@wrongstack/core 0.275.0 → 0.276.2

package/dist/execution/index.d.ts

@@ -1,24 +1,24 @@
-export { A as AutoCompactionMiddleware, a as AutonomousRunner, b as AutonomousRunnerOptions, c as AutonomyPromptContributorOptions, C as CompactorStrategy, U as ContextWindowBudgetSnapshot, D as DefaultDesignKitLoader, d as DefaultPromptLoader, e as DefaultSkillLoader, f as DesignKitLoaderOptions, h as DoneCheckResult, i as DoneConditionChecker, I as IntelligentCompactor, j as IntelligentCompactorOptions, P as PersistedActiveKit, k as PromptLoaderOptions, S as SelectiveCompactor, l as SelectiveCompactorOptions, m as SkillLoaderOptions, n as StrategyCompactorOptions, _ as _resetDesignKitLoaderMemo, o as _resetDesignRulesCache, p as activateDesign, r as buildGoalPreamble, s as clearActiveKit, t as clearPersistedActiveKit, u as createStrategyCompactor, v as designProjectDir, w as detectFrontendFile, x as detectFrontendIntent, y as getDesignKitLoader, z as getDesignState, B as installDesignStudioMiddleware, E as loadActiveKit, F as loadProjectDesignRules, G as makeAutonomyPromptContributor, H as makeDesignDetectToolCallMiddleware, J as makeDesignDetectUserInputMiddleware, K as makeDesignStudioRequestMiddleware, M as recordKitChoice, O as renderPrompt, Q as resolveBundledDesignKitsDir, R as setActiveKit } from '../strategy-compactor-C2bmlWYg.js';
-import { P as Provider, a as ReasoningConfig, n as Capabilities, c as Request } from '../context-BrLe8pJy.js';
-import { e as BrainDecision, h as BrainDecisionRequest, B as BrainArbiter } from '../brain-DCkB5_e7.js';
-export { C as CompactorOptions, h as DefaultErrorHandler, i as DefaultRetryPolicy, E as EternalAutonomyEngine, k as EternalAutonomyOptions, l as EternalEngineState, H as HybridCompactor, I as IterationStage, P as ParallelEngineState, m as ParallelEternalEngine, n as ParallelEternalOptions, o as ParallelIterationStage, T as ToolExecutor } from '../tool-executor-D4YdaJ-M.js';
-import { N as ModelRuntimeConfig } from '../config-BzFRKkg7.js';
-import '../pipeline-BYR-Vdau.js';
-import '../mailbox-types-_7gaY0Rl.js';
-import '../session-event-bridge-m7y--I-H.js';
-import '../compactor-CzSvxM1g.js';
-import '../agent-subagent-runner-CArSFKFl.js';
-import '../index-CtlizLTK.js';
+export { A as AutoCompactionMiddleware, a as AutonomousRunner, b as AutonomousRunnerOptions, c as AutonomyPromptContributorOptions, C as CompactorStrategy, U as ContextWindowBudgetSnapshot, D as DefaultDesignKitLoader, d as DefaultPromptLoader, e as DefaultSkillLoader, f as DesignKitLoaderOptions, h as DoneCheckResult, i as DoneConditionChecker, I as IntelligentCompactor, j as IntelligentCompactorOptions, P as PersistedActiveKit, k as PromptLoaderOptions, S as SelectiveCompactor, l as SelectiveCompactorOptions, m as SkillLoaderOptions, n as StrategyCompactorOptions, _ as _resetDesignKitLoaderMemo, o as _resetDesignRulesCache, p as activateDesign, r as buildGoalPreamble, s as clearActiveKit, t as clearPersistedActiveKit, u as createStrategyCompactor, v as designProjectDir, w as detectFrontendFile, x as detectFrontendIntent, y as getDesignKitLoader, z as getDesignState, B as installDesignStudioMiddleware, E as loadActiveKit, F as loadProjectDesignRules, G as makeAutonomyPromptContributor, H as makeDesignDetectToolCallMiddleware, J as makeDesignDetectUserInputMiddleware, K as makeDesignStudioRequestMiddleware, M as recordKitChoice, O as renderPrompt, Q as resolveBundledDesignKitsDir, R as setActiveKit } from '../strategy-compactor-DI1OHVbB.js';
