opencrowd 0.10.0 → 0.11.0

package/bundle/opencrowd.js

@@ -16717,12 +16717,19 @@ var init_session_defaults = __esm({
 // ../../packages/core/dist/shell.js
 import { spawn } from "node:child_process";
 import { isAbsolute as isAbsolute2, join as join7, resolve as resolve3 } from "node:path";
+function maxShellTimeoutMs(env = process.env) {
+  const requested = Number(env.OPENCROWD_MAX_SHELL_TIMEOUT_MS);
+  if (!Number.isInteger(requested) || requested <= DEFAULT_MAX_SHELL_TIMEOUT_MS) {
+    return DEFAULT_MAX_SHELL_TIMEOUT_MS;
+  }
+  return Math.min(SHELL_TIMEOUT_CEILING_MS, requested);
+}
 async function runShell(state, command, cwd = state.workspaceRoot, timeoutMs = 1e4, options = {}) {
   if (!state.shellEnabled) {
     throw new Error("shell tool is disabled for this session");
   }
   options.signal?.throwIfAborted();
-  const maxTimeout = options.maxTimeoutMs ?? 3e5;
+  const maxTimeout = options.maxTimeoutMs ?? maxShellTimeoutMs();
   if (!Number.isInteger(timeoutMs) || timeoutMs < 1 || timeoutMs > maxTimeout) {
     throw new Error(`timeout_ms must be between 1 and ${maxTimeout}`);
   }
@@ -16853,10 +16860,13 @@ function capOutput(output3, maxBytes) {
   }
   return output3.slice(0, maxBytes) + "\n[output truncated]\n";
 }
+var DEFAULT_MAX_SHELL_TIMEOUT_MS, SHELL_TIMEOUT_CEILING_MS;
 var init_shell = __esm({
   "../../packages/core/dist/shell.js"() {
     "use strict";
     init_ledger();
+    DEFAULT_MAX_SHELL_TIMEOUT_MS = 3e5;
+    SHELL_TIMEOUT_CEILING_MS = 27e5;
   }
 });
 
@@ -16907,7 +16917,7 @@ function toolParameters(name) {
       return objectSchema({
         command: stringSchema("Shell command."),
         cwd: stringSchema("Working directory inside the workspace. Use artifacts/<dir> to run where save_file writes (cd inside the command does not reach it)."),
-        timeout_ms: integerSchema("Timeout in milliseconds (default 10000, maximum 300000).")
+        timeout_ms: integerSchema(`Timeout in milliseconds (default 10000, maximum ${maxShellTimeoutMs()}). Pass the real time a long build or download needs; the command is killed at the timeout.`)
       }, ["command"]);
     case "spawn_subagent":
       return objectSchema({
@@ -17039,6 +17049,7 @@ var TOOL_NAMES, HOSTED_ONLY_TOOL_NAMES, SUBAGENT_TOOL_NAMES, OPEN_CROWD_TOOLS;
 var init_tool_definitions = __esm({
   "../../packages/core/dist/tool-definitions.js"() {
     "use strict";
+    init_shell();
     TOOL_NAMES = [
       "get_budget_status",
       "save_file",
@@ -17167,7 +17178,7 @@ function shellTimeoutMs(value) {
   if (typeof value !== "number" || !Number.isInteger(value)) {
     return DEFAULT_SHELL_TIMEOUT_MS;
   }
-  return Math.min(MAX_SHELL_TIMEOUT_MS, Math.max(1, value));
+  return Math.min(maxShellTimeoutMs(), Math.max(1, value));
 }
 function objectValue(value) {
   if (value === void 0 || value === null) {
@@ -17184,7 +17195,7 @@ function sumCents(rows) {
     return total + (Number.isFinite(value) ? Math.round(value) : 0);
   }, 0);
 }
-var DEFAULT_SHELL_TIMEOUT_MS, MAX_SHELL_TIMEOUT_MS;
+var DEFAULT_SHELL_TIMEOUT_MS;
 var init_tools = __esm({
   "../../packages/core/dist/tools.js"() {
     "use strict";
@@ -17193,7 +17204,6 @@ var init_tools = __esm({
     init_ledger();
     init_shell();
     DEFAULT_SHELL_TIMEOUT_MS = 1e4;
-    MAX_SHELL_TIMEOUT_MS = 3e5;
   }
 });
 
@@ -80894,7 +80904,7 @@ function economyTools(economy, options = {}) {
       const status = await economy.execute("crowdcode_status", {});
       if (!status.ok || !status.data?.enabled)
         return review || void 0;
-      return [review, "Before your final answer, reflect once: What concrete services would have been worth paying for to solve this task? For each observed failure, poor result, excessive cost or avoidable detour with a reusable paid remedy, call request_service. Specify exact inputs, the paid deliverable, acceptance criteria, the actual use case and why paying would help. A prior purchase or spending permission is not required. Do not invent budgets. Skip generic local computation, ordinary agent mistakes without a sellable remedy, and services such as web search that already worked well. If everything worked well and cheaply, submit nothing. Then finish the user's task."].filter(Boolean).join(" ");
