@pixelbyte-software/pixcode 1.42.1 → 1.42.2

package/scripts/smoke/workflow-fallback-replay.mjs

@@ -0,0 +1,56 @@
+#!/usr/bin/env node
+
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+
+const root = process.cwd();
+
+function read(relativePath) {
+  return fs.readFileSync(path.join(root, relativePath), 'utf8');
+}
+
+const fallbackPolicy = read('server/modules/orchestration/workflows/workflow-fallback-policy.ts');
+assert.match(fallbackPolicy, /PIXCODE_FALLBACK_POLICY_PROTOCOL/, 'Fallback policy should declare a stable protocol id.');
+assert.match(fallbackPolicy, /pixcode\.fallback-policy\.v1/, 'Fallback policy should use the v1 protocol id.');
+assert.match(fallbackPolicy, /provider_failure/, 'Fallback policy should classify provider failures.');
+assert.match(fallbackPolicy, /timeout/, 'Fallback policy should classify timeouts.');
+assert.match(fallbackPolicy, /tool_failure/, 'Fallback policy should classify tool failures.');
+assert.match(fallbackPolicy, /invalid_output/, 'Fallback policy should classify invalid output.');
+assert.match(fallbackPolicy, /resolveWorkflowFallbackDecision/, 'Fallback policy should expose a decision helper.');
+
+const replay = read('server/modules/orchestration/workflows/workflow-replay.ts');
+assert.match(replay, /PIXCODE_REPLAY_PROTOCOL/, 'Replay support should declare a stable protocol id.');
+assert.match(replay, /pixcode\.workflow-replay\.v1/, 'Replay support should use the v1 protocol id.');
+assert.match(replay, /buildWorkflowReplayPlan/, 'Replay support should build a replay plan from stored run data.');
+assert.match(replay, /requiresApproval/, 'Replay plans should expose approval requirements.');
+assert.match(replay, /file-write/, 'Replay safety should detect file-write actions.');
+assert.match(replay, /shell/, 'Replay safety should detect shell actions.');
+assert.match(replay, /network/, 'Replay safety should detect network actions.');
+
+const runner = read('server/modules/orchestration/workflows/workflow-runner.ts');
+assert.match(runner, /resolveWorkflowFallbackDecision/, 'Workflow runner should use policy-driven fallback decisions.');
+assert.match(runner, /fallbackTrigger/, 'Fallback nodes should record the trigger that launched them.');
+assert.match(runner, /fallbackSkippedEvents/, 'Skipped fallback decisions should be recorded.');
+
+const trace = read('server/modules/orchestration/workflows/workflow-trace.ts');
+assert.match(trace, /workflow\.trace\.fallback/, 'Trace timeline should surface fallback events.');
+assert.match(trace, /workflow\.trace\.replay/, 'Trace timeline should surface replay metadata.');
+
+const routes = read('server/modules/orchestration/workflows/workflow.routes.ts');
+assert.match(routes, /replay-plan/, 'Workflow routes should expose a replay plan endpoint.');
+assert.match(routes, /REPLAY_APPROVAL_REQUIRED/, 'Replay route should require approval before unsafe replay.');
+assert.match(routes, /workflowRunner\.start\(\s*replayPlan\.workflow/s, 'Replay route should start from the generated replay workflow.');
+
+const panel = read('src/components/orchestration/workflows/WorkflowRunPanel.tsx');
+assert.match(panel, /loadReplayPlan/, 'Workflow run panel should load replay plans.');
+assert.match(panel, /replayRun/, 'Workflow run panel should expose a replay action.');
+assert.match(panel, /approveReplay/, 'Workflow run panel should expose explicit approval for guarded replay.');
+assert.match(panel, /orchestration\.replayRun/, 'Workflow UI should render replay labels.');
+
+const en = read('src/i18n/locales/en/common.json');
+const tr = read('src/i18n/locales/tr/common.json');
+assert.match(en, /"replayRun"/, 'English replay translation is missing.');