+import { P as Provider, a as ReasoningConfig, o as Capabilities, c as Request } from '../context-CLnUMW5g.js';
+import { e as BrainDecision, h as BrainDecisionRequest, B as BrainArbiter } from '../brain-BLOyN5ZP.js';
+export { C as CompactorOptions, h as DefaultErrorHandler, i as DefaultRetryPolicy, E as EternalAutonomyEngine, k as EternalAutonomyOptions, l as EternalEngineState, H as HybridCompactor, I as IterationStage, P as ParallelEngineState, m as ParallelEternalEngine, n as ParallelEternalOptions, o as ParallelIterationStage, T as ToolExecutor } from '../tool-executor-Bmd5Ygoo.js';
+import { N as ModelRuntimeConfig } from '../config-Bf5mj-ad.js';
+import '../pipeline-Dl6XbfE7.js';
+import '../mailbox-types-DTl7bRH3.js';
+import '../session-event-bridge-CMvIO59_.js';
+import '../compactor-DcBpaJsI.js';
+import '../agent-subagent-runner-CEuw4ATz.js';
+import '../index-Cn0NOshr.js';
 import '../logger-B63L5bTg.js';
 import '../observability-D-HZN_mF.js';
-import '../permission-CJR1qfOi.js';
-import '../retry-policy-BmY5ooh3.js';
-import '../selector-qjpee9BF.js';
+import '../permission-BCbQDR2s.js';
+import '../retry-policy-Dt3_z8Aj.js';
+import '../selector-BCkWgdwy.js';
 import '../skill-DGIXCtdv.js';
 import '../wstack-paths-_NrRovdr.js';
 import '../prompt-DLd35n4Q.js';
-import '../goal-store-DUwdbdoY.js';
-import '../multi-agent-coordinator-Dc_HuG9p.js';
+import '../goal-store-DGb6b5Ed.js';
+import '../multi-agent-coordinator-CiRtKVTk.js';
 import 'node:events';
 
 /**

package/dist/execution/index.js

@@ -1515,6 +1515,7 @@ var ERROR_CODES = {
   SDD_NOT_READY: "SDD_NOT_READY",
   // General
   VALIDATION_ERROR: "VALIDATION_ERROR",
+  PARSE_FAILED: "PARSE_FAILED",
   UNKNOWN: "UNKNOWN"
 };
 var WrongStackError = class extends Error {
@@ -1545,6 +1546,21 @@ function formatContext(ctx) {
   const parts = Object.entries(ctx).filter(([, v]) => v !== void 0).slice(0, 3).map(([k, v]) => `${k}=${String(v)}`);
   return parts.length > 0 ? `[${parts.join(" ")}]` : "";
 }
+var ToolError = class extends WrongStackError {
+  toolName;
+  constructor(opts) {
+    super({
+      message: opts.message,
+      code: opts.code,
+      subsystem: "tool",
+      recoverable: opts.recoverable,
+      context: { tool: opts.toolName, ...opts.context },
+      cause: opts.cause
+    });
+    this.name = "ToolError";
+    this.toolName = opts.toolName;
+  }
+};
 var AgentError = class extends WrongStackError {
   constructor(opts) {
     super({
@@ -1614,6 +1630,9 @@ var ToolValidationError = class extends WrongStackError {
     this.name = "ToolValidationError";
   }
 };
+function isWrongStackError(err) {
+  return err instanceof WrongStackError;
+}
 
 // src/execution/auto-compaction-middleware.ts
 var LEVEL_RANK = { warn: 0, soft: 1, hard: 2 };
@@ -1812,6 +1831,7 @@ var AutoCompactionMiddleware = class _AutoCompactionMiddleware {
       const report = await this.compactor.compact(ctx, { aggressive });
       this.recordAttempt(pressure.level, pressure.tokens, report);
       this.events?.emit("compaction.fired", {
+        sessionId: ctx.session.id,
         level: pressure.level,
         tokens: pressure.tokens,
         load: pressure.load,
@@ -1846,6 +1866,7 @@ var AutoCompactionMiddleware = class _AutoCompactionMiddleware {
           `Auto-compaction left context above the hard threshold after ${pressure.level} compaction`
