diff --git a/panel/src/app/(dashboard)/prompter/page.tsx b/panel/src/app/(dashboard)/prompter/page.tsx
index 2b4beaa1..94d1f92c 100644
--- a/panel/src/app/(dashboard)/prompter/page.tsx
+++ b/panel/src/app/(dashboard)/prompter/page.tsx
@@ -4,10 +4,12 @@ import { useEffect, useRef } from "react";
 import { Loader2, Sparkles, X } from "lucide-react";
 import { Button } from "@/components/ui/button";
 import { usePrompter } from "@/hooks/use-prompter";
+import { Team } from "@/types";
 import {
   ChatMessages,
   ChatComposer,
   SuccessCard,
+  BoardReviewSentCard,
   IntakeForm,
   BatchReviewCard,
 } from "@/components/prompter";
@@ -42,7 +44,7 @@ export default function PrompterPage() {
     batch,
     batchWaves,
     batchResult,
-    updateBatchDraftProject,
+    setBatchDraftProjects,
     confirmBatch,
   } = usePrompter();
 
@@ -113,22 +115,40 @@ export default function PrompterPage() {
           createdTaskTeam ? (
             <div className="flex flex-1 flex-col items-center justify-center px-8 py-8">
               <div className="w-full max-w-md space-y-3">
-                <SuccessCard
-                  taskId={createdTaskId}
-                  taskTitle={createdTaskTitle}
-                  team={createdTaskTeam}
-                  onStartAnother={startAnother}
-                />
-                {batchResult && (
-                  <p className="text-center text-xs text-muted-foreground">
-                    {batchResult.root_subtask_ids.length} tasks sequenced into{" "}
-                    {batchResult.waves.length} wave
-                    {batchResult.waves.length === 1 ? "" : "s"}.
-                    {batchResult.warnings.length > 0 &&
-                      ` ${batchResult.warnings.length} advisory note${
-                        batchResult.warnings.length === 1 ? "" : "s"
-                      }.`}
-                  </p>
+                {/* Board-routed MegaTask: created HELD for PO+HoM review, not
+                    dispatched — the CEO releases it with Approve & Start on the
+                    umbrella task. Every other path is a real "created/launched"
+                    success. ``createdTaskTeam === BOARD`` is set only by the
+                    batch board route (the single-draft board route parks and
+                    never reaches success). */}
+                {createdTaskTeam === Team.BOARD && batchResult ? (
+                  <BoardReviewSentCard
+                    taskId={createdTaskId}
+                    taskTitle={createdTaskTitle}
+                    rootSubtaskCount={batchResult.root_subtask_ids.length}
+                    waveCount={batchResult.waves.length}
+                    onStartAnother={startAnother}
+                  />
+                ) : (
+                  <>
+                    <SuccessCard
+                      taskId={createdTaskId}
+                      taskTitle={createdTaskTitle}
+                      team={createdTaskTeam}
+                      onStartAnother={startAnother}
+                    />
+                    {batchResult && (
+                      <p className="text-center text-xs text-muted-foreground">
+                        {batchResult.root_subtask_ids.length} tasks sequenced
+                        into {batchResult.waves.length} wave
+                        {batchResult.waves.length === 1 ? "" : "s"}.
+                        {batchResult.warnings.length > 0 &&
+                          ` ${batchResult.warnings.length} advisory note${
+                            batchResult.warnings.length === 1 ? "" : "s"
+                          }.`}
+                      </p>
+                    )}
+                  </>
                 )}
               </div>
             </div>
@@ -149,7 +169,7 @@ export default function PrompterPage() {
                 waves={batchWaves}
                 projectIds={projectIds}
                 onKeepChatting={keepChatting}
-                onProjectChange={updateBatchDraftProject}
+                onSetProjects={setBatchDraftProjects}
                 onConfirm={confirmBatch}
                 isLaunching={isLaunching}
               />
diff --git a/panel/src/components/prompter/batch-review-card.tsx b/panel/src/components/prompter/batch-review-card.tsx
index ebd4e325..570e08a7 100644
--- a/panel/src/components/prompter/batch-review-card.tsx
+++ b/panel/src/components/prompter/batch-review-card.tsx
@@ -11,15 +11,10 @@ import {
 import { Button } from "@/components/ui/button";
 import { Card, CardContent, CardHeader, CardTitle } from "@/components/ui/card";
 import { Badge } from "@/components/ui/badge";
-import {
-  Select,
-  SelectContent,
-  SelectItem,
-  SelectTrigger,
-  SelectValue,
-} from "@/components/ui/select";
+import { Checkbox } from "@/components/ui/checkbox";
 import { useProjects } from "@/hooks/use-projects";
 import type { BatchProposal, StartRoute } from "@/hooks/use-prompter";
+import type { ProjectSummary } from "@/types";
 import type { CellWork, DraftProposal } from "@/lib/api/prompter";
 import { Team } from "@/types";
 
@@ -32,38 +27,40 @@ const CELL_LABEL: Record<string, string> = {
   ux_ui: "UX/UI",
 };
 
-/** One the_work entry whose team is a delivery cell — a per-cell project picker. */
-interface CellEntry {
-  entry: CellWork;
-  entryIndex: number;
-  team: Team;
-}
-
-/** A draft's per-cell entries (the the_work slots that carry a cell team), in
- *  the_work order. Empty for a legacy single-cell draft with no cell the_work. */
-function cellEntries(draft: DraftProposal): CellEntry[] {
-  return (draft.the_work ?? [])
-    .map((entry, entryIndex) => ({ entry, entryIndex, team: entry?.team }))
+/** The project_ids a draft currently targets: the per-cell ``the_work[].project_id``
+ *  set (the multi-select model), falling back to a legacy top-level project_id.
