fix: stop triage loop at completion-summary node + i18n object-key crashes (#1863) (#1866)

Fixes #1863.

The issue conflated **two distinct v0.52.0 regressions** (the reporter's
i18n root-cause guess was only half of it).

## 1. The real triage loop — engine (primary fix)
The `completion-summary` graph node was added to every built-in workflow
in v0.52.0 (FN-7228/FN-7233), wired with a **success-only edge — no
failure edge**. It's a *best-effort* node
(`ensureWorkflowCompletionSummary` already backfills `task.summary`),
but two failure paths bypass its advisory `!blocking → success`
coercion:

- a **thrown handler exception** (e.g. missing worktree) →
`executeNodeWithRetries` returns `value:"exception"`
- a **failed summary projection write** →
`publishTaskProjectionFromResult` returns `value:"projection-error"`

With no failure edge, either **terminates the graph at
`completion-summary`**. `routeGraphFailureToExecutionResume` then treats
the in-review failure as recoverable and moves the task back to `todo`,
which replays the already-passed nodes (0 new tokens) and fails again —
the **infinite `triage → in-progress → todo` loop** in the report (token
usage 0, execution NOT STARTED).

**Fix:** the graph executor now degrades a `completion-summary` node
failure to `success` so the graph always advances, with a
`routeGraphFailureToExecutionResume` backstop that parks `failed`
instead of looping. Shared `isCompletionSummaryNode` predicate exported
from `@fusion/core`.

## 2. The i18n crash the reporter saw — dashboard
`t("taskDetail.executionMode")` resolves to a **nested object** (the
inline mode-toggle copy), so i18next returns *"key
'taskDetail.executionMode (en)' returned an object instead of string"*
and crashes the Stats tab — the exact error in the report. Surface
enumeration found **two more** callers of the same class:
`routing.source` and `nodes.dockerHost`. Added leaf label keys across
all 6 locales and switched the callers.

## Tests
- `workflow-graph-completion-summary-nonfatal.test.ts` — a failing
summary node (both modes) never terminates/loops the graph, a
non-summary node still fails, and all 5 built-ins have no failure edge.
**Fails without the engine fix.**
- `i18n-string-keys-not-objects.test.ts` — invariant guard scanning
every `t("literal")` caller against the real `en/app.json`. **Catches a
reverted caller.**
- `TaskTokenStatsPanel.test.tsx` — reproduces the exact i18next error
against the real bundle.

## Verification
- Merge gate: **317 engine-core + 63 CI-shape pass**
- Core/engine typecheck clean; i18n parity + typecheck pass; changeset
validates
- All touched dashboard/core/engine suites green

🤖 Generated with [Claude Code](https://claude.com/claude-code)

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **Bug Fixes**
* Fixed an issue where completion-summary failures could cause tasks to
loop indefinitely during resume.
* Made completion-summary behavior best-effort so workflows keep
progressing even when summary handling fails.
* Prevented dashboard crashes by ensuring UI labels use leaf translation
keys (not nested objects).

* **New Features**
* Added missing i18n label keys for Docker host, routing source, and
execution mode across multiple languages.

* **Tests**
* Added regression coverage for completion-summary non-fatal handling
and translation-key validation.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

.changeset/fix-stats-execution-mode-object-key.md

@@ -0,0 +1,7 @@
+---
+"@runfusion/fusion": patch
+---
+
+summary: Fix tasks looping through triage at the completion-summary node, plus Stats/Routing/Node labels crashing.
+category: fix
+dev: Issue #1863 (v0.52.0 regression). (1) The best-effort completion-summary graph node is wired with a success-only edge, so a thrown handler exception or a failed summary projection write terminated the graph and the in-review→todo resume router bounced the task forever. The graph executor now degrades a completion-summary node failure to success (ensureWorkflowCompletionSummary still backfills task.summary), with a routeGraphFailureToExecutionResume backstop. (2) Three views called t() with keys that resolve to nested objects (taskDetail.executionMode, routing.source, nodes.dockerHost); added leaf label keys across all locales and a dashboard invariant test that scans t("literal") callers against the real en/app.json.

packages/core/src/builtin-completion-summary-node.ts

@@ -2,6 +2,21 @@ import type { WorkflowIrNode } from "./workflow-ir-types.js";
 
 export const COMPLETION_SUMMARY_NODE_ID = "completion-summary";
 
+/*
+ * FNXC:WorkflowCompletion 2026-07-01-16:20:
+ * Single source of truth for "is this the advisory completion-summary projection
+ * node?". The engine, graph executor, and lifecycle validation all key summary
+ * behavior off `summaryTarget: "task"` (built-in id `completion-summary`). This
+ * node is BEST-EFFORT: `ensureWorkflowCompletionSummary` deterministically
+ * backfills `task.summary` at the review/done boundary, and every built-in
+ * workflow wires it with a success-only edge (no failure edge — see
+ * `builtin-workflows.ts` `linear()` and the `*-workflow-ir.ts` graphs). So a
+ * summary-node failure must never terminate or loop the graph (issue #1863).
+ */
+export function isCompletionSummaryNode(node: { id?: string; config?: Record<string, unknown> }): boolean {
+  return node.config?.summaryTarget === "task" || node.id === COMPLETION_SUMMARY_NODE_ID;
+}
+
 const COMPLETION_SUMMARY_PROMPT = `Generate the final completion summary for this task.
 