         );
         this.events?.emit("compaction.failed", {
+          sessionId: ctx.session.id,
           err: error,
           aggressive,
           level: pressure.level,
@@ -1873,6 +1894,7 @@ var AutoCompactionMiddleware = class _AutoCompactionMiddleware {
       const error = err instanceof Error ? err : new Error(String(err));
       const fatal = this.failureMode === "throw" || this.failureMode === "throw_on_hard" && pressure.level === "hard";
       this.events?.emit("compaction.failed", {
+        sessionId: ctx.session.id,
         err: error,
         aggressive,
         level: pressure.level,
@@ -2498,6 +2520,18 @@ function isTextBlock(b) {
 }
 
 // src/execution/compaction-core.ts
+var FAILURE_PATTERN = /(error|fail|exception|timeout|enonet|eacces|eperm|enoent|abort)/i;
+var CORRECTION_PATTERN = /\b(wrong|no\b|stop\b|don'?t\b|actually|fix that|undo|revert|forget|ignore|skip)\b/i;
+var ERROR_LANG_PATTERN = /\b(error|exception|fatal|critical|crash|panic|abort|segfault|core dump|undefined is not|null pointer|typeerror|referenceerror|syntaxerror)\b/i;
+var SECURITY_PATTERN = /\b(security|vulnerability|injection|xss|csrf|secret|apikey|api.key|hardcoded|leak|exploit|cve)\b/i;
+var ARCHITECTURE_PATTERN = /\b(architecture|design|approach|strategy|pattern|refactor|migrate|restructure|decision|trade.?off)\b/i;
+var BOILERPLATE_PATTERN = /\b(files_with_matches|count|found \d+ match|directory tree|\.\.\. and \d+ more)\b/i;
+var PATH_HINT_PATTERN = /(?:(?:[A-Za-z]:)?[./\\]?[\w@.-]+(?:[\\/][\w@(). -]+)+\.[A-Za-z0-9]{1,12})/g;
+var PATH_BACKSLASH_PATTERN = /\\/g;
+var PATH_TRIM_PATTERN = /^["'`]+|["'`),;:]+$/g;
+var ERROR_LINE_PATTERN = /\b(error|exception|failed|failure|fatal|panic|timeout|denied|enoent|eacces|eperm)\b/i;
+var NEWLINE_SPLIT_PATTERN = /\r?\n/;
+var WHITESPACE_COLLAPSE_PATTERN = /\s+/g;
 function compactionDebugEnabled() {
   return process.env["NODE_ENV"] === "development" || process.env["WRONGSTACK_DEBUG"] === "1";
 }
@@ -2733,9 +2767,8 @@ function summarizeToolResultElision(block, tokens) {
 function extractPathHints(content) {
   const text = typeof content === "string" ? content : JSON.stringify(content);
   const out = /* @__PURE__ */ new Set();
-  const re = /(?:(?:[A-Za-z]:)?[./\\]?[\w@.-]+(?:[\\/][\w@(). -]+)+\.[A-Za-z0-9]{1,12})/g;
-  for (const match of text.matchAll(re)) {
-    const clean = match[0]?.replace(/\\/g, "/").replace(/^["'`]+|["'`),;:]+$/g, "");
+  for (const match of text.matchAll(PATH_HINT_PATTERN)) {
+    const clean = match[0]?.replace(PATH_BACKSLASH_PATTERN, "/").replace(PATH_TRIM_PATTERN, "");
     if (clean && clean.length <= 220) out.add(clean);
     if (out.size >= 5) break;
   }
@@ -2743,12 +2776,9 @@ function extractPathHints(content) {
 }
 function firstErrorLine(content) {
   const text = typeof content === "string" ? content : JSON.stringify(content);
-  for (const line of text.split(/\r?\n/)) {
-    if (!/\b(error|exception|failed|failure|fatal|panic|timeout|denied|enoent|eacces|eperm)\b/i.test(
-      line
-    ))
-      continue;
-    const trimmed = line.replace(/\s+/g, " ").trim();
+  for (const line of text.split(NEWLINE_SPLIT_PATTERN)) {
+    if (!ERROR_LINE_PATTERN.test(line)) continue;
+    const trimmed = line.replace(WHITESPACE_COLLAPSE_PATTERN, " ").trim();
     if (trimmed) return trimmed.slice(0, 180);
   }
   return void 0;
@@ -2795,9 +2825,9 @@ function scoreMessage(m, context) {
