@prestyj/agent 5.12.0 → 5.16.0

package/dist/index.cjs

@@ -23,6 +23,8 @@ __export(index_exports, {
   Agent: () => Agent,
   AgentStream: () => AgentStream,
   agentLoop: () => agentLoop,
+  cancelledBeforeStartText: () => cancelledBeforeStartText,
+  indeterminateOutcomeText: () => indeterminateOutcomeText,
   isAbortError: () => isAbortError,
   isBillingError: () => isBillingError,
   isContextOverflow: () => isContextOverflow,
@@ -61,6 +63,60 @@ function isLocalBackendUrl(baseUrl) {
   return false;
 }
 
+// src/output-ceiling.ts
+var TTL_MS = 24 * 60 * 60 * 1e3;
+var MIN_PLAUSIBLE_CEILING = 256;
+var MAX_PLAUSIBLE_CEILING = 1e7;
+var ceilings = /* @__PURE__ */ new Map();
+function outputRouteKey(route) {
+  return `${route.provider}\0${route.baseUrl ?? "default"}\0${route.model}`;
+}
+function parseOutputTokenCeiling(err) {
+  if (!(err instanceof Error)) return null;
+  const msg = err.message;
+  if (!/max_tokens|max output tokens|output tokens|completion tokens/i.test(msg)) return null;
+  if (err.statusCode === 402) return null;
+  if (/credit|billing|payment|insufficient|balance|quota/i.test(msg)) return null;
+  const patterns = [
+    // Anthropic: "max_tokens: 100000 > 64000, which is the maximum allowed…"
+    /max_tokens:\s*\d+\s*>\s*(\d+)/i,
+    // OpenAI: "…this model supports at most 16384 completion tokens"
+    /(?:at most|maximum of|limit of)\s+([\d,_]+)\s*(?:output|completion)?\s*tokens/i,
+    // Generic: "max output tokens is 8192" / "max_tokens must be <= 4096"
+    /(?:max_tokens|max output tokens)\b[^\d]{0,30}?([\d,_]+)/i
+  ];
+  for (const pattern of patterns) {
+    const raw = msg.match(pattern)?.[1];
+    if (raw === void 0) continue;
+    const limit = Number(raw.replace(/[,_]/g, ""));
+    if (Number.isFinite(limit) && limit >= MIN_PLAUSIBLE_CEILING && limit <= MAX_PLAUSIBLE_CEILING) {
+      return limit;
+    }
+  }
+  return null;
+}
+function rememberOutputCeiling(key, limit) {
+  const known = outputTokenCeiling(key);
+  ceilings.set(key, {
+    limit: known === void 0 ? limit : Math.min(known, limit),
+    expiresAt: Date.now() + TTL_MS
+  });
+}
+function outputTokenCeiling(key) {
+  const entry = ceilings.get(key);
+  if (!entry) return void 0;
+  if (entry.expiresAt <= Date.now()) {
+    ceilings.delete(key);
+    return void 0;
+  }
+  return entry.limit;
+}
+function clampOutputTokens(key, requested) {
+  const ceiling = outputTokenCeiling(key);
+  if (ceiling === void 0) return requested;
+  return requested === void 0 ? ceiling : Math.min(requested, ceiling);
+}
+
 // src/agent-loop.ts
 var DEFAULT_MAX_TURNS = 300;
 var DEFAULT_TOOL_TIMEOUT_MS = 3e5;
@@ -345,6 +401,16 @@ async function* agentLoop(messages, options) {
   const MAX_OUTPUT_CONTINUATIONS = 2;
   const MAX_TOKENS_CONTINUATION_PROMPT = "[Your previous response hit the output-token limit and was cut off. The text above is what was already delivered to the user. Continue exactly from where it stopped \u2014 do not repeat or restart it.]";
   let maxTokensContinuations = 0;
+  let providerCalls = 0;
+  let nonStreamingCalls = 0;
+  let warnedNonStreaming = false;
+  const MAX_OUTPUT_CEILING_RETRIES = 1;
+  let outputCeilingRetries = 0;
+  const ceilingKey = outputRouteKey({
+    provider: options.provider,
+    model: options.model,
+    baseUrl: options.baseUrl
+  });
   const OVERLOAD_BASE_DELAY_MS = 2e3;
   const OVERLOAD_MAX_DELAY_MS = 3e4;
   const STREAM_FIRST_EVENT_TIMEOUT_MS = 45e3;
@@ -474,6 +540,18 @@ async function* agentLoop(messages, options) {
       let streamIterator = null;
       try {
         diag("stream_call", { nonStreaming: useNonStreamingFallback });
+        providerCalls++;
+        if (useNonStreamingFallback) nonStreamingCalls++;
+        if (!warnedNonStreaming && nonStreamingCalls >= 3) {
+          warnedNonStreaming = true;
+          diag("non_streaming_session", {
+            nonStreamingCalls,
+            providerCalls,
+            provider: options.provider,
+            model: options.model,