+ *  Only scoped ids count — an out-of-scope id is treated as unselected. */
+function selectedProjectIds(
+  draft: DraftProposal,
+  scoped: Set<string>,
+): string[] {
+  const work = Array.isArray(draft.the_work) ? draft.the_work : [];
+  const pids = work
     .filter(
-      (e): e is CellEntry =>
-        !!e.team && (CELL_TEAMS as readonly string[]).includes(e.team),
-    );
+      (w): w is CellWork & { project_id: string } =>
+        !!w?.team &&
+        (CELL_TEAMS as readonly string[]).includes(w.team) &&
+        typeof w.project_id === "string" &&
+        w.project_id !== "" &&
+        scoped.has(w.project_id),
+    )
+    .map((w) => w.project_id);
+  if (pids.length > 0) return pids;
+  if (draft.project_id && scoped.has(draft.project_id))
+    return [draft.project_id];
+  return [];
 }
 
 interface BatchReviewCardProps {
   batch: BatchProposal;
   /** The conflict-free waves (lists of draft indices), once previewed. */
   waves: number[][] | null;
-  /** The repos this MegaTask is scoped to — each cell must target one of them. */
+  /** The repos this MegaTask is scoped to — each task targets a subset of them. */
   projectIds: string[];
   onKeepChatting: () => void;
-  /** `entryIndex` is the the_work slot (the cell); -1 for a legacy single-cell
-   *  draft with no per-cell map (sets the top-level project_id). */
-  onProjectChange: (
-    index: number,
-    entryIndex: number,
-    projectId: string,
-  ) => void;
+  /** Set the whole set of projects one task targets (multi-select across cells,
+   *  one repo per cell — the backend stores one project per cell). */
+  onSetProjects: (index: number, ids: string[]) => void;
   onConfirm: (route: StartRoute) => void;
   /** A launch is in flight — disable the actions so a double-click can't dupe. */
   isLaunching?: boolean;
@@ -71,38 +68,55 @@ interface BatchReviewCardProps {
 
 /**
  * The MegaTask review card: every task the agent proposed in one batch, each
- * with its per-cell target projects (editable) and collision surface, plus the
- * conflict-free wave plan. A multi-cell task (be+fe, fe+uxui) shows one project
- * picker per cell, scoped to that cell's repos — a RoboCo project is per-cell,
- * so each cell lands in its own repo. The human reviews the whole batch and the
- * sequencing, fixes any cell in the wrong repo, then picks one start path.
+ * with its target projects (a multi-select checkbox list — one task can span
+ * several repos, one repo per delivery cell) and collision surface, plus the
+ * conflict-free wave plan. The human reviews the whole batch and the sequencing,
+ * picks the repos each task lands in, then picks one start path.
  */
 export function BatchReviewCard({
   batch,
   waves,
   projectIds,
   onKeepChatting,
-  onProjectChange,
+  onSetProjects,
   onConfirm,
   isLaunching = false,
 }: BatchReviewCardProps) {
   const { data: allProjects = [] } = useProjects();
   // Only the scoped repos are valid targets (the agent read only those).
   const scoped = new Set(projectIds);
+  const scopedByCell = (cell: Team): ProjectSummary[] =>
+    allProjects.filter((p) => scoped.has(p.id) && p.assigned_cell === cell);
   const titleOf = (i: number): string =>
     batch.drafts[i]?.title ?? `Task ${i + 1}`;
-  // A task is mis-targeted when any of its cells lacks a scoped project (a
-  // multi-cell draft checks every the_work entry; a legacy single-cell draft
-  // with no cell map checks its top-level project_id).
-  const missingProject = batch.drafts.some((d) => {
-    const entries = cellEntries(d);
-    if (entries.length > 0) {
-      return entries.some(
-        (ce) => !ce.entry.project_id || !scoped.has(ce.entry.project_id),
+  // A task is mis-targeted when it has no project selected at all (the backend
+  // re-asserts each targeted project is in scope and the batch spans ≥2 repos).