 Use the task description, executed workflow context, changed files/diff, verification notes, and any produced artifacts.

packages/core/src/index.ts

@@ -396,6 +396,11 @@ export {
   isBuiltinWorkflowId,
   isBuiltinWorkflowPluginGated,
 } from "./builtin-workflows.js";
+export {
+  COMPLETION_SUMMARY_NODE_ID,
+  completionSummaryNode,
+  isCompletionSummaryNode,
+} from "./builtin-completion-summary-node.js";
 export {
   resolveWorkflowIrForTask,
   resolveWorkflowIrById,

packages/core/src/workflow-lifecycle-validation.ts

@@ -1,5 +1,6 @@
 import type { WorkflowDefinitionKind } from "./workflow-definition-types.js";
 import type { WorkflowIr, WorkflowIrEdge, WorkflowIrNode } from "./workflow-ir-types.js";
+import { isCompletionSummaryNode } from "./builtin-completion-summary-node.js";
 
 /**
  * FNXC:WorkflowLifecycle 2026-07-01-00:00:
@@ -36,7 +37,7 @@ export interface AnalyzeWorkflowLifecycleOptions {
 }
 
 function isSummaryNode(node: WorkflowIrNode): boolean {
-  return node.config?.summaryTarget === "task" || node.id === "completion-summary";
+  return isCompletionSummaryNode(node);
 }
 
 function isMergeNode(node: WorkflowIrNode): boolean {

packages/dashboard/app/__tests__/i18n-string-keys-not-objects.test.ts

@@ -0,0 +1,72 @@
+import { describe, expect, it } from "vitest";
+import { readFileSync, readdirSync } from "node:fs";
+import { resolve } from "node:path";
+import realEnApp from "../../../i18n/locales/en/app.json";
+
+/*
+ * FNXC:TaskStats 2026-07-01-00:00:
+ * Invariant guard for issue #1863. A dashboard component that calls
+ * `t("some.key")` where `some.key` resolves to a NESTED OBJECT (not a leaf
+ * string) makes i18next return "key 'some.key (en)' returned an object instead
+ * of string" and crashes that view's render. Three surfaces hit this class:
+ * the Task Stats "Execution mode" row, the Routing tab source label, and the
+ * Node Detail Docker host label. The bug is invisible to component tests
+ * because the shared test i18n bundle only carries a handful of keys, so calls
+ * fall through to their inline English defaults instead of resolving the object.
+ *
+ * This test scans the real dashboard source against the real en/app.json so a
+ * new object-key-as-string caller fails here regardless of test-bundle gaps.
+ */
+
+// Dotted paths under the `app` namespace whose value is an object (non-leaf).
+function collectObjectKeyPaths(obj: unknown, prefix: string, out: Set<string>): void {
+  if (obj === null || typeof obj !== "object" || Array.isArray(obj)) return;
+  if (prefix) out.add(prefix);
+  for (const [key, value] of Object.entries(obj as Record<string, unknown>)) {
+    collectObjectKeyPaths(value, prefix ? `${prefix}.${key}` : key, out);
+  }
+}
+
+function collectSourceFiles(dir: string, out: string[]): void {
+  for (const entry of readdirSync(dir, { withFileTypes: true })) {
+    if (entry.name === "node_modules" || entry.name === "__tests__" || entry.name.startsWith(".")) continue;
+    const full = resolve(dir, entry.name);
+    if (entry.isDirectory()) {
+      collectSourceFiles(full, out);
+    } else if (/\.(ts|tsx)$/.test(entry.name) && !/\.test\.(ts|tsx)$/.test(entry.name)) {
+      out.push(full);
+    }
+  }
+}
+
+describe("dashboard i18n string-context callers", () => {
+  it("never calls t() with a key that resolves to a nested object", () => {
+    const objectKeyPaths = new Set<string>();
+    collectObjectKeyPaths(realEnApp, "", objectKeyPaths);
+
+    const appDir = resolve(__dirname, "..");
+    const files: string[] = [];
+    collectSourceFiles(appDir, files);
+
+    // Match `t("dotted.key"` where the leading char is not part of another
+    // identifier (so `getT(`, `handleSort(`, `params.set(` are excluded).
+    const callRe = /(?<![\w$.])t\(\s*["']([\w.]+)["']/g;
+    const violations: string[] = [];
+
+    for (const file of files) {
+      const text = readFileSync(file, "utf8");
+      let match: RegExpExecArray | null;
+      while ((match = callRe.exec(text)) !== null) {
+        const key = match[1];
+        if (!objectKeyPaths.has(key)) continue;
+        // `t(key, opts)` with returnObjects:true intentionally reads the object.
+        const tail = text.slice(match.index, match.index + 200);
+        if (/returnObjects\s*:\s*true/.test(tail)) continue;
+        const line = text.slice(0, match.index).split("\n").length;
+        violations.push(`${file.slice(appDir.length + 1)}:${line} → t("${key}") resolves to an object`);
+      }
+    }
+
+    expect(violations, `Object-valued i18n keys used in string context:\n${violations.join("\n")}`).toEqual([]);
+  });
+});