+            impact: "streaming is disabled for this session after repeated stalls: failed turns are re-billed in full instead of resuming from partial output, and replies appear only when complete"
+          });
+        }
         streamCallStart = Date.now();
         providerAttemptStartedAt = streamCallStart;
         let liveApiKey = options.apiKey;
@@ -499,7 +577,9 @@ async function* agentLoop(messages, options) {
           serverTools: options.serverTools,
           toolChoice: options.toolChoice,
           webSearch: options.webSearch,
-          maxTokens: options.maxTokens,
+          // Clamped to whatever ceiling this route has already rejected us for
+          // (identity when nothing has been learned).
+          maxTokens: clampOutputTokens(ceilingKey, options.maxTokens),
           temperature: options.temperature,
           thinking: options.thinking,
           apiKey: liveApiKey,
@@ -656,6 +736,29 @@ async function* agentLoop(messages, options) {
           });
           throw err;
         }
+        const statedCeiling = parseOutputTokenCeiling(err);
+        if (statedCeiling !== null) {
+          rememberOutputCeiling(ceilingKey, statedCeiling);
+          diag("output_ceiling_learned", {
+            ceiling: statedCeiling,
+            requested: options.maxTokens,
+            provider: options.provider,
+            model: options.model
+          });
+          if (outputCeilingRetries < MAX_OUTPUT_CEILING_RETRIES) {
+            outputCeilingRetries++;
+            yield {
+              type: "retry",
+              reason: "provider_error",
+              attempt: outputCeilingRetries,
+              maxAttempts: MAX_OUTPUT_CEILING_RETRIES,
+              delayMs: 0,
+              silent: true
+            };
+            turn--;
+            continue;
+          }
+        }
         if (isContextOverflow(err)) {
           const overflowDetails = extractContextOverflowDetails(err);
           diag("context_overflow_detected", {
@@ -951,6 +1054,7 @@ async function* agentLoop(messages, options) {
           continue;
         }
       }
+      const emptyExhausted = !hasActionableContent;
       emptyResponseRetries = 0;
       useNonStreamingFallback = false;
       totalUsage.inputTokens += response.usage.inputTokens;
@@ -964,9 +1068,11 @@ async function* agentLoop(messages, options) {
       if (response.usage.cacheWrite) {
         totalUsage.cacheWrite = (totalUsage.cacheWrite ?? 0) + response.usage.cacheWrite;
       }
-      messages.push(response.message);
-      latestProviderUsage = response.usage;
-      usageAnchorIndex = messages.length - 1;
+      if (!emptyExhausted) {
+        messages.push(response.message);
+        latestProviderUsage = response.usage;
+        usageAnchorIndex = messages.length - 1;
+      }
       const completedAt = Date.now();
       const outputTokensPerSecond = providerDurationMs > 0 && response.usage.outputTokens > 0 ? response.usage.outputTokens / (providerDurationMs / 1e3) : void 0;
       const timing = {
@@ -998,8 +1104,15 @@ async function* agentLoop(messages, options) {
       }
       consecutivePauses = 0;
       const allToolCalls = extractToolCalls(response.message.content);
-      if (response.stopReason !== "tool_use" && allToolCalls.length === 0) {
-        if (response.stopReason === "max_tokens") {
+      if (emptyExhausted || response.stopReason !== "tool_use" && allToolCalls.length === 0) {
+        if (emptyExhausted) {
+          diag("empty_response_exhausted", {
+            provider: options.provider,
+            model: options.model,
+            stopReason: response.stopReason
+          });
+          yield { type: "truncated", reason: "empty_response", continued: false };
+        } else if (response.stopReason === "max_tokens") {
           if (maxTokensContinuations < MAX_OUTPUT_CONTINUATIONS) {
             maxTokensContinuations++;
             diag("max_tokens_continuation", {
@@ -1221,6 +1334,7 @@ async function executeSingleToolCall(toolCall, options, pushEvent) {