+assert.match(tr, /"replayRun"/, 'Turkish replay translation is missing.');
+
+console.log('workflow fallback replay smoke passed');

package/server/modules/orchestration/workflows/workflow-fallback-policy.ts

@@ -0,0 +1,161 @@
+import type { WorkflowNode, WorkflowRun } from '@/modules/orchestration/workflows/workflow.types.js';
+
+export const PIXCODE_FALLBACK_POLICY_PROTOCOL = 'pixcode.fallback-policy.v1';
+
+export type WorkflowFallbackTrigger = 'provider_failure' | 'timeout' | 'tool_failure' | 'invalid_output';
+
+export interface WorkflowFallbackPolicy {
+  protocol: typeof PIXCODE_FALLBACK_POLICY_PROTOCOL;
+  enabled: boolean;
+  triggers: WorkflowFallbackTrigger[];
+  maxFallbacksPerRun: number;
+  requireDifferentAgent: boolean;
+}
+
+export interface WorkflowFallbackDecision {
+  shouldFallback: boolean;
+  trigger: WorkflowFallbackTrigger;
+  reason: string;
+  policy: WorkflowFallbackPolicy;
+  skippedReason?: string;
+}
+
+const DEFAULT_TRIGGERS: WorkflowFallbackTrigger[] = [
+  'provider_failure',
+  'timeout',
+  'tool_failure',
+  'invalid_output',
+];
+
+export const DEFAULT_WORKFLOW_FALLBACK_POLICY: WorkflowFallbackPolicy = {
+  protocol: PIXCODE_FALLBACK_POLICY_PROTOCOL,
+  enabled: true,
+  triggers: DEFAULT_TRIGGERS,
+  maxFallbacksPerRun: 3,
+  requireDifferentAgent: true,
+};
+
+function readRecord(value: unknown): Record<string, unknown> | undefined {
+  return value && typeof value === 'object' ? value as Record<string, unknown> : undefined;
+}
+
+function readBoolean(value: unknown): boolean | undefined {
+  return typeof value === 'boolean' ? value : undefined;
+}
+
+function readNumber(value: unknown): number | undefined {
+  return typeof value === 'number' && Number.isFinite(value) ? value : undefined;
+}
+
+function isWorkflowFallbackTrigger(value: unknown): value is WorkflowFallbackTrigger {
+  return value === 'provider_failure'
+    || value === 'timeout'
+    || value === 'tool_failure'
+    || value === 'invalid_output';
+}
+
+function readFallbackTriggers(value: unknown): WorkflowFallbackTrigger[] | undefined {
+  if (!Array.isArray(value)) return undefined;
+  const triggers = value.filter(isWorkflowFallbackTrigger);
+  return triggers.length > 0 ? [...new Set(triggers)] : undefined;
+}
+
+export function readWorkflowFallbackPolicy(metadata?: Record<string, unknown>): WorkflowFallbackPolicy {
+  const settings = readRecord(metadata?.settings) ?? {};
+  const configured = readRecord(settings.fallbackPolicy) ?? {};
+  const maxFallbacksPerRun = readNumber(configured.maxFallbacksPerRun);
+
+  return {
+    protocol: PIXCODE_FALLBACK_POLICY_PROTOCOL,
+    enabled: readBoolean(configured.enabled) ?? DEFAULT_WORKFLOW_FALLBACK_POLICY.enabled,
+    triggers: readFallbackTriggers(configured.triggers) ?? DEFAULT_WORKFLOW_FALLBACK_POLICY.triggers,
+    maxFallbacksPerRun: maxFallbacksPerRun === undefined
+      ? DEFAULT_WORKFLOW_FALLBACK_POLICY.maxFallbacksPerRun
+      : Math.max(0, Math.min(8, Math.round(maxFallbacksPerRun))),
+    requireDifferentAgent: readBoolean(configured.requireDifferentAgent)
+      ?? DEFAULT_WORKFLOW_FALLBACK_POLICY.requireDifferentAgent,
+  };
+}
+
+export function classifyWorkflowFailure(
+  reason: string,
+  explicitTrigger?: WorkflowFallbackTrigger,
+): WorkflowFallbackTrigger {
+  if (explicitTrigger) return explicitTrigger;
+
+  const text = reason.toLocaleLowerCase('en');
+  if (/timed out|timeout|deadline/u.test(text)) return 'timeout';
+  if (/invalid (handoff|output|artifact|json|schema)|parse|protocol/u.test(text)) return 'invalid_output';
+  if (
+    /tool|command|shell|exit code|permission|file write|write failed|network|fetch|curl|wget|gh |npm |git /u
+      .test(text)
+  ) {
+    return 'tool_failure';
+  }