packages/dashboard/app/components/NodeDetailModal.tsx

@@ -760,7 +760,8 @@ export function NodeDetailModal({
               <div className="node-detail-modal__grid docker-management__info-grid">
                 <div className="node-detail-modal__field"><span>{t("nodes.dockerFieldImage", "Image")}</span><strong><code>{managedDockerNode.imageName}:{managedDockerNode.imageTag}</code></strong></div>
                 <div className="node-detail-modal__field"><span>{t("nodes.dockerContainerId", "Container ID")}</span><strong><code>{managedDockerNode.containerId ? managedDockerNode.containerId.slice(0, 12) : "—"}</code></strong></div>
-                <div className="node-detail-modal__field"><span>{t("nodes.dockerHost", "Host")}</span><strong>{dockerHost}</strong></div>
+                {/* FNXC:Nodes 2026-07-01-00:00: `nodes.dockerHost` is a nested object (local/remote); read the `nodes.dockerHostLabel` leaf so i18next returns a string instead of crashing (issue #1863 bug class). */}
+                <div className="node-detail-modal__field"><span>{t("nodes.dockerHostLabel", "Host")}</span><strong>{dockerHost}</strong></div>
                 <div className="node-detail-modal__field"><span>{t("nodes.dockerPersistentStorage", "Persistent Storage")}</span><strong>{managedDockerNode.persistentStorage ? t("nodes.yes", "Yes") : t("nodes.no", "No")}</strong></div>
                 <div className="node-detail-modal__field"><span>{t("nodes.dockerPort", "Port")}</span><strong>{parsePortFromReachableUrl(managedDockerNode.reachableUrl)}</strong></div>
                 <div className="node-detail-modal__field"><span>{t("nodes.dockerResourceSizing", "Resource Sizing")}</span><strong>{dockerResourceSizing}</strong></div>

packages/dashboard/app/components/RoutingTab.tsx

@@ -142,7 +142,8 @@ export function RoutingTab({ task, settings, addToast, onTaskUpdated }: RoutingT
             </span>
           </div>
           <div className="routing-summary-row" role="listitem">
-            <span className="routing-summary-label">{t("routing.source", "Routing source")}</span>
+            {/* FNXC:Routing 2026-07-01-00:00: `routing.source` is a nested object (noRouting/override/projectDefault); the label must read the `routing.sourceLabel` leaf or i18next returns "returned an object instead of string" and crashes this tab (issue #1863 bug class). */}
+            <span className="routing-summary-label">{t("routing.sourceLabel", "Routing source")}</span>
             <span className="routing-summary-value">{routingSource}</span>
           </div>
           <div className="routing-summary-row" role="listitem">

packages/dashboard/app/components/TaskTokenStatsPanel.tsx

@@ -214,7 +214,11 @@ export function TaskTokenStatsPanel({ tokenUsage, loading, task }: TaskTokenStat
         <h5>{t("taskDetail.executionDetails", "Execution Details")}</h5>
         <dl className="task-token-stats-panel__details">
           <div className="task-token-stats-panel__detail-row">
-            <dt>{t("taskDetail.executionMode", "Execution mode")}</dt>
+            {/*
+            FNXC:TaskStats 2026-07-01-00:00:
+            Use the leaf `taskDetail.executionModeLabel`, not `taskDetail.executionMode`, which is a nested object (ariaLabel/fast/standard/replan* copy for the inline mode toggle). Calling `t()` on the object key makes i18next return "key 'taskDetail.executionMode (en)' returned an object instead of string" and crashes the Stats tab render (issue #1863).
+            */}
+            <dt>{t("taskDetail.executionModeLabel", "Execution mode")}</dt>
             <dd>{task?.executionMode === "fast" ? t("taskDetail.executionModeFast", "Fast") : t("taskDetail.executionModeStandard", "Standard")}</dd>
           </div>
           <div className="task-token-stats-panel__detail-row">

packages/dashboard/app/components/__tests__/TaskTokenStatsPanel.test.tsx

@@ -1,7 +1,10 @@
 import { describe, expect, it, vi } from "vitest";
 import { render, screen } from "@testing-library/react";
+import { createInstance } from "i18next";
+import { I18nextProvider, initReactI18next } from "react-i18next";
 import { TaskTokenStatsPanel } from "../TaskTokenStatsPanel";
 import type { Task } from "@fusion/core";
+import realEnApp from "../../../../i18n/locales/en/app.json";
 