   let resultContent;
   let details;
   let isError = false;
+  let invalidArgAttempt;
   const tool = options.toolMap.get(toolCall.name);
   if (!tool) {
     resultContent = `Unknown tool: ${toolCall.name}`;
@@ -1258,6 +1372,7 @@ async function executeSingleToolCall(toolCall, options, pushEvent) {
         const failureKey = `${toolCall.name}:${prettyError}`;
         const failureCount = (options.invalidToolArgumentCounts.get(failureKey) ?? 0) + 1;
         options.invalidToolArgumentCounts.set(failureKey, failureCount);
+        invalidArgAttempt = failureCount;
         resultContent = `Invalid arguments for tool \`${toolCall.name}\`:
 ` + prettyError + "\nRe-issue the call with each field as the correct type.";
         if (failureCount >= 3) {
@@ -1274,6 +1389,8 @@ async function executeSingleToolCall(toolCall, options, pushEvent) {
             toolCall.name
           );
         }
+      } else if (options.signal?.aborted && isAbortError(err)) {
+        resultContent = indeterminateOutcomeText(toolCall.name);
       } else {
         resultContent = (0, import_ai.redactValue)(err instanceof Error ? err.message : String(err));
       }
@@ -1288,7 +1405,8 @@ async function executeSingleToolCall(toolCall, options, pushEvent) {
     result: toolResultPreview(resultContent),
     details,
     isError,
-    durationMs
+    durationMs,
+    ...invalidArgAttempt === void 0 ? {} : { invalidArgAttempt }
   });
   return { toolCallId: toolCall.id, content: resultContent, isError };
 }
@@ -1296,6 +1414,7 @@ async function* executeToolCallsMixed(toolCalls, initialToolResults, options) {
   const eventStream = new import_ai.EventStream();
   const state = { finalized: false };
   const resultsById = /* @__PURE__ */ new Map();
+  const dispatchedIds = /* @__PURE__ */ new Set();
   const abortHandler = () => eventStream.abort(new Error("aborted"));
   options.signal?.addEventListener("abort", abortHandler, { once: true });
   const phases = [];
@@ -1320,6 +1439,7 @@ async function* executeToolCallsMixed(toolCalls, initialToolResults, options) {
       for (const phase of phases) {
         if (options.signal?.aborted) break;
         if (phase.sequential) {
+          dispatchedIds.add(phase.sequential.id);
           const record = await executeSingleToolCall(
             phase.sequential,
             options,
@@ -1327,6 +1447,7 @@ async function* executeToolCallsMixed(toolCalls, initialToolResults, options) {
           );
           resultsById.set(record.toolCallId, record);
         } else if (phase.parallel.length === 1) {
+          dispatchedIds.add(phase.parallel[0].id);
           const record = await executeSingleToolCall(
             phase.parallel[0],
             options,
@@ -1336,6 +1457,7 @@ async function* executeToolCallsMixed(toolCalls, initialToolResults, options) {
         } else {
           await Promise.all(
             phase.parallel.map(async (toolCall) => {
+              dispatchedIds.add(toolCall.id);
               const record = await executeSingleToolCall(
                 toolCall,
                 options,
@@ -1366,7 +1488,7 @@ async function* executeToolCallsMixed(toolCalls, initialToolResults, options) {
     options.signal?.removeEventListener("abort", abortHandler);
     state.finalized = true;
   }
-  const toolResults = buildToolResults(initialToolResults, toolCalls, resultsById);
+  const toolResults = buildToolResults(initialToolResults, toolCalls, resultsById, dispatchedIds);
   capToolResults(toolResults, options.maxToolResultChars);
   capTurnToolResults(toolResults, options.maxTurnToolResultChars);
   return { toolResults, aborted };
@@ -1375,10 +1497,12 @@ async function* executeToolCallsParallel(toolCalls, initialToolResults, options)
   const eventStream = new import_ai.EventStream();
   const state = { finalized: false };
   const resultsById = /* @__PURE__ */ new Map();
+  const dispatchedIds = /* @__PURE__ */ new Set();
   const abortHandler = () => eventStream.abort(new Error("aborted"));