+  return 'provider_failure';
+}
+
+function fallbackEventCount(run: WorkflowRun): number {
+  return Array.isArray(run.metadata?.fallbackEvents) ? run.metadata.fallbackEvents.length : 0;
+}
+
+export function resolveWorkflowFallbackDecision({
+  run,
+  node,
+  reason,
+  trigger,
+  fallbackAgentInstanceId,
+}: {
+  run: WorkflowRun;
+  node: WorkflowNode;
+  reason: string;
+  trigger?: WorkflowFallbackTrigger;
+  fallbackAgentInstanceId?: string;
+}): WorkflowFallbackDecision {
+  const fallbackTrigger = classifyWorkflowFailure(reason, trigger);
+  const policy = readWorkflowFallbackPolicy(run.metadata);
+
+  if (!policy.enabled) {
+    return {
+      shouldFallback: false,
+      trigger: fallbackTrigger,
+      reason,
+      policy,
+      skippedReason: 'Fallback policy is disabled.',
+    };
+  }
+  if (!policy.triggers.includes(fallbackTrigger)) {
+    return {
+      shouldFallback: false,
+      trigger: fallbackTrigger,
+      reason,
+      policy,
+      skippedReason: `Fallback trigger ${fallbackTrigger} is not enabled.`,
+    };
+  }
+  if (fallbackEventCount(run) >= policy.maxFallbacksPerRun) {
+    return {
+      shouldFallback: false,
+      trigger: fallbackTrigger,
+      reason,
+      policy,
+      skippedReason: `Fallback limit ${policy.maxFallbacksPerRun} reached.`,
+    };
+  }
+  if (policy.requireDifferentAgent && fallbackAgentInstanceId && fallbackAgentInstanceId === node.agentInstanceId) {
+    return {
+      shouldFallback: false,
+      trigger: fallbackTrigger,
+      reason,
+      policy,
+      skippedReason: 'Fallback agent must be different from the failed agent.',
+    };
+  }
+
+  return {
+    shouldFallback: true,
+    trigger: fallbackTrigger,
+    reason,
+    policy,
+  };
+}

package/server/modules/orchestration/workflows/workflow-replay.ts

@@ -0,0 +1,254 @@
+import type {
+  Workflow,
+  WorkflowNode,
+  WorkflowNodeRun,
+  WorkflowRun,
+} from '@/modules/orchestration/workflows/workflow.types.js';
+import { redactTraceText } from '@/modules/orchestration/workflows/workflow-trace.js';
+
+export const PIXCODE_REPLAY_PROTOCOL = 'pixcode.workflow-replay.v1';
+
+export type WorkflowReplayScope = 'run' | 'node';
+export type WorkflowReplaySafetyKind = 'file-write' | 'shell' | 'network';
+
+export interface WorkflowReplayOperation {
+  kind: WorkflowReplaySafetyKind;
+  nodeId?: string;
+  summary: string;
+}
+
+export interface WorkflowReplayPlan {
+  protocol: typeof PIXCODE_REPLAY_PROTOCOL;
+  sourceRunId: string;
+  sourceWorkflowId: string;
+  scope: WorkflowReplayScope;
+  fromNodeId?: string;
+  selectedNodeIds: string[];
+  requiresApproval: boolean;
+  approvalReasons: string[];
+  destructiveOperations: WorkflowReplayOperation[];
+  limitations: string[];
+  input: string;
+  workflow: Workflow;
+  metadata: Record<string, unknown>;
+}
+
+function safeNodeId(value: string): string {
+  return value.replace(/[^a-zA-Z0-9_]+/g, '_').slice(0, 48) || 'node';
+}
+
+function readRecord(value: unknown): Record<string, unknown> | undefined {
+  return value && typeof value === 'object' ? value as Record<string, unknown> : undefined;
+}
+
+function visibleNodes(run: WorkflowRun): WorkflowNodeRun[] {
+  return run.nodeRuns.filter((node) => !node.internal);
+}
+
+function defaultReplayNode(run: WorkflowRun): WorkflowNodeRun | undefined {
+  return visibleNodes(run).find((node) => node.status === 'failed')
+    ?? [...visibleNodes(run)].reverse().find((node) => node.status !== 'skipped')
+    ?? visibleNodes(run)[0];
+}
+
+function selectReplayNodes(run: WorkflowRun, scope: WorkflowReplayScope, fromNodeId?: string): WorkflowNodeRun[] {
+  const nodes = visibleNodes(run);
+  if (scope === 'run') return nodes;
+