+      return [review, "Before this run ends, reflect once: What concrete services would have been worth paying for to solve this task? For each observed failure, poor result, excessive cost or avoidable detour with a reusable paid remedy, call request_service. Specify exact inputs, the paid deliverable, acceptance criteria, the actual use case and why paying would help. A prior purchase or spending permission is not required. Do not invent budgets. Skip generic local computation, ordinary agent mistakes without a sellable remedy, and services such as web search that already worked well. If everything worked well and cheaply, submit nothing. Then call complete_session again to finish."].filter(Boolean).join(" ");
     }
   };
 }
@@ -82475,6 +82485,8 @@ async function runAgentTaskDetailed(session, task, options = {}) {
   ];
   if (options.hosted)
     systemPromptParts.push("You operate this service for its owner. Inference and paid tools debit its USDC wallet. When authorized to build or sell a service, take concrete steps within the owner's configured scope and budget. Ask only for missing information or permissions that materially block progress. Report actual costs and outcomes. Financial sustainability is the owner's business objective; respect their stop controls.", "You are a continuous autonomous agent, not a turn-based assistant. Maintain unfinished goals, promised follow-ups, recurring work, and conditions worth monitoring. When future work would clearly help, proactively create a persistent automation instead of waiting for the owner to say schedule. Use a one-time automation for a future reminder or follow-up and a cron automation for monitoring, recurring reports, or maintenance. It runs in this same conversation. Never create duplicates or noisy schedules, and always tell the owner exactly what you created, its timezone, and next run.", "Before finishing a substantive task, consider whether a follow-up, reminder, monitoring loop, recurring report, or maintenance check would materially advance the owner's goal; create one when the need is clear, but do not schedule ordinary one-off work.", "Financial state: the current hosted financial snapshot is authoritative; historical local session budget figures are not wallet balances and may omit inference spending. Wallet balance, pending payment holds, settled inference spend, settled service spend, and remaining run budget are separate. When the snapshot lists wallet networks, each network's balance is separate: a payment on a network needs that network's available balance. Report the wallet to the owner as one total, mentioning the per-network split only when it matters. Spending totals cover only the identified current run. A budget limit does not supply funds. An unavailable balance is unknown, not zero. The snapshot before a model call excludes that call's eventual charge; never claim a response was free or that you spent nothing merely because no paid tool ran. Use get_budget_status or get_wallet_status for refreshed financial state. Respect approval mode and per-call limits; over-limit requests require the owner's approval.");
+  if (options.hosted)
+    systemPromptParts.push("Files you produce: every file in this session's workspace, whether written with save_file or by a shell command, is published to the owner as a downloadable, previewable file. To point at one in your final message, write its exact workspace-relative path as plain text, for example generated/report.pdf \u2014 the owner's interface turns that path into a working download and preview link. Never invent a download URL, a sandbox: link, a /mnt/data path, or any other path the owner cannot open, and never claim a file exists before you have verified it.");
   if (options.dynamicTools) {