+  const missingProject = batch.drafts.some(
+    (d) => selectedProjectIds(d, scoped).length === 0,
+  );
+
+  /** Toggle one project in a task's selection. A RoboCo project is per-cell and
+   *  the backend stores one project per cell, so checking a repo in a cell that
+   *  already has a different repo checked swaps it (unchecks the sibling). */
+  const toggle = (index: number, projectId: string) => {
+    const draft = batch.drafts[index];
+    const current = selectedProjectIds(draft, scoped);
+    const proj = allProjects.find((p) => p.id === projectId);
+    const cell = proj?.assigned_cell;
+    if (current.includes(projectId)) {
+      onSetProjects(
+        index,
+        current.filter((id) => id !== projectId),
       );
+      return;
     }
-    return !d.project_id || !scoped.has(d.project_id);
-  });
+    // Checking: drop any other project in the same cell (one repo per cell).
+    const next = current.filter((id) => {
+      const other = allProjects.find((p) => p.id === id);
+      return other?.assigned_cell !== cell;
+    });
+    onSetProjects(index, [...next, projectId]);
+  };
 
   return (
     <Card className="border-primary/40 bg-primary/5">
@@ -124,14 +138,22 @@ export function BatchReviewCard({
       <CardContent className="space-y-3 pb-3">
         <ol className="space-y-2">
           {batch.drafts.map((draft, i) => {
-            const entries = cellEntries(draft);
+            const selected = selectedProjectIds(draft, scoped);
             return (
               <li
                 key={i}
                 className="rounded-md border bg-background/60 px-3 py-2 text-sm"
               >
                 <div className="flex items-start justify-between gap-2">
-                  <span className="font-medium leading-tight">
+                  {/* Title is clamped to a fixed 2-line space so a long title
+                      can't grow the row (or, as a long unbroken token, blow the
+                      card width out and wreck the layout). min-w-0 lets the flex
+                      item shrink below min-content; break-words stops a token
+                      from overflowing; the full title is on the tooltip. */}
+                  <span
+                    className="min-w-0 flex-1 break-words font-medium leading-tight line-clamp-2"
+                    title={`${i + 1}. ${draft.title}`}
+                  >
                     {i + 1}. {draft.title}
                   </span>
                   <div className="flex shrink-0 items-center gap-1">
@@ -154,87 +176,48 @@ export function BatchReviewCard({
                     {draft.objective || draft.description}
                   </p>
                 )}
-                {entries.length > 0 ? (
-                  /* Per-cell project picker — one Select per the_work entry,
-                     scoped to that cell's repos (a project is per-cell). */
-                  <div className="mt-1.5 space-y-1">
-                    {entries.map(({ entry, entryIndex, team }) => {
-                      const cellProjects = allProjects.filter(
-                        (p) => scoped.has(p.id) && p.assigned_cell === team,
-                      );
-                      const pid = entry.project_id ?? "";
-                      const ok = pid !== "" && scoped.has(pid);
-                      return (
-                        <div
-                          key={entryIndex}
-                          className="flex items-center gap-2"
-                        >
-                          <span className="w-16 shrink-0 text-xs text-muted-foreground">
-                            {CELL_LABEL[team] ?? team}
-                          </span>
-                          <Select
-                            value={ok ? pid : ""}
-                            onValueChange={(v) =>
-                              onProjectChange(i, entryIndex, v)
-                            }
-                            disabled={isLaunching}
-                          >
-                            <SelectTrigger
-                              className={`h-7 flex-1 text-xs ${
-                                ok ? "" : "border-destructive"
-                              }`}
-                            >
-                              <SelectValue placeholder="Pick a project…" />
-                            </SelectTrigger>
-                            <SelectContent>
-                              {cellProjects.map((p) => (
-                                <SelectItem key={p.id} value={p.id}>
-                                  {p.name}
-                                </SelectItem>
-                              ))}
-                            </SelectContent>
-                          </Select>
+                {/* Multi-select project picker — one task can span several repos
+                    (one per delivery cell). Grouped by cell; one repo per cell. */}
+                <div className="mt-1.5 space-y-1.5">
+                  <p
+                    className={`text-xs ${
+                      selected.length === 0
+                        ? "text-destructive"
+                        : "text-muted-foreground"
+                    }`}
+                  >
+                    Projects {selected.length === 0 && "— pick at least one"}
+                  </p>
+                  {CELL_TEAMS.map((cell) => {
+                    const repos = scopedByCell(cell);
+                    if (repos.length === 0) return null;