+  const requested = fromNodeId
+    ? nodes.find((node) => node.nodeId === fromNodeId)
+    : defaultReplayNode(run);
+  return requested ? [requested] : [];
+}
+
+function compact(value: string | undefined, run: WorkflowRun, maxLength = 1_200): string | undefined {
+  return redactTraceText(value, run, maxLength);
+}
+
+function nodeTraceSummary(run: WorkflowRun, node: WorkflowNodeRun): string {
+  const artifactTypes = (node.artifacts ?? []).map((artifact) => artifact.type).filter(Boolean);
+  return [
+    `Step: ${node.agentLabel || node.nodeId}`,
+    `Node id: ${node.nodeId}`,
+    `Status: ${node.status}`,
+    node.stage ? `Stage: ${node.stage}` : undefined,
+    node.adapterId ? `Adapter: ${node.adapterId}` : undefined,
+    node.model ? `Model: ${node.model}` : undefined,
+    node.error ? `Error: ${compact(node.error, run, 800)}` : undefined,
+    artifactTypes.length > 0 ? `Artifacts: ${artifactTypes.join(', ')}` : undefined,
+    node.outputText ? `Output excerpt:\n${compact(node.outputText, run)}` : undefined,
+  ].filter(Boolean).join('\n');
+}
+
+function replayTraceSummary(run: WorkflowRun, nodes: WorkflowNodeRun[]): string {
+  return nodes.map((node) => nodeTraceSummary(run, node)).join('\n\n---\n\n');
+}
+
+function pushOperation(
+  operations: WorkflowReplayOperation[],
+  kind: WorkflowReplaySafetyKind,
+  nodeId: string | undefined,
+  summary: string,
+): void {
+  if (operations.some((operation) =>
+    operation.kind === kind && operation.nodeId === nodeId && operation.summary === summary,
+  )) {
+    return;
+  }
+  operations.push({ kind, nodeId, summary });
+}
+
+function detectReplayOperations(run: WorkflowRun, nodes: WorkflowNodeRun[]): WorkflowReplayOperation[] {
+  const operations: WorkflowReplayOperation[] = [];
+
+  for (const node of nodes) {
+    for (const artifact of node.artifacts ?? []) {
+      if (artifact.type === 'file-diff') {
+        pushOperation(operations, 'file-write', node.nodeId, 'Prior step produced a file diff artifact.');
+      }
+      if (artifact.type === 'command-output') {
+        pushOperation(operations, 'shell', node.nodeId, 'Prior step produced command output.');
+      }
+      const text = [artifact.text, artifact.data ? JSON.stringify(artifact.data) : undefined]
+        .filter(Boolean)
+        .join('\n')
+        .toLocaleLowerCase('en');
+      if (/https?:\/\/|curl |wget |gh |npm publish|npm install|git push|ssh /u.test(text)) {
+        pushOperation(operations, 'network', node.nodeId, 'Prior artifact references a network-capable operation.');
+      }
+    }
+
+    const text = [node.outputText, node.error, node.promptPreview].filter(Boolean).join('\n').toLocaleLowerCase('en');
+    if (/apply_patch|write file|file write|modified files|changed files/u.test(text)) {
+      pushOperation(operations, 'file-write', node.nodeId, 'Prior step text references file-write activity.');
+    }
+    if (/shell|command|terminal|npm run|node |python |php |go test|cargo |make |exit code/u.test(text)) {
+      pushOperation(operations, 'shell', node.nodeId, 'Prior step text references shell execution.');
+    }
+    if (/https?:\/\/|curl |wget |gh |npm publish|npm install|git push|ssh |network/u.test(text)) {
+      pushOperation(operations, 'network', node.nodeId, 'Prior step text references a network-capable operation.');
+    }
+  }
+
+  return operations;
+}
+
+function replayNodeFromRunNode(
+  node: WorkflowNodeRun,
+  index: number,
+  previousReplayNodeId: string | undefined,
+  traceSummary: string,
+  limitations: string[],
+  requiresApproval: boolean,
+): WorkflowNode {
+  const replayNodeId = `replay_${index + 1}_${safeNodeId(node.nodeId)}`;
+  return {
+    id: replayNodeId,