     if (hasToolUse2(m) || hasResult) return 0;
   }
   if (context?.failureCounts && m.role === "user" && hasToolUse2(m) === false) {
-    const isFailure = /error|fail|exception|timeout|enonet|eacces|eperm|enoent|abort/i.test(text);
-    if (isFailure) {
-      const errKey = /(error|fail|exception|timeout|enonet|eacces|eperm|enoent|abort)/i.exec(text)?.[0]?.toLowerCase() ?? "error";
+    const failureMatch = FAILURE_PATTERN.exec(text);
+    if (failureMatch) {
+      const errKey = failureMatch[0]?.toLowerCase() ?? "error";
       const count = (context.failureCounts.get(errKey) ?? 0) + 1;
       context.failureCounts.set(errKey, count);
       if (count >= 5) return 0;
@@ -2805,29 +2835,21 @@ function scoreMessage(m, context) {
     }
   }
   if (m.role === "user") {
-    if (/\b(wrong|no\b|stop\b|don'?t\b|actually|fix that|undo|revert|forget|ignore|skip)\b/i.test(
-      text
-    )) {
+    if (CORRECTION_PATTERN.test(text)) {
       return 5;
     }
   }
-  if (/\b(error|exception|fatal|critical|crash|panic|abort|segfault|core dump|undefined is not|null pointer|typeerror|referenceerror|syntaxerror)\b/i.test(
-    text
-  )) {
+  if (ERROR_LANG_PATTERN.test(text)) {
     return 5;
   }
-  if (/\b(security|vulnerability|injection|xss|csrf|secret|apikey|api.key|hardcoded|leak|exploit|cve)\b/i.test(
-    text
-  )) {
+  if (SECURITY_PATTERN.test(text)) {
     return 5;
   }
-  if (m.role === "assistant" && /\b(architecture|design|approach|strategy|pattern|refactor|migrate|restructure|decision|trade.?off)\b/i.test(
-    text
-  )) {
+  if (m.role === "assistant" && ARCHITECTURE_PATTERN.test(text)) {
     return 5;
   }
   if (hasLargeToolResult(m)) return 1;
-  if (m.role === "user" && !hasToolUse2(m) && /\b(files_with_matches|count|found \d+ match|directory tree|\.\.\. and \d+ more)\b/i.test(text)) {
+  if (m.role === "user" && !hasToolUse2(m) && BOILERPLATE_PATTERN.test(text)) {
     return 1;
   }
   return 3;
@@ -2837,6 +2859,11 @@ function buildSmartDigest(messages) {
   const failureCounts = /* @__PURE__ */ new Map();
   let noiseCount = 0;
   for (const m of messages) {
+    const isPureToolIO = Array.isArray(m.content) && m.content.length > 0 && !hasToolUse2(m) && m.content.every((b) => b.type === "tool_result");
+    if (isPureToolIO) {
+      noiseCount++;
+      continue;
+    }
     const score = scoreMessage(m, { failureCounts });
     const text = extractText2(m);
     const toolCount = countToolBlocks(m);
@@ -3880,7 +3907,7 @@ function buildRecoveryStrategies(opts) {
       label: "downgrade_model",
       async attempt(err, ctx) {
         if (!(err instanceof ProviderError)) return null;
-        if (err.status !== 429 && err.status !== 529 && err.status < 500) return null;
+        if (err.status !== 529 && err.status < 500) return null;
         return null;
       }
     }
@@ -4544,6 +4571,7 @@ Deliverables: ${goal.deliverables.length} total, progress ${goal.progress ?? "un
     try {
       const decision = await this.opts.brain.decide({
         id: `goal-done-${goal.iterations}`,
+        sessionId: this.opts.agent.ctx.session?.id,
         source: "system",
         question: `Brainstorm returned DONE ${this.consecutiveBrainstormDone}x. Is the goal truly complete?`,
         context: [
@@ -4937,6 +4965,7 @@ var SubagentBudget = class _SubagentBudget {
    */
   lastActivityTime = null;
   _onThreshold;
+  _sessionId;
   /**