+                    return (
+                      <div key={cell} className="space-y-1">
+                        <span className="text-[10px] font-medium uppercase tracking-wide text-muted-foreground">
+                          {CELL_LABEL[cell] ?? cell}
+                        </span>
+                        <div className="flex flex-wrap gap-x-3 gap-y-1">
+                          {repos.map((p) => {
+                            const checked = selected.includes(p.id);
+                            return (
+                              <label
+                                key={p.id}
+                                className="flex cursor-pointer items-center gap-1.5 text-xs disabled:cursor-not-allowed"
+                              >
+                                <Checkbox
+                                  checked={checked}
+                                  disabled={isLaunching}
+                                  onCheckedChange={() => toggle(i, p.id)}
+                                />
+                                <span>{p.name}</span>
+                              </label>
+                            );
+                          })}
                         </div>
-                      );
-                    })}
-                  </div>
-                ) : (
-                  /* Legacy single-cell draft (no per-cell the_work) — one Select
-                     bound to the top-level project_id, scoped to all repos. */
-                  <div className="mt-1.5 flex items-center gap-2">
-                    <span className="text-xs text-muted-foreground">
-                      Project
-                    </span>
-                    <Select
-                      value={
-                        draft.project_id && scoped.has(draft.project_id)
-                          ? draft.project_id
-                          : ""
-                      }
-                      onValueChange={(v) => onProjectChange(i, -1, v)}
-                      disabled={isLaunching}
-                    >
-                      <SelectTrigger
-                        className={`h-7 flex-1 text-xs ${
-                          draft.project_id && scoped.has(draft.project_id)
-                            ? ""
-                            : "border-destructive"
-                        }`}
-                      >
-                        <SelectValue placeholder="Pick a project…" />
-                      </SelectTrigger>
-                      <SelectContent>
-                        {allProjects
-                          .filter((p) => scoped.has(p.id))
-                          .map((p) => (
-                            <SelectItem key={p.id} value={p.id}>
-                              {p.name}
-                            </SelectItem>
-                          ))}
-                      </SelectContent>
-                    </Select>
-                  </div>
-                )}
+                      </div>
+                    );
+                  })}
+                </div>
               </li>
             );
           })}
@@ -248,7 +231,11 @@ export function BatchReviewCard({
             </p>
             <ol className="space-y-0.5">
               {waves.map((wave, w) => (
-                <li key={w} className="text-xs">
+                <li
+                  key={w}
+                  className="break-words text-xs line-clamp-2"
+                  title={wave.map((i) => titleOf(i)).join(", ")}
+                >
                   <span className="font-medium">Wave {w + 1}:</span>{" "}
                   {wave.map((i) => titleOf(i)).join(", ")}
                 </li>
@@ -259,7 +246,7 @@ export function BatchReviewCard({
 
         {missingProject && (
           <p className="text-xs text-destructive">
-            Pick a project for every cell of every task before launching the
+            Pick at least one project for every task before launching the
             MegaTask.
           </p>
         )}
diff --git a/panel/src/components/prompter/board-review-sent-card.tsx b/panel/src/components/prompter/board-review-sent-card.tsx
new file mode 100644
index 00000000..59b4618b
--- /dev/null
+++ b/panel/src/components/prompter/board-review-sent-card.tsx
@@ -0,0 +1,91 @@
+"use client";
+
+import Link from "next/link";
+import { Users, ExternalLink, RefreshCw } from "lucide-react";
+import { Button } from "@/components/ui/button";
+import {
+  Card,
+  CardContent,
+  CardFooter,
+  CardHeader,
+  CardTitle,
+} from "@/components/ui/card";
+import { Badge } from "@/components/ui/badge";
+
+interface BoardReviewSentCardProps {
+  /** The umbrella task id — the single board-review / CEO-approve unit. */
+  taskId: string;
+  taskTitle: string;
+  rootSubtaskCount: number;
+  waveCount: number;
+  onStartAnother: () => void;
+}
+
+/**
+ * The MegaTask "Board review & Start" confirmation: the umbrella + root-subtasks
+ * were created HELD (umbrella assigned to the Product Owner, root-subtasks in
+ * BACKLOG) for the Product Owner + Head of Marketing to review. Nothing is
+ * dispatched yet — the CEO releases the sequenced tasks with Approve & Start on
+ * the umbrella task once the board finishes. This is the batch analogue of the
+ * single-draft board route's "sent to the board" wait, reusing the existing CEO
+ * Approve & Start gate on the umbrella (which fires ``approve_and_start`` →
+ * ``_activate_batch_root_subtasks``).