+    adapterId: node.adapterId || 'claude-code',
+    agentInstanceId: node.agentInstanceId,
+    agentLabel: node.agentLabel ? `${node.agentLabel} Replay` : 'Replay agent',
+    assignment: node.assignment ? `Replay: ${node.assignment}` : `Replay source node ${node.nodeId}`,
+    stage: node.stage ? `replay_${node.stage}` : 'replay',
+    model: node.model,
+    permissionMode: node.permissionMode === 'bypassPermissions' ? 'default' : node.permissionMode,
+    timeoutMs: node.timeoutMs,
+    inputs: previousReplayNodeId ? [previousReplayNodeId] : [],
+    output: 'both',
+    onFail: 'abort',
+    prompt: [
+      'This is a Pixcode workflow replay run.',
+      `Replay protocol: ${PIXCODE_REPLAY_PROTOCOL}`,
+      `Source node: ${node.nodeId}`,
+      requiresApproval
+        ? 'Replay safety review found prior shell, network, or file-write activity. Do not repeat any such action unless the current CLI permission flow asks for and receives user approval.'
+        : 'Replay safety review did not find prior shell, network, or file-write artifacts, but still avoid destructive actions unless they are required and approved.',
+      'Use the trace summary to continue from the failure or inspect the run. Do not expose secrets, local-only paths, raw tool protocol, or irrelevant logs.',
+      `Known limitations:\n- ${limitations.join('\n- ')}`,
+      `Trace summary:\n${traceSummary}`,
+      `Original step prompt:\n${node.promptPreview || '(No source prompt was stored.)'}`,
+    ].join('\n\n'),
+  };
+}
+
+export function buildWorkflowReplayPlan(
+  run: WorkflowRun,
+  options: {
+    scope?: WorkflowReplayScope;
+    fromNodeId?: string;
+  } = {},
+): WorkflowReplayPlan {
+  const scope = options.scope ?? 'node';
+  const nodes = selectReplayNodes(run, scope, options.fromNodeId);
+  if (nodes.length === 0) {
+    throw new Error('No replayable workflow steps were found.');
+  }
+
+  const limitations = [
+    'Replay uses stored run traces, prompt previews, messages, and artifacts; it cannot reproduce hidden provider state.',
+    'Replay reconstructs selected steps as a new workflow run instead of mutating the source run.',
+    'Shell, network, and file-write actions stay under the current CLI permission flow and require explicit replay approval when detected.',
+  ];
+  const destructiveOperations = detectReplayOperations(run, nodes);
+  const requiresApproval = destructiveOperations.length > 0;
+  const traceSummary = replayTraceSummary(run, nodes);
+  const replayNodes = nodes.reduce<WorkflowNode[]>((accumulator, node, index) => {
+    const previousReplayNodeId = accumulator[accumulator.length - 1]?.id;
+    accumulator.push(replayNodeFromRunNode(
+      node,
+      index,
+      scope === 'run' ? previousReplayNodeId : undefined,
+      traceSummary,
+      limitations,
+      requiresApproval,
+    ));
+    return accumulator;
+  }, []);
+  const settings = readRecord(run.metadata?.settings) ?? {};
+  const replayMetadata = {
+    protocol: PIXCODE_REPLAY_PROTOCOL,
+    sourceRunId: run.id,
+    sourceWorkflowId: run.workflowId,
+    scope,
+    fromNodeId: options.fromNodeId,
+    selectedNodeIds: nodes.map((node) => node.nodeId),
+    requiresApproval,
+    destructiveOperations,
+    limitations,
+    createdAt: Date.now(),
+  };
+
+  return {
+    protocol: PIXCODE_REPLAY_PROTOCOL,
+    sourceRunId: run.id,
+    sourceWorkflowId: run.workflowId,
+    scope,
+    fromNodeId: options.fromNodeId,
+    selectedNodeIds: nodes.map((node) => node.nodeId),
+    requiresApproval,
+    approvalReasons: destructiveOperations.map((operation) =>
+      `${operation.kind}${operation.nodeId ? ` in ${operation.nodeId}` : ''}: ${operation.summary}`,
+    ),