    * Hard cap on how long `_negotiateExtension` waits for the coordinator to
    * respond before defaulting to 'stop'. Without this fallback an absent
@@ -5005,10 +5034,15 @@ var SubagentBudget = class _SubagentBudget {
   get mode() {
     return this._mode;
   }
-  constructor(limits = {}, mode = "auto") {
+  constructor(limits = {}, mode = "auto", options = {}) {
     this._mode = mode;
+    this._sessionId = options.sessionId;
     this.limits = { ...limits };
   }
+  currentSessionId() {
+    const value = typeof this._sessionId === "function" ? this._sessionId() : this._sessionId;
+    return typeof value === "string" && value.length > 0 ? value : void 0;
+  }
   start() {
     this.startTime = Date.now();
     this.lastActivityTime = this.startTime;
@@ -5176,7 +5210,9 @@ var SubagentBudget = class _SubagentBudget {
         resolve5(d);
       };
       const fallback = setTimeout(() => respond("stop"), _SubagentBudget.DECISION_TIMEOUT_MS);
+      const sessionId = this.currentSessionId();
       bus.emit("budget.threshold_reached", {
+        ...sessionId ? { sessionId } : {},
         kind: entry.kind,
         used: entry.used,
         limit: entry.limit,
@@ -7427,6 +7463,7 @@ var DefaultMultiAgentCoordinator = class _DefaultMultiAgentCoordinator extends E
   coordinatorId;
   config;
   runner;
+  sessionId;
   fleetBus;
   subagents = /* @__PURE__ */ new Map();
   /**
@@ -7461,6 +7498,11 @@ var DefaultMultiAgentCoordinator = class _DefaultMultiAgentCoordinator extends E
     this.coordinatorId = config.coordinatorId;
     this.config = config;
     this.runner = options.runner;
+    this.sessionId = options.sessionId;
+  }
+  currentSessionId() {
+    const value = typeof this.sessionId === "function" ? this.sessionId() : this.sessionId;
+    return typeof value === "string" && value.length > 0 ? value : void 0;
   }
   /**
    * Replace the runner after construction. Used when the runner depends
@@ -7619,11 +7661,16 @@ var DefaultMultiAgentCoordinator = class _DefaultMultiAgentCoordinator extends E
       status: s.status,
       assigned: s.context.parentBridge !== null
     }));
+    const sessionId = this.currentSessionId();
     this.fleetBus?.emit({
       subagentId: this.coordinatorId,
       ts: Date.now(),
       type: "coordinator.stats",
-      payload: { ...stats, subagentStatuses }
+      payload: {
+        ...sessionId ? { sessionId } : {},
+        ...stats,
+        subagentStatuses
+      }
     });
   }
   getStatus() {
@@ -7826,16 +7873,20 @@ var DefaultMultiAgentCoordinator = class _DefaultMultiAgentCoordinator extends E
     const rawTimeoutMs = subagent.config.timeoutMs;
     const rawIdleTimeoutMs = subagent.config.idleTimeoutMs;
     const configWithRosterDefaults = applyRosterBudget(subagent.config);
-    const budget = new SubagentBudget({
-      maxIterations: rawMaxIterations ?? this.config.defaultBudget?.maxIterations ?? configWithRosterDefaults.maxIterations,
-      maxToolCalls: rawMaxToolCalls ?? this.config.defaultBudget?.maxToolCalls ?? configWithRosterDefaults.maxToolCalls,
-      maxTokens: rawMaxTokens ?? this.config.defaultBudget?.maxTokens ?? configWithRosterDefaults.maxTokens,
-      maxCostUsd: rawMaxCostUsd ?? this.config.defaultBudget?.maxCostUsd ?? configWithRosterDefaults.maxCostUsd,
-      // Wall-clock cap is opt-in (explicit config / defaultBudget only); the
-      // roster no longer supplies one. Idle is the default reaper.