+ */
+export function BoardReviewSentCard({
+  taskId,
+  taskTitle,
+  rootSubtaskCount,
+  waveCount,
+  onStartAnother,
+}: BoardReviewSentCardProps) {
+  return (
+    <Card className="border-primary/30 bg-primary/5">
+      <CardHeader className="pb-2">
+        <div className="flex items-center gap-2">
+          <Users className="h-5 w-5 text-primary" />
+          <CardTitle className="text-sm font-semibold text-primary">
+            Sent to the Board for review
+          </CardTitle>
+        </div>
+      </CardHeader>
+
+      <CardContent className="pb-3 space-y-2">
+        <p className="text-sm font-medium">{taskTitle}</p>
+        <div className="flex items-center gap-2">
+          <Badge variant="secondary" className="text-xs">
+            {rootSubtaskCount} task{rootSubtaskCount === 1 ? "" : "s"}
+          </Badge>
+          <Badge variant="outline" className="text-xs">
+            {waveCount} wave{waveCount === 1 ? "" : "s"}
+          </Badge>
+          <span className="text-xs text-muted-foreground">
+            ID: {taskId.slice(0, 8)}…
+          </span>
+        </div>
+        <p className="text-xs text-muted-foreground">
+          The Product Owner and Head of Marketing are reviewing this MegaTask.
+          Nothing is dispatched yet. Once they finish, open the umbrella task
+          and use <span className="font-medium">Approve &amp; Start</span> to
+          release the sequenced tasks. You can leave and come back.
+        </p>
+      </CardContent>
+
+      <CardFooter className="gap-2 pt-0">
+        <Button variant="outline" size="sm" asChild className="flex-1">
+          <Link
+            href={`/tasks/${taskId}`}
+            target="_blank"
+            rel="noopener noreferrer"
+          >
+            <ExternalLink className="mr-1.5 h-3.5 w-3.5" />
+            View umbrella task
+          </Link>
+        </Button>
+        <Button size="sm" className="flex-1" onClick={onStartAnother}>
+          <RefreshCw className="mr-1.5 h-3.5 w-3.5" />
+          Start another
+        </Button>
+      </CardFooter>
+    </Card>
+  );
+}
diff --git a/panel/src/components/prompter/draft-proposal-card.tsx b/panel/src/components/prompter/draft-proposal-card.tsx
index 1a75c690..d1577f3d 100644
--- a/panel/src/components/prompter/draft-proposal-card.tsx
+++ b/panel/src/components/prompter/draft-proposal-card.tsx
@@ -45,7 +45,7 @@ function draftToText(draft: DraftProposal): string {
       "",
     );
   }
-  if (draft.the_work?.length) {
+  if (Array.isArray(draft.the_work) && draft.the_work.length) {
     lines.push("## The Work");
     for (const cell of draft.the_work) {
       lines.push(`### ${cellLabel(cell.team)}`, cell.summary);
@@ -73,7 +73,7 @@ export function DraftProposalCard({
   isLaunching = false,
 }: DraftProposalCardProps) {
   const priorityLabel = PRIORITY_LABELS[draft.priority ?? 2] ?? "Medium";
-  const cells = draft.the_work ?? [];
+  const cells = Array.isArray(draft.the_work) ? draft.the_work : [];
   // Distinct cells only: the_work has one entry per work item, so a cell with
   // several items would otherwise show its badge repeated (Backend Backend …).
   const distinctTeams = Array.from(new Set(cells.map((c) => c.team)));
diff --git a/panel/src/components/prompter/index.ts b/panel/src/components/prompter/index.ts
index 55d6b98b..5b1ac1c4 100644
--- a/panel/src/components/prompter/index.ts
+++ b/panel/src/components/prompter/index.ts
@@ -3,4 +3,5 @@ export { ChatComposer } from "./chat-composer";
 export { DraftProposalCard } from "./draft-proposal-card";
 export { BatchReviewCard } from "./batch-review-card";
 export { SuccessCard } from "./success-card";
+export { BoardReviewSentCard } from "./board-review-sent-card";
 export { IntakeForm } from "./intake-form";
diff --git a/panel/src/components/tasks/task-table.tsx b/panel/src/components/tasks/task-table.tsx
index b378f5cd..5532eb92 100644
--- a/panel/src/components/tasks/task-table.tsx
+++ b/panel/src/components/tasks/task-table.tsx
@@ -544,9 +544,9 @@ export function TaskTable({
                     )}
                     onClick={handleRowClick}
                   >
-                    <TableCell>
+                    <TableCell className="max-w-[22rem]">
                       <div
-                        className="flex items-center gap-1"
+                        className="flex items-center gap-1 min-w-0"
                         style={{ paddingLeft: `${node.depth * 1.5}rem` }}
                       >
                         {hasChildren ? (
@@ -569,8 +569,10 @@ export function TaskTable({
                           href={"/tasks/" + task.id}
                           className="block hover:underline min-w-0"
                         >