     systemPromptParts.push("When the local computer is not the right environment, or after one clear local capability failure, buy external capability: find_paid_service to discover, inspect_paid_service to see the exact schema/price/reputation, call_paid_service to execute through the enforced purchase lifecycle, and review_paid_service for the required review after every confirmed paid call (success or failure). The same review tool accepts unpaid experiences; review relevant free results or failed attempts without inventing payment evidence. Distinguish provider faults from local errors and uncertain failures.", `Choosing a service: find_paid_service with a short capability query (for example "web search" or "object storage", not the task's subject) searches every catalog (AgentCash, Coinbase's x402 catalog, the MPP catalog and CrowdCode's reviews) across Base, Solana and Tempo and returns services you can pay, ranked best first by CrowdCode: quality (score and verified reviews), then effective price on the service's best network, including about $0.03 to move USDC between your own networks when that network's balance is short, then your balance there. Prefer the highest-ranked service that fits the task, whatever its network; a better or cheaper service on Solana or Tempo beats a Base default. When the chosen service is marked transfer_needed, move USDC between your own networks first if you have a tool for that; otherwise tell the user which network needs funds. When the ranking is unavailable, compare candidates yourself with inspect_paid_service (reputation, price and network) against your per-network balances.`, ...options.hosted ? [
       "You can use new services without prior CrowdCode reviews. Payable: x402 USDC on Base or Solana and MPP USDC.e on Tempo; the hosted gateway pays the cheapest offered network your balance on that network covers. Other networks, including Monad, are not payable here. Payments stay within the user's configured limits; over-limit quotes request approval. Use read_service for authenticated job polling and never resubmit a pending generation job.",
@@ -82494,7 +82506,7 @@ async function runAgentTaskDetailed(session, task, options = {}) {
     systemPromptParts.push("Using secrets from chat: call vault_request to send one HTTPS request with a vault secret injected, writing {{SECRET_NAME}} where the value belongs (for example an authorization header). The platform substitutes the real value only for hosts the user approved for that secret, refreshes OAuth tokens automatically, and never returns the value to you. If it reports that a host needs approval, tell the user to allow it in the Vault tab and stop; never look for paid third-party services to do what an approved secret already lets you do. To upload a file (an image or video you rendered with run_shell, where ffmpeg is installed), pass body_artifact with its path under artifacts/ and, for multipart APIs, body_field plus form_fields; files over an API's chunk limit can be split with run_shell first.");
   }
   if (options.hosted && hostedTools.includes("run_shell")) {
-    systemPromptParts.push("Tooling: ffmpeg, python3 with pip and venv, node and npm are installed; the system image is read-only, but `pip install --user <package>` and `npm install` persist on this machine's volume between runs, so install once and reuse.", "Resource recovery: before installing large ML dependencies, check available memory, disk space and Python version. A killed install or OOM is a failed approach, not the end of the task. Inspect partial installs and preserve user files, then choose a materially lighter backend or an external service within the user's budget. Verify package metadata: optional extras add dependencies and do not remove default ones; never invent an extra such as basic-pitch[tflite]. For audio-to-MIDI, prefer a verified TFLite or ONNX inference setup in an isolated venv over full TensorFlow; test on a short audio segment before processing the whole file. If no viable path remains, report the concrete blocker and useful artifacts. Do not claim success until the output has been checked.");