   options.signal?.addEventListener("abort", abortHandler, { once: true });
   Promise.all(
     toolCalls.map(async (toolCall) => {
+      dispatchedIds.add(toolCall.id);
       const record = await executeSingleToolCall(
         toolCall,
         options,
@@ -1406,12 +1530,18 @@ async function* executeToolCallsParallel(toolCalls, initialToolResults, options)
     options.signal?.removeEventListener("abort", abortHandler);
     state.finalized = true;
   }
-  const toolResults = buildToolResults(initialToolResults, toolCalls, resultsById);
+  const toolResults = buildToolResults(initialToolResults, toolCalls, resultsById, dispatchedIds);
   capToolResults(toolResults, options.maxToolResultChars);
   capTurnToolResults(toolResults, options.maxTurnToolResultChars);
   return { toolResults, aborted };
 }
-function buildToolResults(initialToolResults, toolCalls, resultsById) {
+function cancelledBeforeStartText(name) {
+  return `\`${name}\` was cancelled before it started, so it had no effect. Safe to retry.`;
+}
+function indeterminateOutcomeText(name) {
+  return `\`${name}\` started running and was cut off before it reported back, so its outcome is UNKNOWN \u2014 it may have completed. Check the real state (re-read the file, re-run a status command) before retrying it, and do not tell the user it did not happen.`;
+}
+function buildToolResults(initialToolResults, toolCalls, resultsById, dispatchedIds) {
   const toolResults = [...initialToolResults];
   for (const toolCall of toolCalls) {
     const result = resultsById.get(toolCall.id);
@@ -1423,10 +1553,11 @@ function buildToolResults(initialToolResults, toolCalls, resultsById) {
         isError: result.isError || void 0
       });
     } else {
+      const dispatched = dispatchedIds?.has(toolCall.id) ?? true;
       toolResults.push({
         type: "tool_result",
         toolCallId: toolCall.id,
-        content: "Tool execution was aborted.",
+        content: dispatched ? indeterminateOutcomeText(toolCall.name) : cancelledBeforeStartText(toolCall.name),
         isError: true
       });
     }
@@ -1574,32 +1705,22 @@ function repairToolPairingAdjacent(messages) {
     const msg = messages[i];
     if (msg.role !== "assistant") continue;
     if (typeof msg.content === "string" || !Array.isArray(msg.content)) continue;
-    const toolCallIds = msg.content.filter((p) => p.type === "tool_call").map((p) => p.id);
-    if (toolCallIds.length === 0) continue;
+    const orphanCalls = msg.content.filter((p) => p.type === "tool_call").map((p) => p);
+    if (orphanCalls.length === 0) continue;
+    const repaired = (call) => ({
+      type: "tool_result",
+      toolCallId: call.id,
+      content: indeterminateOutcomeText(call.name),
+      isError: true
+    });
     const next = messages[i + 1];
     if (next?.role === "tool" && Array.isArray(next.content)) {
       const existingIds = new Set(next.content.map((r) => r.toolCallId));
-      const missing = toolCallIds.filter((id) => !existingIds.has(id));
-      if (missing.length > 0) {
-        for (const id of missing) {
-          next.content.push({
-            type: "tool_result",
-            toolCallId: id,
-            content: "Tool execution was interrupted.",
-            isError: true
-          });
-        }
+      for (const call of orphanCalls) {
+        if (!existingIds.has(call.id)) next.content.push(repaired(call));
       }
     } else {
-      messages.splice(i + 1, 0, {
-        role: "tool",
-        content: toolCallIds.map((id) => ({
-          type: "tool_result",
-          toolCallId: id,
-          content: "Tool execution was interrupted.",
-          isError: true
-        }))
-      });
+      messages.splice(i + 1, 0, { role: "tool", content: orphanCalls.map(repaired) });
     }
   }
   const toolCallIdSet = /* @__PURE__ */ new Set();
@@ -1759,6 +1880,8 @@ var Agent = class {
   Agent,
   AgentStream,
   agentLoop,
+  cancelledBeforeStartText,
+  indeterminateOutcomeText,
   isAbortError,
   isBillingError,
   isContextOverflow,