+    destructiveOperations,
+    limitations,
+    input: [
+      `Replay ${scope === 'run' ? 'full workflow run' : 'workflow step'} from source run ${run.id}.`,
+      run.input ? `Original request:\n${compact(run.input, run, 2_000)}` : undefined,
+    ].filter(Boolean).join('\n\n'),
+    workflow: {
+      id: `${run.workflowId}_replay`,
+      name: `Replay ${run.workflowId}`,
+      description: 'Replay generated from stored Pixcode workflow trace data.',
+      trigger: 'manual',
+      nodes: replayNodes,
+    },
+    metadata: {
+      ...run.metadata,
+      workflowName: `Replay: ${String(run.metadata?.workflowName ?? run.workflowId)}`,
+      replay: replayMetadata,
+      settings: {
+        ...settings,
+        replayMode: true,
+      },
+    },
+  };
+}

package/server/modules/orchestration/workflows/workflow-runner.ts

@@ -16,6 +16,11 @@ import {
   buildWorkflowContextPacket,
   formatContextPacketForPrompt,
 } from '@/modules/orchestration/workflows/context-packet.js';
+import {
+  type WorkflowFallbackTrigger,
+  classifyWorkflowFailure,
+  resolveWorkflowFallbackDecision,
+} from '@/modules/orchestration/workflows/workflow-fallback-policy.js';
 import {
   type ResolvedWorkspaceTarget,
   resolveWorkflowWorkspace,
@@ -1123,6 +1128,8 @@ function nodeRunFromNode(node: WorkflowNode): WorkflowNodeRun {
     timeoutMs: node.timeoutMs,
     stage: node.stage,
     internal: node.internal,
+    fallbackTrigger: node.fallbackTrigger,
+    fallbackSourceNodeId: node.fallbackSourceNodeId,
     status: 'queued',
   };
 }
@@ -1276,7 +1283,12 @@ class WorkflowRunner {
     return readAgentAssignments(run.metadata).find((agent) => agent.instanceId === fallbackAgentInstanceId);
   }
 
-  private createFallbackNode(node: WorkflowNode, fallbackAgent: AgentAssignment, reason: string): WorkflowNode {
+  private createFallbackNode(
+    node: WorkflowNode,
+    fallbackAgent: AgentAssignment,
+    reason: string,
+    fallbackTrigger: WorkflowFallbackTrigger,
+  ): WorkflowNode {
     const fallbackSuffix = safeNodeId(fallbackAgent.instanceId, 'fallback');
     return {
       ...node,
@@ -1289,9 +1301,12 @@ class WorkflowRunner {
       model: fallbackAgent.model,
       permissionMode: fallbackAgent.permissionMode,
       toolsSettings: fallbackAgent.toolsSettings,
+      fallbackTrigger,
+      fallbackSourceNodeId: node.id,
       prompt: [
         'The previous CLI agent failed on this orchestration step.',
         `Failed step: ${node.agentLabel || node.id}`,
+        `Fallback trigger: ${fallbackTrigger}`,
         `Failure: ${reason}`,
         'Take over the same assignment as the backup CLI. Use the original goal and upstream context.',
         'Do not repeat unrelated work; complete the failed step and report what you did.',
@@ -1301,6 +1316,32 @@ class WorkflowRunner {
     };
   }
 
+  private recordFallbackSkipped(
+    run: WorkflowRun,
+    node: WorkflowNode,
+    reason: string,
+    fallbackTrigger: WorkflowFallbackTrigger,
+    skippedReason: string,
+  ): void {
+    const fallbackSkippedEvents = Array.isArray(run.metadata?.fallbackSkippedEvents)
+      ? run.metadata.fallbackSkippedEvents
+      : [];
+    run.metadata = {
+      ...run.metadata,
+      fallbackSkippedEvents: [
+        ...fallbackSkippedEvents,
+        {
+          nodeId: node.id,
+          trigger: fallbackTrigger,
+          reason,
+          skippedReason,
+          createdAt: Date.now(),
+        },
+      ],
+    };
+    workflowStore.setRun(run);
+  }
+
   private async runFallbackAfterFailure(
     node: WorkflowNode,
     workflow: Workflow,
@@ -1309,9 +1350,29 @@ class WorkflowRunner {
     started: Set<string>,
     completed: Set<string>,
     reason: string,