+    systemPromptParts.push("Tooling: ffmpeg, python3 with pip and venv, node and npm are installed; the system image is read-only, but `pip install --user <package>` and `npm install` persist on this machine's volume between runs, so install once and reuse.", "Heavy toolchains: this machine's CPU, memory and disk come from the agent's compute tier, and a multi-gigabyte toolchain needs the heavy tier. Measure before installing one (`df -h \"$HOME\"`, `free -m`, `nproc`) and compare the numbers with what it needs. Install it under `$HOME`: that is the persistent volume and is excluded from the per-run workspace backup, while the session workspace is archived after every run with a 25 MB cap. Give a long download or compile the time it really needs through run_shell `timeout_ms` instead of accepting the 10 s default, redirect build logs to a file under `$HOME` and tail it rather than printing everything, and re-check free space between steps. `No space left on device` or an OOM kill twice means the tier is too small, not that the command was wrong: stop retrying, and tell the owner exactly which limit you hit, how much you need, and that they can switch this agent to the heavy compute tier from Machine in the agent menu (which also raises the shell timeout).", 'Lean 4 and Mathlib, when a task asks for machine-checked proofs: install elan into `$HOME` (`ELAN_HOME="$HOME/.elan" curl -fsSL https://elan.lean-lang.org/elan-init.sh | sh -s -- -y --default-toolchain none`, then source `$HOME/.elan/env`), create or clone the project under `$HOME` so `lake-manifest.json` pins the toolchain, and always run `lake exe cache get` to fetch prebuilt `.olean` files before `lake build` \u2014 compiling Mathlib from source is not viable here. Expect roughly 2 GB for the toolchain, 10 GB for the Mathlib cache, several minutes for `cache get`, and a few GB of memory to elaborate one file that imports Mathlib. Report the compiler\'s real output; a file containing `sorry` is not a proof.', "Resource recovery: before installing large ML dependencies, check available memory, disk space and Python version. A killed install or OOM is a failed approach, not the end of the task. Inspect partial installs and preserve user files, then choose a materially lighter backend or an external service within the user's budget. Verify package metadata: optional extras add dependencies and do not remove default ones; never invent an extra such as basic-pitch[tflite]. For audio-to-MIDI, prefer a verified TFLite or ONNX inference setup in an isolated venv over full TensorFlow; test on a short audio segment before processing the whole file. If no viable path remains, report the concrete blocker and useful artifacts. Do not claim success until the output has been checked.");
   }
   systemPromptParts.push("End by calling complete_session with a concise final message.");
   if (options.subagent) {
@@ -82523,6 +82535,8 @@ async function runAgentTaskDetailed(session, task, options = {}) {
   const repeatedFailures = new Map(options.resume?.repeatedFailures);
   let completionNudges = options.resume?.completionNudges ?? 0;
   let pendingFinalMessage = options.resume?.pendingFinalMessage;
+  let finalAnswerHeld = options.resume?.finalAnswerHeld ?? false;
+  let supersededFinalMessage;
   let outputContinuationNudges = options.resume?.outputContinuationNudges ?? 0;
   let outputContinuationPrefix = options.resume?.outputContinuationPrefix ?? "";
   let serviceCallFailures = options.resume?.serviceCallFailures ?? 0;
@@ -82537,6 +82551,7 @@ async function runAgentTaskDetailed(session, task, options = {}) {
       deliveredMessageIds: [...deliveredMessageIds],
       completionNudges,
       pendingFinalMessage,
+      finalAnswerHeld,
       outputContinuationNudges,
       outputContinuationPrefix,
       serviceCallFailures,
@@ -82552,6 +82567,8 @@ async function runAgentTaskDetailed(session, task, options = {}) {
     }
     if (fresh.length) {
       pendingFinalMessage = void 0;
+      finalAnswerHeld = false;
+      supersededFinalMessage = void 0;
       outputContinuationPrefix = "";
       await checkpoint(turn);
     }
@@ -82561,6 +82578,16 @@ async function runAgentTaskDetailed(session, task, options = {}) {
       await options.inbox?.delivered(pending.map((m) => m.id));
     return fresh.length > 0;
   };
+  const recordFinalMessage = (text) => {
+    if (!text?.trim())
+      return;
+    if (!finalAnswerHeld)
+      pendingFinalMessage = text;
+    else if (text.trim() !== pendingFinalMessage?.trim())
+      supersededFinalMessage = text;
+  };
+  const gateNudge = (blocker) => finalAnswerHeld ? `Your final answer is already saved and will be delivered to the user exactly as written; nothing you write now can replace it. Before this run ends: ${blocker} Then call complete_session again to finish; a one-line acknowledgement is enough, and do not restate the answer.` : `You cannot finish yet: ${blocker}`;
