@foldrun/core 11.1.0 → 11.2.1

package/dist/src/evals.d.ts

@@ -164,6 +164,15 @@ export interface EvalResult {
 /** How long one judge may take before it is cut off: a verdict on one
  *  sentence is seconds; a stalled model must not hold a grade pass. */
 export declare const JUDGE_TIMEOUT_MS = 120000;
+/**
+ * The environment a judge's model process starts with: the host's PATH, HOME
+ * and temp folder under the model's own variables. An account's provider env
+ * is only its URL, key and models, and a CLI started with no HOME dies on
+ * uv_os_homedir in a pod whose user has no passwd entry (FOL-107). Never the
+ * host's secrets: hostSafeEnv is an allowlist, as for drafts and the
+ * assistant's tool loop.
+ */
+export declare function judgeSdkEnv(env: Record<string, string | undefined>): Record<string, string | undefined>;
 /** One judge call: its words and what it cost (the SDK's result line). */
 export declare function judgeCall(rubric: string, output: string, model: string, effort: Effort | null, env: Record<string, string | undefined>, 
 /** What the output was an answer to — a live run's task, so the judge

package/dist/src/evals.js

@@ -31,7 +31,7 @@
 // they pass — no point paying a model to grade output already known to be
 // wrong.
 import { accountModelEnv, withModelEnv, ModelKeyError } from "./model-env.js";
-import { sdkEnv } from "./step-exec.js";
+import { sdkEnv, modelProcessError } from "./step-exec.js";
 import { platformModelEnv } from "./runner.js";
 import fs from "node:fs";
 import crypto from "node:crypto";
@@ -467,6 +467,17 @@ function shell(command, cwd) {
 /** How long one judge may take before it is cut off: a verdict on one
  *  sentence is seconds; a stalled model must not hold a grade pass. */
 export const JUDGE_TIMEOUT_MS = 120_000;
+/**
+ * The environment a judge's model process starts with: the host's PATH, HOME
+ * and temp folder under the model's own variables. An account's provider env
+ * is only its URL, key and models, and a CLI started with no HOME dies on
+ * uv_os_homedir in a pod whose user has no passwd entry (FOL-107). Never the
+ * host's secrets: hostSafeEnv is an allowlist, as for drafts and the
+ * assistant's tool loop.
+ */
+export function judgeSdkEnv(env) {
+    return { ...hostSafeEnv(), ...env };
+}
 /** One judge call: its words and what it cost (the SDK's result line). */
 export async function judgeCall(rubric, output, model, effort, env, 
 /** What the output was an answer to — a live run's task, so the judge
@@ -491,7 +502,7 @@ context) {
                 allowedTools: [],
                 settingSources: [],
                 strictMcpConfig: true,
-                env,
+                env: judgeSdkEnv(env),
                 maxTurns: 1,
                 abortController: abort,
             },
@@ -511,7 +522,7 @@ context) {
     catch (err) {
         if (abort.signal.aborted)
             throw new Error(`the judge took longer than ${JUDGE_TIMEOUT_MS / 1000}s and was stopped`);
-        throw err;
+        throw modelProcessError(err, "judge");
     }
     finally {
         clearTimeout(timer);

package/dist/src/step-exec.d.ts

@@ -332,6 +332,14 @@ export declare function judgePrompt(rubric: string, reply: string, work: string)
  * "the only tools I have are the Claude Docs ones". `strictMcpConfig` keeps
  * the rest of the on-disk MCP configuration out; this keeps the account's.
  */
+/**
+ * A model process's failure as a person should read it: one short line. The
+ * SDK's error for a CLI that died carries the CLI's stderr, which can be KBs
+ * of its minified bundle — internal prompt text included (FOL-107) — so that
+ * part goes to the server log, never to an eval result or a trace. An error
+ * that is already short (an API error the model answered) is kept as it is.
+ */
+export declare function modelProcessError(err: unknown, where: string): Error;
 export declare function sdkEnv(base: Record<string, string | undefined> | undefined, extra?: Record<string, string>): Record<string, string | undefined>;
 /** The text of a tool result, whichever shape it came in: a string, or a
  *  list of content blocks of which only the text ones say anything. */