-      timeoutMs: rawTimeoutMs ?? this.config.defaultBudget?.timeoutMs ?? configWithRosterDefaults.timeoutMs,
-      idleTimeoutMs: rawIdleTimeoutMs ?? this.config.defaultBudget?.idleTimeoutMs ?? configWithRosterDefaults.idleTimeoutMs
-    });
+    const budget = new SubagentBudget(
+      {
+        maxIterations: rawMaxIterations ?? this.config.defaultBudget?.maxIterations ?? configWithRosterDefaults.maxIterations,
+        maxToolCalls: rawMaxToolCalls ?? this.config.defaultBudget?.maxToolCalls ?? configWithRosterDefaults.maxToolCalls,
+        maxTokens: rawMaxTokens ?? this.config.defaultBudget?.maxTokens ?? configWithRosterDefaults.maxTokens,
+        maxCostUsd: rawMaxCostUsd ?? this.config.defaultBudget?.maxCostUsd ?? configWithRosterDefaults.maxCostUsd,
+        // Wall-clock cap is opt-in (explicit config / defaultBudget only); the
+        // roster no longer supplies one. Idle is the default reaper.
+        timeoutMs: rawTimeoutMs ?? this.config.defaultBudget?.timeoutMs ?? configWithRosterDefaults.timeoutMs,
+        idleTimeoutMs: rawIdleTimeoutMs ?? this.config.defaultBudget?.idleTimeoutMs ?? configWithRosterDefaults.idleTimeoutMs
+      },
+      "auto",
+      { sessionId: () => this.currentSessionId() }
+    );
     subagent.activeBudget = budget;
     if (!this.runner) {
       return;
@@ -7938,7 +7989,9 @@ var DefaultMultiAgentCoordinator = class _DefaultMultiAgentCoordinator extends E
                 resolveDecision(d);
               };
               const fallback = setTimeout(() => resolve5("stop"), DECISION_TIMEOUT_MS);
+              const sessionId = this.currentSessionId();
               budget._events?.emit("budget.threshold_reached", {
+                ...sessionId ? { sessionId } : {},
                 kind: "timeout",
                 used,
                 limit,
@@ -7972,7 +8025,9 @@ var DefaultMultiAgentCoordinator = class _DefaultMultiAgentCoordinator extends E
         const wallExceeded = wallLimit !== void 0 && elapsed >= wallLimit;
         const idleExceeded = idleLimit !== void 0 && budget.idleMs() >= idleLimit;
         if (idleExceeded && !wallExceeded) {
+          const sessionId = this.currentSessionId();
           budget._events?.emit("budget.threshold_reached", {
+            ...sessionId ? { sessionId } : {},
             kind: "idle_timeout",
             used: budget.idleMs(),
             limit: idleLimit ?? 0,
@@ -9762,6 +9817,64 @@ var ToolExecutor = class _ToolExecutor {
     const mode = resolveToolResultRenderMode(modes, toolName);
     renderer.setResultRenderMode(toolName, mode);
   }
+  toolLogBase(ctx, use, toolName, durationMs) {
+    return {
+      event: "tool.execution",
+      traceId: ctx.traceId,
+      sessionId: ctx.session.id,
+      agentId: ctx.agentId ?? "<unknown>",
+      toolName,
+      toolUseId: use.id,
+      durationMs
+    };
+  }
+  logToolSuccess(ctx, use, toolName, durationMs, outputChars) {
+    this.opts.events?.emit("tool.completed", {
+      name: toolName,
+      id: use.id,
+      sessionId: ctx.session.id,
+      ...ctx.traceId ? { traceId: ctx.traceId } : {},
+      agentId: ctx.agentId ?? "<unknown>",
+      durationMs,
+      outputChars
+    });
+    this.opts.logger?.info("tool execution completed", {
+      ...this.toolLogBase(ctx, use, toolName, durationMs),
+      outcome: "success",
+      isError: false,
+      outputChars
+    });
+  }
+  logToolFailure(ctx, use, toolName, durationMs, err) {
+    const { category, retryable, detail } = classifyToolError(err);
+    const structured = isWrongStackError(err);
+    const structuredError = structured ? {
+      errorCode: err.code,
+      errorSubsystem: err.subsystem,
+      errorSeverity: err.severity
+    } : {};
+    this.opts.events?.emit("tool.failed", {
+      name: toolName,
+      id: use.id,
+      sessionId: ctx.session.id,
+      ...ctx.traceId ? { traceId: ctx.traceId } : {},
+      agentId: ctx.agentId ?? "<unknown>",
+      durationMs,
+      category,
+      retryable,
+      ...detail ? { detail } : {},
+      ...structuredError
+    });
+    this.opts.logger?.warn("tool execution failed", {
+      ...this.toolLogBase(ctx, use, toolName, durationMs),
+      outcome: "failure",
+      isError: true,
+      errorCategory: category,
+      retryable,
+      errorDetail: detail,
+      ...structuredError
+    });
+  }
   /**
    * Execute a batch of tool uses using the configured strategy.