-                          <div className="font-medium flex items-center gap-2">
-                            <span className="truncate">{task.title}</span>
+                          <div className="font-medium flex items-center gap-2 min-w-0">
+                            <span className="truncate" title={task.title}>
+                              {task.title}
+                            </span>
                             {task.batch_id && !task.parent_task_id && (
                               <Badge
                                 variant="outline"
diff --git a/panel/src/hooks/__tests__/use-prompter.test.ts b/panel/src/hooks/__tests__/use-prompter.test.ts
new file mode 100644
index 00000000..9d94609e
--- /dev/null
+++ b/panel/src/hooks/__tests__/use-prompter.test.ts
@@ -0,0 +1,186 @@
+import { describe, expect, it } from "vitest";
+import type { CellWork, DraftProposal } from "@/lib/api/prompter";
+import { Team } from "@/types";
+import {
+  fillBatchProjects,
+  rebuildCellWork,
+  type BatchProposal,
+} from "@/hooks/use-prompter";
+
+const BE = "be-repo";
+const FE = "fe-repo";
+const BE2 = "be-core";
+
+function proj(id: string, cell: Team) {
+  return { id, name: id, assigned_cell: cell };
+}
+
+function draft(partial: Partial<DraftProposal>): DraftProposal {
+  return {
+    title: "T",
+    description: "objective objective objective",
+    acceptance_criteria: ["a"],
+    team: Team.BACKEND,
+    ...partial,
+  } as DraftProposal;
+}
+
+function mk(drafts: DraftProposal[]): BatchProposal {
+  return { title: "batch", drafts, dropped: 0 };
+}
+
+describe("fillBatchProjects", () => {
+  it("fills a single-cell draft from the top-level project_id when it matches the cell", () => {
+    const batch = mk([
+      draft({ the_work: [{ team: Team.BACKEND, summary: "s", items: [] }] }),
+    ]);
+    const out = fillBatchProjects(batch, [BE], [proj(BE, Team.BACKEND)]);
+    expect(out.drafts[0].the_work![0].project_id).toBe(BE);
+  });
+
+  it("falls back to the single scoped repo for the cell when no project_id is set", () => {
+    const batch = mk([
+      draft({ the_work: [{ team: Team.FRONTEND, summary: "s", items: [] }] }),
+    ]);
+    const out = fillBatchProjects(
+      batch,
+      [BE, FE],
+      [proj(BE, Team.BACKEND), proj(FE, Team.FRONTEND)],
+    );
+    expect(out.drafts[0].the_work![0].project_id).toBe(FE);
+  });
+
+  it("leaves a cell empty when 2+ scoped repos belong to that cell (real ambiguity)", () => {
+    const batch = mk([
+      draft({ the_work: [{ team: Team.BACKEND, summary: "s", items: [] }] }),
+    ]);
+    const out = fillBatchProjects(
+      batch,
+      [BE, BE2],
+      [proj(BE, Team.BACKEND), proj(BE2, Team.BACKEND)],
+    );
+    expect(out.drafts[0].the_work![0].project_id).toBeFalsy();
+  });
+
+  it("does not overwrite an explicit per-cell project_id", () => {
+    const batch = mk([
+      draft({
+        the_work: [
+          { team: Team.BACKEND, summary: "s", items: [], project_id: BE2 },
+        ],
+      }),
+    ]);
+    const out = fillBatchProjects(
+      batch,
+      [BE, BE2],
+      [proj(BE, Team.BACKEND), proj(BE2, Team.BACKEND)],
+    );
+    expect(out.drafts[0].the_work![0].project_id).toBe(BE2);
+  });
+
+  it("does not trust a top-level project_id whose cell mismatches the work entry", () => {
+    const batch = mk([
+      draft({
+        project_id: FE, // frontend repo, but the work says backend
+        the_work: [{ team: Team.BACKEND, summary: "s", items: [] }],
+      }),
+    ]);
+    const out = fillBatchProjects(
+      batch,
+      [FE, BE],
+      [proj(FE, Team.FRONTEND), proj(BE, Team.BACKEND)],
+    );
+    expect(out.drafts[0].the_work![0].project_id).toBe(BE);
+  });
+
+  it("is idempotent (returns the same batch reference on a second pass)", () => {
+    const batch = mk([
+      draft({ the_work: [{ team: Team.BACKEND, summary: "s", items: [] }] }),
+    ]);
+    const once = fillBatchProjects(batch, [BE], [proj(BE, Team.BACKEND)]);
+    const twice = fillBatchProjects(once, [BE], [proj(BE, Team.BACKEND)]);
+    expect(twice).toBe(once);
+  });
+
+  it("legacy draft (no the_work) keeps a scoped top-level project_id as-is", () => {
+    const batch = mk([draft({ the_work: [], project_id: BE })]);
+    const out = fillBatchProjects(batch, [BE], [proj(BE, Team.BACKEND)]);
+    expect(out).toBe(batch); // nothing to fill → same reference
+  });
+});
+
+describe("rebuildCellWork", () => {
+  // The multi-select review card: the human picks a SET of repos for a task
+  // (one repo per delivery cell). rebuildCellWork turns that set into the
+  // the_work[] cell map the backend stores (task_cell_projects is unique per
+  // (task, team) — one repo per cell).