 function makeTask(overrides: Partial<Task> = {}): Task {
   return {
@@ -271,4 +274,47 @@ describe("TaskTokenStatsPanel", () => {
     const metric = screen.getByText("Total execution time").closest(".task-token-stats-panel__metric");
     expect(metric).toHaveTextContent("3m 0s");
   });
+
+  /*
+   * FNXC:TaskStats 2026-07-01-00:00:
+   * Regression for issue #1863. The shared component-test i18n bundle only carries
+   * a handful of `app` keys, so every other key falls through to its inline default
+   * and this panel rendered fine in tests while crashing in production. Load the
+   * REAL en/app.json — where `taskDetail.executionMode` is a nested object (the
+   * inline-toggle copy) — behind an I18nextProvider that mirrors production init
+   * options, and assert the Execution Details label resolves to the leaf string
+   * without i18next's "returned an object instead of string" fallback.
+   */
+  it("renders the execution-mode label against the real locale bundle without an object-key crash", async () => {
+    const instance = createInstance();
+    await instance.use(initReactI18next).init({
+      lng: "en",
+      fallbackLng: "en",
+      ns: ["app"],
+      defaultNS: "app",
+      // Mirror production so a key that resolves to an object surfaces the same
+      // way it does at runtime instead of being silently swallowed.
+      returnNull: false,
+      returnEmptyString: false,
+      returnObjects: false,
+      react: { useSuspense: false },
+      interpolation: { escapeValue: false },
+      resources: { en: { app: realEnApp } },
+    });
+
+    // Sanity-check the fixture still encodes the collision this test guards.
+    expect(typeof (realEnApp as { taskDetail: { executionMode: unknown } }).taskDetail.executionMode).toBe("object");
+
+    render(
+      <I18nextProvider i18n={instance}>
+        <TaskTokenStatsPanel loading={false} tokenUsage={undefined} task={makeTask({ executionMode: "fast" })} />
+      </I18nextProvider>,
+    );
+
+    const label = screen.getByText("Execution mode");
+    expect(label.tagName).toBe("DT");
+    expect(label.textContent).not.toMatch(/returned an object/i);
+    // The value cell resolves the fast/standard leaf, proving the row rendered.
+    expect(label.nextElementSibling?.textContent).toBe("Fast");
+  });
 });

packages/engine/src/__tests__/workflow-graph-completion-summary-nonfatal.test.ts

@@ -0,0 +1,136 @@
+import { describe, expect, it, vi } from "vitest";
+import {
+  BUILTIN_CODING_WORKFLOW_IR,
+  BUILTIN_MARKETING_WORKFLOW_IR,
+  BUILTIN_STEPWISE_CODING_WORKFLOW_IR,
+  BUILTIN_STEPWISE_FINAL_REVIEW_CODING_WORKFLOW_IR,
+  BUILTIN_LEAD_GENERATION_WORKFLOW_IR,
+  COMPLETION_SUMMARY_NODE_ID,
+} from "@fusion/core";
+import type { TaskDetail, WorkflowIr } from "@fusion/core";
+
+import { WorkflowGraphExecutor } from "../workflow-graph-executor.js";
+
+/*
+ * FNXC:WorkflowCompletion 2026-07-01-16:30:
+ * Regression for issue #1863 (v0.52.0 triage loop). The built-in
+ * completion-summary node is best-effort — a deterministic fallback backfills
+ * task.summary — and every built-in workflow wires it with a success-only edge.
+ * A thrown handler exception or a failed summary projection write bypasses the
+ * advisory `!blocking → success` coercion and, absent a failure edge, terminates
+ * the graph at 'completion-summary'; the in-review→todo resume router then bounces
+ * the task back to execution forever. These tests assert the invariant across BOTH
+ * failure modes: a failing completion-summary node never terminates or loops the
+ * graph, while a non-summary node still fails normally.
+ */
+
+const task = { id: "FN-1863" } as TaskDetail;
+
+function settingsOn() {
+  return { experimentalFeatures: { workflowGraphExecutor: true } };
+}
+
+function summaryGraph(): WorkflowIr {
+  // Mirror the built-in wiring: completion-summary has a success-only outgoing
+  // edge and no failure edge.
+  return {
+    version: "v1",
+    name: "summary-nonfatal",
+    nodes: [
+      { id: "start", kind: "start" },
+      { id: COMPLETION_SUMMARY_NODE_ID, kind: "prompt", config: { summaryTarget: "task" } },
+      { id: "end", kind: "end" },
+    ],
+    edges: [
+      { from: "start", to: COMPLETION_SUMMARY_NODE_ID },
+      { from: COMPLETION_SUMMARY_NODE_ID, to: "end", condition: "success" },
+    ],
+  };
+}
+
+describe("completion-summary node is non-fatal (issue #1863)", () => {
+  it("advances past a completion-summary node whose handler throws", async () => {
+    const prompt = vi.fn(async () => {
+      throw new Error("worktree missing during summary");
+    });
+    const executor = new WorkflowGraphExecutor({ handlers: { prompt }, maxRetriesPerNode: 1 });
+
+    const result = await executor.run(task, settingsOn(), summaryGraph());
+
+    // Success (not a terminal graph failure) proves the node was degraded and
+    // the graph advanced instead of terminating at completion-summary. (`end`
+    // nodes are not recorded in visitedNodeIds, so assert on outcome.)
+    expect(result.outcome).toBe("success");
+    expect(result.visitedNodeIds).toContain(COMPLETION_SUMMARY_NODE_ID);
+    expect(prompt).toHaveBeenCalled();
+  });
+
+  it("advances past a completion-summary node whose summary projection write fails", async () => {
+    const prompt = vi.fn(async () => ({
+      outcome: "success" as const,
+      contextPatch: { summary: "Did the thing." },
+    }));
+    const publishTaskProjection = vi.fn(async () => {
+      throw new Error("db write failed");
+    });
+    const executor = new WorkflowGraphExecutor({
+      handlers: { prompt },
+      publishTaskProjection,
+      maxRetriesPerNode: 1,
+    });
+
+    const result = await executor.run(task, settingsOn(), summaryGraph());
+
+    expect(result.outcome).toBe("success");
+    expect(result.visitedNodeIds).toContain(COMPLETION_SUMMARY_NODE_ID);
+    // The write was attempted (and swallowed) — not skipped.
+    expect(publishTaskProjection).toHaveBeenCalled();
+  });
+
+  it("still fails the graph when a NON-summary node throws (degrade is summary-scoped)", async () => {
+    const ir: WorkflowIr = {
+      version: "v1",
+      name: "plain-throws",
+      nodes: [
+        { id: "start", kind: "start" },
+        { id: "work", kind: "prompt" },
+        { id: "end", kind: "end" },
+      ],
+      edges: [
+        { from: "start", to: "work" },
+        { from: "work", to: "end", condition: "success" },
+      ],
+    };
+    const executor = new WorkflowGraphExecutor({
+      handlers: {
+        prompt: async () => {
+          throw new Error("boom");
+        },
+      },
+      maxRetriesPerNode: 1,
+    });
+
+    const result = await executor.run(task, settingsOn(), ir);
+
+    expect(result.outcome).toBe("failure");
+  });
+
+  // Surface enumeration: the fix matters because NONE of the built-in workflows
+  // give completion-summary a failure edge — so any failure there terminates the
+  // graph unless the executor degrades it.
+  const builtins: Array<[string, WorkflowIr]> = [
+    ["coding", BUILTIN_CODING_WORKFLOW_IR],
+    ["stepwise-coding", BUILTIN_STEPWISE_CODING_WORKFLOW_IR],
+    ["stepwise-final-review-coding", BUILTIN_STEPWISE_FINAL_REVIEW_CODING_WORKFLOW_IR],
+    ["marketing", BUILTIN_MARKETING_WORKFLOW_IR],
+    ["lead-generation", BUILTIN_LEAD_GENERATION_WORKFLOW_IR],
+  ];
+
+  it.each(builtins)("built-in %s workflow wires completion-summary with no failure edge", (_name, ir) => {
+    const node = ir.nodes.find((n) => n.id === COMPLETION_SUMMARY_NODE_ID);
+    expect(node, "built-in workflow should contain a completion-summary node").toBeTruthy();
+    const outgoing = ir.edges.filter((e) => e.from === COMPLETION_SUMMARY_NODE_ID);
+    expect(outgoing.length).toBeGreaterThan(0);
+    expect(outgoing.some((e) => e.condition === "failure")).toBe(false);
+  });
+});