    * Returns the execution results and the remaining output budget.
@@ -9892,7 +10005,8 @@ ${errorDetails}`,
         "tool.has_dangerous_capabilities": toolCapsForAudit.length > 0
       });
       try {
-        let { block: result, bytes } = await this.executeTool(tool, use, ctx, budget);
+        const producedText = await this.produceToolOutput(tool, use, ctx, budget);
+        let { block: result, bytes } = this.settleToolOutput(tool, use, producedText, budget);
         budget -= bytes;
         if (this.opts.hookRunner?.has("PostToolUse")) {
           const post = await this.opts.hookRunner.postToolUse(
@@ -9909,13 +10023,18 @@ ${post.additionalContext}`;
             budget = Math.max(0, budget - Buffer.byteLength(appended, "utf8"));
           }
         }
+        const outputChars = typeof result.content === "string" ? result.content.length : 0;
         span?.setAttribute("tool.is_error", !!result.is_error);
-        span?.setAttribute(
-          "tool.output_bytes",
-          typeof result.content === "string" ? result.content.length : 0
-        );
+        span?.setAttribute("tool.output_bytes", outputChars);
+        this.logToolSuccess(ctx, use, tool.name, Date.now() - start, outputChars);
         return { result, tool, durationMs: Date.now() - start };
       } catch (err) {
+        if (isWrongStackError(err)) {
+          if (err instanceof Error) span?.recordError(err);
+          span?.setAttribute("tool.is_error", true);
+          this.logToolFailure(ctx, use, tool.name, Date.now() - start, err);
+          throw err;
+        }
         const msg = toErrorMessage(err);
         const scrubbed = this.opts.secretScrubber.scrub(msg);
         const { category, retryable, detail } = classifyToolError(err);
@@ -9933,6 +10052,7 @@ ${post.additionalContext}`;
         span?.setAttribute("tool.error_category", category);
         span?.setAttribute("tool.error_retryable", retryable);
         if (detail) span?.setAttribute("tool.error_detail", detail);
+        this.logToolFailure(ctx, use, tool.name, Date.now() - start, err);
         return { result, tool, durationMs: Date.now() - start };
       } finally {
         span?.end();
@@ -9942,7 +10062,8 @@ ${post.additionalContext}`;
       try {
         return await runOne(use);
       } catch (err) {
-        const msg = toErrorMessage(err);
+        const isStructured = isWrongStackError(err);
+        const msg = isStructured ? err.describe() : toErrorMessage(err);
         const scrubbed = this.opts.secretScrubber.scrub(msg);
         const { category, retryable, detail } = classifyToolError(err);
         const tool = this.registry.get(use.name);
@@ -9952,7 +10073,7 @@ ${post.additionalContext}`;
         const result = {
           type: "tool_result",
           tool_use_id: use.id,
-          content: `Tool "${use.name}" execution failed: ${scrubbed}`,
+          content: isStructured ? scrubbed : `Tool "${use.name}" execution failed: ${scrubbed}`,
           is_error: true
         };
         budget = this.budgetForString(result.content, budget);
@@ -9995,7 +10116,23 @@ ${post.additionalContext}`;
    * to call a tool" and "tool.executed".
    */
   async executeTool(tool, use, ctx, budget) {
+    const text = await this.produceToolOutput(tool, use, ctx, budget);
+    return this.settleToolOutput(tool, use, text, budget);
+  }
+  /**
+   * Async "produce" phase: run the tool, serialize, scrub, and spill an
+   * oversized output to a disk artifact. Returns the pre-cap text. This is
+   * the long-running, concurrency-safe part — it touches NO shared budget
+   * state, so multiple invocations can run in parallel without racing.