+
+  const projects = [
+    proj(BE, Team.BACKEND),
+    proj(BE2, Team.BACKEND),
+    proj(FE, Team.FRONTEND),
+    proj("ux-repo", Team.UX_UI),
+  ];
+
+  it("maps one selected repo per cell and clears the top-level project_id", () => {
+    const out = rebuildCellWork([BE, FE], projects, []);
+    expect(out.project_id).toBeNull();
+    const teams = out.the_work.map((w) => w.team);
+    expect(teams).toEqual([Team.BACKEND, Team.FRONTEND]);
+    expect(out.the_work[0].project_id).toBe(BE);
+    expect(out.the_work[1].project_id).toBe(FE);
+  });
+
+  it("keeps only the first selected repo when two are picked for the same cell (one repo per cell)", () => {
+    const out = rebuildCellWork([BE, BE2], projects, []);
+    expect(out.the_work).toHaveLength(1);
+    expect(out.the_work[0].team).toBe(Team.BACKEND);
+    expect(out.the_work[0].project_id).toBe(BE); // first wins
+  });
+
+  it("preserves an existing cell entry's summary/items when the cell stays selected", () => {
+    const current: CellWork[] = [
+      {
+        team: Team.BACKEND,
+        summary: "build API",
+        items: ["x", "y"],
+        project_id: BE2,
+      },
+    ];
+    const out = rebuildCellWork([BE], projects, current);
+    expect(out.the_work).toHaveLength(1);
+    expect(out.the_work[0].summary).toBe("build API");
+    expect(out.the_work[0].items).toEqual(["x", "y"]);
+    expect(out.the_work[0].project_id).toBe(BE); // repo swapped, summary kept
+  });
+
+  it("drops a cell entry when its cell is no longer selected", () => {
+    const current: CellWork[] = [
+      { team: Team.BACKEND, summary: "s", items: [] },
+      { team: Team.FRONTEND, summary: "s", items: [] },
+    ];
+    const out = rebuildCellWork([FE], projects, current);
+    expect(out.the_work.map((w) => w.team)).toEqual([Team.FRONTEND]);
+  });
+
+  it("appends a minimal entry for a newly-selected cell not in the_work", () => {
+    const out = rebuildCellWork([FE, "ux-repo"], projects, [
+      { team: Team.BACKEND, summary: "s", items: [] },
+    ]);
+    const teams = out.the_work.map((w) => w.team);
+    expect(teams).toEqual([Team.FRONTEND, Team.UX_UI]); // backend dropped, fe+ux added
+    expect(out.the_work[0]).toMatchObject({
+      team: Team.FRONTEND,
+      summary: "",
+      items: [],
+    });
+  });
+
+  it("empty selection empties the_work (the task then has no project → blocked)", () => {
+    const out = rebuildCellWork([], projects, [
+      { team: Team.BACKEND, summary: "s", items: [] },
+    ]);
+    expect(out.the_work).toEqual([]);
+    expect(out.project_id).toBeNull();
+  });
+});
diff --git a/panel/src/hooks/use-prompter.ts b/panel/src/hooks/use-prompter.ts
index 792693b7..fd3f19a4 100644
--- a/panel/src/hooks/use-prompter.ts
+++ b/panel/src/hooks/use-prompter.ts
@@ -16,6 +16,7 @@ import {
 } from "@/lib/api/prompter";
 import { getErrorMessage } from "@/lib/api/client";
 import { tasksApi } from "@/lib/api/tasks";
+import { useProjects } from "@/hooks/use-projects";
 import { Team } from "@/types";
 import type { TaskType, TaskNature, Complexity } from "@/types";
 
@@ -68,7 +69,8 @@ const CELL_TEAMS: Team[] = [Team.BACKEND, Team.FRONTEND, Team.UX_UI];
  *  is a delivery cell and that carries a project_id. Empty for a legacy
  *  single-cell draft that uses a top-level project_id instead. */
 function draftCellProjectIds(draft: DraftProposal): string[] {
-  return (draft.the_work ?? [])
+  const work = Array.isArray(draft.the_work) ? draft.the_work : [];
+  return work
     .filter(
       (w): w is CellWork & { project_id: string } =>
         !!w?.team &&
@@ -173,6 +175,17 @@ function draftFromEvent(data: Record<string, unknown> | undefined): {
   return { draft: d as unknown as DraftProposal, scale };
 }
 
+/** Coerce a value into a `CellWork[]`: a bare object → one-element array, a
+ *  non-array non-object → empty. The backend coerces `the_work` before it
+ *  reaches SSE, but a stale localStorage payload or a malformed frame could
+ *  still carry a non-array, and the batch card does `the_work.map(...)` — so
+ *  normalize here rather than crash. */
+function asCellWork(value: unknown): CellWork[] {
+  if (Array.isArray(value)) return value as CellWork[];