packages/engine/src/executor.ts

@@ -11,7 +11,7 @@ import { existsSync, lstatSync, realpathSync } from "node:fs";
 import { readFile, rm, writeFile } from "node:fs/promises";
 import type { TaskStore, Task, TaskDetail, TaskTokenUsage, StepStatus, Settings, WorkflowStep, MissionStore, Slice, AgentState, AgentCapability, RunMutationContext, AgentHeartbeatConfig, Agent, AgentMemoryInclusionMode, ProjectSettings, MergeResult, WorkflowIrNode, WorkflowIrNodeKind, WorkflowStepResult as CoreWorkflowStepResult } from "@fusion/core";
 import { getUnmetSchedulingDependencies } from "./scheduler.js";
-import { RetryStormError, TaskDeletedError, serializeRetryStormError, isExperimentalFeatureEnabled, resolveWorkflowIrForTask, resolveColumnAgentBinding, resolveEffectiveAgent, instanceNodeId, getWorkflowExtensionRegistry, getBuiltinWorkflow, parseNoOpCompletionMarker, allowsAutoMergeProcessing, isSharedBranchGroupMemberIntegration, resolveMaxAutoMergeRetries, resolveOptionalStepRevisionBudget, resolveOptionalReviewRevisionBudget } from "@fusion/core";
+import { RetryStormError, TaskDeletedError, serializeRetryStormError, isExperimentalFeatureEnabled, resolveWorkflowIrForTask, resolveColumnAgentBinding, resolveEffectiveAgent, instanceNodeId, getWorkflowExtensionRegistry, getBuiltinWorkflow, parseNoOpCompletionMarker, allowsAutoMergeProcessing, isSharedBranchGroupMemberIntegration, resolveMaxAutoMergeRetries, resolveOptionalStepRevisionBudget, resolveOptionalReviewRevisionBudget, COMPLETION_SUMMARY_NODE_ID } from "@fusion/core";
 import { finalizeProvenAutoMergeTask } from "./auto-merge-finalization.js";
 import { mergeEffectiveSettings } from "./effective-settings.js";
 import type { TaskStep, WorkflowIr, WorkflowFieldDefinition, WorkflowColumnAgent, EffectiveAgentInput, WorkflowWorkEngineDispatchResult } from "@fusion/core";
@@ -8484,6 +8484,17 @@ export class TaskExecutor {
     if (live.deletedAt) return false;
     if (live.paused || live.userPaused === true) return false;
     if (live.column === "done" || live.column === "archived") return false;
+    /*
+     * FNXC:WorkflowCompletion 2026-07-01-16:26:
+     * Backstop for issue #1863. The advisory completion-summary node must never
+     * drive the in-review→todo resume loop: it has no failure edge, so a failure
+     * here would bounce the task back to execution every run and never stick.
+     * The graph executor now degrades summary-node failures to success, so this
+     * should be unreachable — but if a summary failure ever reaches this router,
+     * let the caller park the task `failed` (a visible terminal state) instead of
+     * looping it forever.
+     */
+    if (failedNode === COMPLETION_SUMMARY_NODE_ID) return false;
     const incompleteSteps = hasNonTerminalWorkflowSteps(live);
     const prematureMergeWithIncompleteSteps = failedNode === "merge" && failureValue === "implementation-incomplete" && incompleteSteps;
     if (live.column !== "in-review" && !(incompleteSteps && live.column === "todo") && !(prematureMergeWithIncompleteSteps && live.column === "in-progress")) return false;