package/dist/src/step-exec.js

@@ -1101,7 +1101,7 @@ async function judgeReply(rubric, reply, work, env, signal, model = "haiku", eff
         }
     }
     catch (err) {
-        return `FAIL — the judge could not run: ${err instanceof Error ? err.message : String(err)}`;
+        return `FAIL — the judge could not run: ${modelProcessError(err, "verify judge").message}`;
     }
     return texts.join("\n").trim();
 }
@@ -1167,6 +1167,22 @@ function runVerify(agentDir, command, env, data, signal) {
  * "the only tools I have are the Claude Docs ones". `strictMcpConfig` keeps
  * the rest of the on-disk MCP configuration out; this keeps the account's.
  */
+/**
+ * A model process's failure as a person should read it: one short line. The
+ * SDK's error for a CLI that died carries the CLI's stderr, which can be KBs
+ * of its minified bundle — internal prompt text included (FOL-107) — so that
+ * part goes to the server log, never to an eval result or a trace. An error
+ * that is already short (an API error the model answered) is kept as it is.
+ */
+export function modelProcessError(err, where) {
+    const full = err instanceof Error ? err.message : String(err);
+    const at = full.indexOf(" stderr:");
+    if (at < 0 && full.length <= 500)
+        return err instanceof Error ? err : new Error(full);
+    console.error(`[foldrun] ${where}: ${full}`);
+    const head = (at >= 0 ? full.slice(0, at) : full.slice(0, 300)).trim().replace(/[.:]$/, "");
+    return new Error(`${head} — the model process failed; the details are in the server log`);
+}
 export function sdkEnv(base, extra = {}) {
     return { ...(base ?? process.env), ...extra, ENABLE_CLAUDEAI_MCP_SERVERS: "false" };
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@foldrun/core",
-  "version": "11.1.0",
+  "version": "11.2.1",
   "description": "Agents are just folders. The format, the runner, the checks.",
   "license": "FSL-1.1-ALv2",
   "type": "module",

package/src/evals.ts

@@ -33,7 +33,7 @@
 
 import { accountModelEnv, withModelEnv, ModelKeyError } from "./model-env.ts";
 import type { TranslatorSpec } from "./translator.ts";
-import { sdkEnv } from "./step-exec.ts";
+import { sdkEnv, modelProcessError } from "./step-exec.ts";
 import { platformModelEnv } from "./runner.ts";
 import fs from "node:fs";
 import crypto from "node:crypto";
@@ -621,6 +621,18 @@ function shell(command: string, cwd: string): Promise<{ code: number | null; out
  *  sentence is seconds; a stalled model must not hold a grade pass. */
 export const JUDGE_TIMEOUT_MS = 120_000;
 
+/**
+ * The environment a judge's model process starts with: the host's PATH, HOME
+ * and temp folder under the model's own variables. An account's provider env
+ * is only its URL, key and models, and a CLI started with no HOME dies on
+ * uv_os_homedir in a pod whose user has no passwd entry (FOL-107). Never the
+ * host's secrets: hostSafeEnv is an allowlist, as for drafts and the
+ * assistant's tool loop.
+ */
+export function judgeSdkEnv(env: Record<string, string | undefined>): Record<string, string | undefined> {
+  return { ...hostSafeEnv(), ...env };
+}
+
 /** One judge call: its words and what it cost (the SDK's result line). */
 export async function judgeCall(
   rubric: string,
@@ -653,7 +665,7 @@ export async function judgeCall(
         allowedTools: [],
         settingSources: [],
         strictMcpConfig: true,
-        env,
+        env: judgeSdkEnv(env),
         maxTurns: 1,
         abortController: abort,
       },
@@ -669,7 +681,7 @@ export async function judgeCall(
     }
   } catch (err) {
     if (abort.signal.aborted) throw new Error(`the judge took longer than ${JUDGE_TIMEOUT_MS / 1000}s and was stopped`);
-    throw err;
+    throw modelProcessError(err, "judge");
   } finally {
     clearTimeout(timer);
   }

package/src/step-exec.ts

@@ -1309,7 +1309,7 @@ async function judgeReply(
       }
     }
   } catch (err) {
-    return `FAIL — the judge could not run: ${err instanceof Error ? err.message : String(err)}`;
+    return `FAIL — the judge could not run: ${modelProcessError(err, "verify judge").message}`;
   }
   return texts.join("\n").trim();
 }
@@ -1377,6 +1377,22 @@ function runVerify(
  * "the only tools I have are the Claude Docs ones". `strictMcpConfig` keeps
  * the rest of the on-disk MCP configuration out; this keeps the account's.
  */
+/**
+ * A model process's failure as a person should read it: one short line. The
+ * SDK's error for a CLI that died carries the CLI's stderr, which can be KBs
+ * of its minified bundle — internal prompt text included (FOL-107) — so that
+ * part goes to the server log, never to an eval result or a trace. An error
+ * that is already short (an API error the model answered) is kept as it is.
+ */
+export function modelProcessError(err: unknown, where: string): Error {
+  const full = err instanceof Error ? err.message : String(err);
+  const at = full.indexOf(" stderr:");
+  if (at < 0 && full.length <= 500) return err instanceof Error ? err : new Error(full);
+  console.error(`[foldrun] ${where}: ${full}`);
+  const head = (at >= 0 ? full.slice(0, at) : full.slice(0, 300)).trim().replace(/[.:]$/, "");
+  return new Error(`${head} — the model process failed; the details are in the server log`);
+}
+
 export function sdkEnv(base: Record<string, string | undefined> | undefined, extra: Record<string, string> = {}): Record<string, string | undefined> {
   return { ...(base ?? process.env), ...extra, ENABLE_CLAUDEAI_MCP_SERVERS: "false" };
 }