+   *
+   * The `budgetHint` only gates the disk-spill threshold (a heuristic for
+   * "is this output large enough to persist"); it is NOT the output cap.
+   * The authoritative cap is applied synchronously in settleToolOutput()
+   * against the live budget.
+   */
+  async produceToolOutput(tool, use, ctx, budgetHint) {
     this.opts.events?.emit("tool.started", {
+      sessionId: ctx.session.id,
       name: tool.name,
       id: use.id,
       input: use.input
@@ -10004,8 +10141,20 @@ ${post.additionalContext}`;
     const output = await this.runWithTimeout(tool, use.input, ctx.signal, ctx, use.id);
     const text = this.serializer.serialize(output, { toolName: tool.name, input: use.input, tool });
     const scrubbed = this.opts.secretScrubber.scrub(text);
-    const withArtifact = await maybePersistLargeToolOutput(tool.name, scrubbed, budget);
-    const { text: capped, newBudget } = this.serializer.enforceCap(withArtifact, budget);
+    return maybePersistLargeToolOutput(tool.name, scrubbed, budgetHint);
+  }
+  /**
+   * Synchronous "settle" phase: enforce the output cap against the CURRENT
+   * budget, render, and build the result block. This MUST stay synchronous
+   * (no awaits) so that, in the parallel/smart strategies, two tools settling
+   * one after another against the shared closure budget can't interleave a
+   * stale read with a write. The first to settle consumes its bytes; the next
+   * settles against the reduced budget; once the budget hits 0, enforceCap
+   * truncates — making the per-iteration cap genuinely CUMULATIVE across
+   * parallel tools instead of degrading into a per-tool cap.
+   */
+  settleToolOutput(tool, use, text, budget) {
+    const { text: capped, newBudget } = this.serializer.enforceCap(text, budget);
     this.hintRenderMode(tool.name);
     this.opts.renderer?.writeToolResult(tool.name, capped, false);
     return {
@@ -10058,7 +10207,12 @@ ${post.additionalContext}`;
     let finalOutput;
     let sawFinal = false;
     if (!tool.executeStream) {
-      throw new Error(`Tool "${tool.name}" does not support streaming execution`);
+      throw new ToolError({
+        message: `Tool "${tool.name}" does not support streaming execution`,
+        code: "TOOL_EXECUTION_FAILED",
+        toolName: tool.name,
+        context: { reason: "streaming_not_supported" }
+      });
     }
     const stream = tool.executeStream(input, ctx, { signal });
     const iter = stream[Symbol.asyncIterator]();
@@ -10068,6 +10222,7 @@ ${post.additionalContext}`;
     let lastProgressEmitAt = 0;
     const emitProgress = (ev) => {
       this.opts.events?.emit("tool.progress", {
+        sessionId: ctx.session.id,
         name: tool.name,
         id: toolUseId ?? "<unknown>",
         event: ev
@@ -10123,7 +10278,12 @@ ${progressTail}`;
       await iter.return?.(void 0);
     }
     if (!sawFinal) {
-      throw new Error(`tool "${tool.name}" executeStream completed without a 'final' event`);
+      throw new ToolError({
+        message: `tool "${tool.name}" executeStream completed without a 'final' event`,
+        code: "TOOL_EXECUTION_FAILED",
+        toolName: tool.name,
+        context: { reason: "missing_final_event" }
+      });
     }
     return finalOutput;
   }
@@ -10247,6 +10407,15 @@ function classifyToolError(err) {
   if (err instanceof Error && err.message.includes("validation")) {
     return { category: "validation" /* VALIDATION */, retryable: false, detail: "validation" };
   }
+  if (err instanceof WrongStackError) {
+    const wse = err;
+    const category = wse.severity === "warning" ? "transient" /* TRANSIENT */ : "fatal" /* FATAL */;
+    return {
+      category,
+      retryable: wse.recoverable,
+      detail: `${wse.code} [${wse.subsystem}]`
+    };
+  }
   return {
     category: "fatal" /* FATAL */,
     retryable: false,