packages/engine/src/workflow-graph-executor.ts

@@ -9,7 +9,7 @@ import type {
   WorkflowNodeExtensionResult,
   WorkflowStepResult,
 } from "@fusion/core";
-import { BUILTIN_CODING_WORKFLOW_IR, PLAN_REVIEW_GROUP_ID, WorkflowIrError, getWorkflowExtensionRegistry, resolveMaxReworkCycles, isExperimentalFeatureEnabled, GRAPH_NATIVE_POST_MERGE_FLAG } from "@fusion/core";
+import { BUILTIN_CODING_WORKFLOW_IR, PLAN_REVIEW_GROUP_ID, WorkflowIrError, getWorkflowExtensionRegistry, resolveMaxReworkCycles, isExperimentalFeatureEnabled, GRAPH_NATIVE_POST_MERGE_FLAG, isCompletionSummaryNode } from "@fusion/core";
 
 import {
   createDefaultNodeHandlers,
@@ -1207,6 +1207,30 @@ export class WorkflowGraphExecutor {
       return this.withEnginePauseAbortContext(node, { outcome: "failure", value: "aborted" });
     }
 
+    /*
+     * FNXC:WorkflowCompletion 2026-07-01-16:24:
+     * A thrown handler exception (missing/pruned worktree, model/provider error,
+     * etc.) bypasses `runGraphCustomNode`'s advisory `!blocking → success`
+     * coercion and lands here as a hard `exception` failure. For the best-effort
+     * completion-summary node that failure has nowhere to go (success-only edge),
+     * so it terminates the graph and loops the in-review task back to todo forever
+     * (issue #1863). Degrade the summary node's exhausted-retry failure to success
+     * — the deterministic `ensureWorkflowCompletionSummary` fallback still fills
+     * `task.summary` — so the graph always advances past it.
+     */
+    if (isCompletionSummaryNode(node)) {
+      const degraded: WorkflowNodeResult = {
+        outcome: "success",
+        value: "summary-unavailable",
+        contextPatch: {
+          [`node:${node.id}:error`]: lastError instanceof Error ? lastError.message : String(lastError),
+        },
+      };
+      if (recordProgress && this.shouldRecordNodeProgress(node)) {
+        await this.recordNodeProgressFinish(task.id, node, null, degraded);
+      }
+      return degraded;
+    }
     const failureResult: WorkflowNodeResult = {
       outcome: "failure",
       value: "exception",
@@ -1327,6 +1351,25 @@ export class WorkflowGraphExecutor {
     try {
       await this.deps.publishTaskProjection?.(taskId, patch, source);
     } catch (error) {
+      /*
+       * FNXC:WorkflowCompletion 2026-07-01-16:24:
+       * The advisory completion-summary node persists its text via a `summary`
+       * projection patch. A failed projection write here must NOT become a
+       * `projection-error` graph failure: that node has no failure edge, so it
+       * would terminate the graph and `routeGraphFailureToExecutionResume` would
+       * bounce the in-review task back to todo forever (issue #1863 v0.52.0
+       * triage loop). Keep the node's success outcome — `ensureWorkflowCompletionSummary`
+       * still backfills `task.summary` deterministically at the review/done boundary.
+       */
+      if (isCompletionSummaryNode(node)) {
+        return {
+          ...result,
+          contextPatch: {
+            ...(result.contextPatch ?? {}),
+            [`node:${node.id}:projectionError`]: error instanceof Error ? error.message : String(error),
+          },
+        };
+      }
       return {
         outcome: "failure",
         value: "projection-error",

packages/i18n/locales/en/app.json

@@ -4163,6 +4163,7 @@
       "local": "Local Docker",
       "remote": "Remote: {{host}}"
     },
+    "dockerHostLabel": "Host",
     "dockerHostConfig": "Host Config",
     "dockerHostPath": "Host path",
     "dockerHostUrl": "Docker host URL",
@@ -5234,6 +5235,7 @@
       "override": "Per-task override",
       "projectDefault": "Project default"
     },
+    "sourceLabel": "Routing source",
     "summarySection": "Routing Summary",
     "title": "Task Routing",
     "unavailablePolicy": "Unavailable-node policy",
@@ -7325,6 +7327,7 @@
       "standard": "Standard",
       "updated": "Execution mode updated to {{mode}}"
     },
+    "executionModeLabel": "Execution mode",
     "executionModeFast": "Fast",
     "executionModeStandard": "Standard",
     "executionStatsAria": "Task execution statistics",

