projected = projectRejectedReasoningArtifact(item);
+      if (Object.keys(projected).length === 1) {
+        changed = true;
+        continue;
+      }
       changed ||= projected !== item;
       out.push(projected as CompactionItem);
       continue;
     }
+    if (type === "reasoning" && Object.keys(item).length === 1) {
+      changed = true;
+      continue;
+    }
     out.push(item);
   }
   return changed ? out : (items as CompactionItem[]);
@@ -1605,16 +1617,20 @@ function oldestLogicalUnitCuts(items: readonly CompactionItem[]): number[] {
 
 /**
  * Build the active history after compaction:
- * the newest real user messages that fit one cumulative 20k-token budget
+ * the newest real user messages that fit one cumulative, model-bounded budget
  * (prior summaries excluded, retained images preserved) plus one marked summary item.
  */
 export function buildCompactionReplacementHistory(
   items: readonly CompactionItem[],
   summaryBody: string,
   retainedItemTokens: (item: CompactionItem) => number = (item) => estimateTokens([item]),
+  retainedMessageBudgetTokens = COMPACT_USER_MESSAGE_MAX_TOKENS,
 ): CompactionItem[] {
   const retainedReversed: CompactionItem[] = [];
-  let remaining = COMPACT_USER_MESSAGE_MAX_TOKENS;
+  let remaining = Math.max(
+    0,
+    Math.min(COMPACT_USER_MESSAGE_MAX_TOKENS, retainedMessageBudgetTokens),
+  );
   for (let index = items.length - 1; index >= 0 && remaining > 0; index -= 1) {
     const item = items[index]!;
     if (!isUserMessage(item) || isCompactionSummary(item) || isAttachmentCatalog(item)) {