+    trigger?: WorkflowFallbackTrigger,
   ): Promise<boolean> {
+    const fallbackTrigger = classifyWorkflowFailure(reason, trigger);
     const fallbackAgent = this.fallbackAgentFor(run, node);
     if (!fallbackAgent) {
+      this.recordFallbackSkipped(run, node, reason, fallbackTrigger, 'No fallback agent is configured for this run.');
+      return false;
+    }
+    const decision = resolveWorkflowFallbackDecision({
+      run,
+      node,
+      reason,
+      trigger: fallbackTrigger,
+      fallbackAgentInstanceId: fallbackAgent.instanceId,
+    });
+    if (!decision.shouldFallback) {
+      this.recordFallbackSkipped(
+        run,
+        node,
+        reason,
+        decision.trigger,
+        decision.skippedReason ?? 'Fallback policy skipped this failure.',
+      );
       return false;
     }
     if (workflow.nodes.length + 1 > 64) {
@@ -1323,7 +1384,7 @@ class WorkflowRunner {
       return false;
     }
 
-    let fallbackNode = this.createFallbackNode(node, fallbackAgent, reason);
+    let fallbackNode = this.createFallbackNode(node, fallbackAgent, reason, decision.trigger);
     let collision = 1;
     while (workflow.nodes.some((candidate) => candidate.id === fallbackNode.id)) {
       collision += 1;
@@ -1357,6 +1418,8 @@ class WorkflowRunner {
           nodeId: node.id,
           fallbackNodeId: fallbackNode.id,
           fallbackAgentInstanceId: fallbackAgent.instanceId,
+          trigger: decision.trigger,
+          policy: decision.policy,
           reason,
           startedAt: Date.now(),
         },
@@ -1658,7 +1721,16 @@ class WorkflowRunner {
         workflowStore.setRun(run);
         return;
       }
-      if (await this.runFallbackAfterFailure(node, workflow, run, outputs, started, completed, nodeRun.error)) {
+      if (await this.runFallbackAfterFailure(
+        node,
+        workflow,
+        run,
+        outputs,
+        started,
+        completed,
+        nodeRun.error,
+        'provider_failure',
+      )) {
         return;
       }
       if (node.onFail === 'continue') {
@@ -1710,7 +1782,16 @@ class WorkflowRunner {
         workflowStore.setRun(run);
         return;
       }
-      if (await this.runFallbackAfterFailure(node, workflow, run, outputs, started, completed, nodeRun.error)) {
+      if (await this.runFallbackAfterFailure(
+        node,
+        workflow,
+        run,
+        outputs,
+        started,
+        completed,
+        nodeRun.error,
+        'timeout',
+      )) {
         return;
       }
       if (node.onFail === 'continue') {
@@ -1744,7 +1825,16 @@ class WorkflowRunner {
           nodeRun.status = 'failed';
           nodeRun.error = visibleHandoffError;
           workflowStore.setRun(run);
-          if (await this.runFallbackAfterFailure(node, workflow, run, outputs, started, completed, visibleHandoffError)) {
+          if (await this.runFallbackAfterFailure(
+            node,
+            workflow,
+            run,
+            outputs,
+            started,
+            completed,
+            visibleHandoffError,
+            'invalid_output',
+          )) {
             return;
           }
           if (node.onFail === 'continue') {
@@ -1783,7 +1873,16 @@ class WorkflowRunner {
       workflowStore.setRun(run);
       return;
     }
-    if (await this.runFallbackAfterFailure(node, workflow, run, outputs, started, completed, nodeRun.error)) {
+    if (await this.runFallbackAfterFailure(
+      node,
+      workflow,
+      run,
+      outputs,
+      started,
+      completed,
+      nodeRun.error,
+      classifyWorkflowFailure(`${nodeRun.error}\n${nodeRun.outputText ?? ''}`),
+    )) {
       return;
     }
     if (node.onFail === 'continue') {