packages/i18n/locales/es/app.json

@@ -4153,6 +4153,7 @@
       "local": "Docker local",
       "remote": "Remoto: {{host}}"
     },
+    "dockerHostLabel": "Host",
     "dockerHostConfig": "Configuración del host",
     "dockerHostPath": "Ruta del host",
     "dockerHostUrl": "URL del host de Docker",
@@ -5224,6 +5225,7 @@
       "override": "Anulación por tarea",
       "projectDefault": "Predeterminado del proyecto"
     },
+    "sourceLabel": "Origen de enrutamiento",
     "summarySection": "Resumen de enrutamiento",
     "title": "Enrutamiento de tareas",
     "unavailablePolicy": "Política de nodo no disponible",
@@ -7315,6 +7317,7 @@
       "replanTitle": "¿Cambiar el modo de ejecución y replanificar?",
       "replanning": "Modo de ejecución actualizado a {{mode}} — {{id}} volvió a Planificación para replanificación"
     },
+    "executionModeLabel": "Modo de ejecución",
     "executionModeFast": "Rápido",
     "executionModeStandard": "Estándar",
     "executionStatsAria": "Estadísticas de ejecución de la tarea",

packages/i18n/locales/fr/app.json

@@ -4153,6 +4153,7 @@
       "local": "Docker local",
       "remote": "Distant : {{host}}"
     },
+    "dockerHostLabel": "Hôte",
     "dockerHostConfig": "Config hôte",
     "dockerHostPath": "Chemin hôte",
     "dockerHostUrl": "URL de l'hôte Docker",
@@ -5224,6 +5225,7 @@
       "override": "Substitution par tâche",
       "projectDefault": "Défaut du projet"
     },
+    "sourceLabel": "Source de routage",
     "summarySection": "Résumé du routage",
     "title": "Routage des tâches",
     "unavailablePolicy": "Politique en cas de nœud indisponible",
@@ -7315,6 +7317,7 @@
       "replanTitle": "Changer le mode d'exécution et replanifier ?",
       "replanning": "Mode d'exécution mis à jour vers {{mode}} — {{id}} est revenue en planification pour replanification"
     },
+    "executionModeLabel": "Mode d'exécution",
     "executionModeFast": "Rapide",
     "executionModeStandard": "Standard",
     "executionStatsAria": "Statistiques d'exécution de la tâche",

packages/i18n/locales/ko/app.json

@@ -4153,6 +4153,7 @@
       "local": "로컬 Docker",
       "remote": "원격: {{host}}"
     },
+    "dockerHostLabel": "호스트",
     "dockerHostConfig": "호스트 구성",
     "dockerHostPath": "호스트 경로",
     "dockerHostUrl": "Docker 호스트 URL",
@@ -5224,6 +5225,7 @@
       "override": "작업별 재정의",
       "projectDefault": "프로젝트 기본값"
     },
+    "sourceLabel": "라우팅 소스",
     "summarySection": "라우팅 요약",
     "title": "작업 라우팅",
     "unavailablePolicy": "사용 불가 노드 정책",
@@ -7315,6 +7317,7 @@
       "replanTitle": "실행 모드를 변경하고 다시 계획할까요?",
       "replanning": "실행 모드가 {{mode}}로 업데이트되었습니다 — {{id}}가 재계획을 위해 계획으로 돌아갔습니다"
     },
+    "executionModeLabel": "실행 모드",
     "executionModeFast": "빠름",
     "executionModeStandard": "표준",
     "executionStatsAria": "작업 실행 통계",

packages/i18n/locales/zh-CN/app.json

@@ -4153,6 +4153,7 @@
       "local": "本地 Docker",
       "remote": "远程: {{host}}"
     },
+    "dockerHostLabel": "主机",
     "dockerHostConfig": "主机配置",
     "dockerHostPath": "主机路径",
     "dockerHostUrl": "Docker 主机地址",
@@ -5224,6 +5225,7 @@
       "override": "每个任务的覆盖",
       "projectDefault": "项目默认值"
     },
+    "sourceLabel": "路由来源",
     "summarySection": "路由摘要",
     "title": "任务路由",
     "unavailablePolicy": "不可用节点策略",
@@ -7315,6 +7317,7 @@
       "replanTitle": "更改执行模式并重新规划？",
       "replanning": "执行模式已更新为 {{mode}} — {{id}} 已返回规划以重新规划"
     },
+    "executionModeLabel": "执行模式",
     "executionModeFast": "快速",
     "executionModeStandard": "标准",
     "executionStatsAria": "任务执行统计",

packages/i18n/locales/zh-TW/app.json

@@ -4153,6 +4153,7 @@
       "local": "本地 Docker",
       "remote": "遠端: {{host}}"
     },
+    "dockerHostLabel": "主機",
     "dockerHostConfig": "主機設定",
     "dockerHostPath": "主機路徑",
     "dockerHostUrl": "Docker 主機網址",
@@ -5224,6 +5225,7 @@
       "override": "每個工作的覆寫",
       "projectDefault": "專案預設值"
     },
+    "sourceLabel": "路由來源",
     "summarySection": "路由摘要",
     "title": "任務路由",
     "unavailablePolicy": "不可用節點政策",
@@ -7315,6 +7317,7 @@
       "replanTitle": "變更執行模式並重新規劃？",
       "replanning": "執行模式已更新為 {{mode}} — {{id}} 已返回規劃以重新規劃"
     },
+    "executionModeLabel": "執行模式",
     "executionModeFast": "快速",
     "executionModeStandard": "標準",
     "executionStatsAria": "任務執行統計",