+  const withReflection = (summary2) => supersededFinalMessage ? { ...summary2, reflection_note: supersededFinalMessage } : summary2;
   let subagentCount = 0;
   const backgroundSubagents = /* @__PURE__ */ new Map();
   const subagentLimiter = createLimiter(options.subagent?.maxParallel ?? 4);
@@ -82642,8 +82669,7 @@ ${outputContinuationPrefix}`);
       if (await injectMessages(turn + 1))
         continue;
       const finalMessage = response.content.trim() ? outputContinuationPrefix + response.content : pendingFinalMessage ?? outputContinuationPrefix;
-      if (finalMessage.trim())
-        pendingFinalMessage = finalMessage;
+      recordFinalMessage(finalMessage);
       const blocker = await options.completionGate?.(completionNudges > 0);
       if (blocker) {
         if (completionNudges >= 1) {
@@ -82651,7 +82677,9 @@ ${outputContinuationPrefix}`);
           return { outcome: "stopped", summary: summary3, turns: turn + 1 };
         }
         completionNudges += 1;
-        const nudge = { role: "user", content: `You cannot finish yet: ${blocker}` };
+        if (pendingFinalMessage?.trim())
+          finalAnswerHeld = true;
+        const nudge = { role: "user", content: gateNudge(blocker) };
         messages.push(nudge);
         await options.onMessage?.(nudge);
         await checkpoint(turn + 1);
@@ -82659,7 +82687,7 @@ ${outputContinuationPrefix}`);
       }
       if (!pendingFinalMessage?.trim())
         throw new Error("The model ended without an answer.");
-      const summary2 = await completeSession(session, pendingFinalMessage);
+      const summary2 = withReflection(await completeSession(session, pendingFinalMessage));
       return { outcome: "completed", summary: summary2, turns: turn + 1 };
     }
     const spawnedThisTurn = /* @__PURE__ */ new Map();
@@ -82795,8 +82823,8 @@ ${outputContinuationPrefix}`);
       if (!result.ok)
         continue;
       const finalMessage = result.data && typeof result.data === "object" ? result.data.final_message : void 0;
-      if (typeof finalMessage === "string" && finalMessage.trim())
-        pendingFinalMessage = finalMessage;
+      if (typeof finalMessage === "string")
+        recordFinalMessage(finalMessage);
       const blocker = await options.completionGate?.(completionNudges > 0);
       if (blocker) {
         if (completionNudges >= 1) {
@@ -82804,9 +82832,11 @@ ${outputContinuationPrefix}`);
           return { outcome: "stopped", summary: summary3, turns: turn + 1 };
         }
         completionNudges += 1;
+        if (pendingFinalMessage?.trim())
+          finalAnswerHeld = true;
         const gateMessage = {
           role: "user",
-          content: `You cannot finish yet: ${blocker}`
+          content: gateNudge(blocker)
         };
         messages.push(gateMessage);
         await options.onMessage?.(gateMessage);
@@ -82817,13 +82847,13 @@ ${outputContinuationPrefix}`);
         throw new Error("The model ended without an answer.");
       const backgroundResults = backgroundSubagents.size > 0 ? await drainBackground() : void 0;
       if (result.data && typeof result.data === "object") {
-        const summary3 = { ...result.data, final_message: pendingFinalMessage };
+        const summary3 = withReflection({ ...result.data, final_message: pendingFinalMessage });
         if (backgroundResults) {
           summary3.background_subagents = backgroundResults;
         }
         return { outcome: "completed", summary: summary3, turns: turn + 1 };
       }
-      const summary2 = await completeSession(session, pendingFinalMessage);
+      const summary2 = withReflection(await completeSession(session, pendingFinalMessage));
       if (backgroundResults) {
         summary2.background_subagents = backgroundResults;
       }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencrowd",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "description": "A CLI agent with its own wallet — discovers x402 services, pays in USDC, and completes tasks under a local budget",
   "type": "module",
   "license": "MIT",