packages/i18n/src/resources.d.ts

@@ -4164,6 +4164,7 @@ export default interface Resources {
         "local": "Local Docker",
         "remote": "Remote: {{host}}"
       },
+      "dockerHostLabel": "Host",
       "dockerHostConfig": "Host Config",
       "dockerHostPath": "Host path",
       "dockerHostUrl": "Docker host URL",
@@ -5217,6 +5218,7 @@ export default interface Resources {
         "override": "Per-task override",
         "projectDefault": "Project default"
       },
+      "sourceLabel": "Routing source",
       "summarySection": "Routing Summary",
       "title": "Task Routing",
       "unavailablePolicy": "Unavailable-node policy",
@@ -7351,6 +7353,7 @@ export default interface Resources {
         "standard": "Standard",
         "updated": "Execution mode updated to {{mode}}"
       },
+      "executionModeLabel": "Execution mode",
       "executionModeFast": "Fast",
       "executionModeStandard": "Standard",
       "executionStatsAria": "Task execution statistics",

packages/pi-claude-cli/index.ts

@@ -5,7 +5,16 @@
  * subprocess using stream-json NDJSON protocol.
  */
 
-import { getModels } from "@earendil-works/pi-ai";
+/*
+ * FNXC:ModelCatalog 2026-07-01-17:30:
+ * pi-ai 0.80 restructured its static catalog API: the top-level `getModels`
+ * export moved to the deprecated `/compat` shim, and the canonical accessor is
+ * `getBuiltinModels` from `@earendil-works/pi-ai/providers/all`. pi-claude-cli
+ * now pins `@earendil-works/pi-ai` and `@earendil-works/pi-coding-agent` to
+ * `^0.80.3` (matching cli/engine) so the whole extension resolves one pi-ai
+ * version and the ExtensionAPI stream types stay compatible.
+ */
+import { getBuiltinModels } from "@earendil-works/pi-ai/providers/all";
 import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
 import { streamViaCli } from "./src/provider.js";
 import { streamViaAcp } from "./src/acp-driver.js";
@@ -175,7 +184,7 @@ export default function (pi: ExtensionAPI) {
     // `claude` subprocess in streamViaCli still reports hard errors on send.
     void runCliValidationOnce();
 
-    const catalogModels = getModels("anthropic").map((model) => ({
+    const catalogModels = getBuiltinModels("anthropic").map((model) => ({
       id: model.id,
       name: model.name,
       reasoning: model.reasoning,

packages/pi-claude-cli/package.json

@@ -27,6 +27,8 @@
     "@earendil-works/pi-coding-agent": "^0.80.3"
   },
   "devDependencies": {
+    "@earendil-works/pi-ai": "^0.80.3",
+    "@earendil-works/pi-coding-agent": "^0.80.3",
     "@types/node": "^22.0.0",
     "typescript": "^5.7.0",
     "vitest": "^4.1.0"

packages/pi-claude-cli/src/__tests__/provider.test.ts

@@ -59,13 +59,18 @@ const { MockAssistantMessageEventStream } = vi.hoisted(() => {
 });
 
 vi.mock("@earendil-works/pi-ai", () => ({
-  getModels: vi.fn(() => mockModels),
   AssistantMessageEventStream: MockAssistantMessageEventStream,
   calculateCost: vi.fn(),
 }));
 
+// pi-ai 0.80 moved the static catalog read to `getBuiltinModels` in the
+// `/providers/all` subpath (see index.ts). Mock it there.
+vi.mock("@earendil-works/pi-ai/providers/all", () => ({
+  getBuiltinModels: vi.fn(() => mockModels),
+}));
+
 import { spawn } from "node:child_process";
-import { getModels } from "@earendil-works/pi-ai";
+import { getBuiltinModels } from "@earendil-works/pi-ai/providers/all";
 import { streamViaCli } from "../provider";
 
 describe("provider registration (default export)", () => {
@@ -148,7 +153,7 @@ describe("provider registration (default export)", () => {
   it("deduplicates extra models when catalog already includes them", async () => {
     const registerProvider = vi.fn();
     const mockPi = { registerProvider, on: vi.fn() } as any;
-    const getModelsMock = vi.mocked(getModels);
+    const getModelsMock = vi.mocked(getBuiltinModels);
     const upstreamSonnet5 = {
       id: "claude-sonnet-5",
       name: "Claude Sonnet 5 Upstream",

pnpm-lock.yaml

@@ -609,13 +609,13 @@ importers:
       '@agentclientprotocol/sdk':
         specifier: 0.24.0
         version: 0.24.0(zod@4.3.6)
+    devDependencies:
       '@earendil-works/pi-ai':
         specifier: ^0.80.3
         version: 0.80.3(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)
       '@earendil-works/pi-coding-agent':
         specifier: ^0.80.3
         version: 0.80.3(@modelcontextprotocol/sdk@1.28.0(zod@4.3.6))(ws@8.20.0)(zod@4.3.6)
-    devDependencies:
       '@types/node':
         specifier: ^25.5.2
         version: 25.5.2