+  if (value && typeof value === "object") return [value as CellWork];
+  return [];
+}
+
 /** Pull a MegaTask ({title, drafts[]}) out of a `batch` SSE event's payload. */
 function batchFromEvent(
   data: Record<string, unknown> | undefined,
@@ -180,10 +193,15 @@ function batchFromEvent(
   if (!data || typeof data !== "object") return null;
   const raw = (data as Record<string, unknown>).drafts;
   if (!Array.isArray(raw)) return null;
-  const drafts = raw.filter(
-    (x): x is DraftProposal =>
-      !!x && typeof (x as DraftProposal).title === "string",
-  );
+  const drafts = raw
+    .filter(
+      (x): x is Record<string, unknown> =>
+        !!x && typeof (x as DraftProposal).title === "string",
+    )
+    .map((d) => ({
+      ...(d as unknown as DraftProposal),
+      the_work: asCellWork((d as Record<string, unknown>).the_work),
+    }));
   if (drafts.length === 0) return null;
   const title = (data as Record<string, unknown>).title;
   // Prefer the backend's dropped count; else compute from what we filtered, so a
@@ -196,6 +214,128 @@ function batchFromEvent(
   return { title: typeof title === "string" ? title : "", drafts, dropped };
 }
 
+/** Rebuild a draft's ``the_work`` from the set of projects the human selected
+ *  for it (the multi-select review card: one task can span several repos, one
+ *  repo per delivery cell — the backend's ``task_cell_projects`` is unique per
+ *  ``(task, team)``). Existing cell entries keep their summary/items; a newly-
+ *  selected cell gets a minimal entry; a deselected cell's entry is dropped
+ *  (that cell no longer participates). The top-level ``project_id`` is cleared
+ *  — a multi-repo task targets via its cell map, not a top-level repo. Pure, so
+ *  ``setBatchDraftProjects`` can stay a thin setState wrapper and this is
+ *  unit-tested directly. */
+export function rebuildCellWork(
+  ids: string[],
+  allProjects: { id: string; assigned_cell?: Team | "" }[],
+  currentWork: CellWork[],
+): { the_work: CellWork[]; project_id: null } {
+  const cellToPid = new Map<Team, string>();
+  for (const id of ids) {
+    const proj = allProjects.find((p) => p.id === id);
+    const cell = proj?.assigned_cell;
+    if (cell && !cellToPid.has(cell)) cellToPid.set(cell, id);
+  }
+  const covered = new Set<Team>();
+  // Update existing cell entries in place, preserve summary/items.
+  const updated = currentWork.map((w) => {
+    const team = w?.team;
+    if (team && cellToPid.has(team)) {
+      covered.add(team);
+      return { ...w, project_id: cellToPid.get(team)! };
+    }
+    return w;
+  });
+  // Append a minimal entry for a newly-selected cell not already in the_work.
+  const appended: CellWork[] = [];
+  for (const [team, pid] of cellToPid) {
+    if (!covered.has(team)) {
+      appended.push({ team, summary: "", items: [], project_id: pid });
+    }
+  }
+  // Drop entries for cells no longer selected.
+  const the_work = [...updated, ...appended].filter((w) => {
+    const team = w?.team;
+    return team && cellToPid.has(team);
+  });
+  return { the_work, project_id: null };
+}
+
+/** Fill each draft's missing project assignment from the MegaTask's scoped
+ *  repos, so the review card is pre-filled instead of forcing the human to
+ *  re-pick the projects they already scoped at intake. The intake prompt tells
+ *  the agent to set a top-level ``project_id`` (each task lives in one repo),
+ *  and the backend's scope validator falls back to it — so this mirrors that:
+ *  an explicit per-cell ``the_work[].project_id`` wins; else the draft's
+ *  top-level ``project_id`` (when scoped); else the single scoped repo that
+ *  belongs to this cell, when unambiguous. An empty field stays empty (real
+ *  ambiguity — 2+ scoped repos for the cell — the human picks). Returns the
+ *  same batch reference when nothing changed (idempotent, no render loop). */
+export function fillBatchProjects(
+  batch: BatchProposal,
+  projectIds: string[],
+  allProjects: { id: string; assigned_cell?: Team | "" }[],
+): BatchProposal {
+  const scoped = new Set(projectIds);
+  const scopedProjects = allProjects.filter((p) => scoped.has(p.id));
+  const byCell = (team: Team) =>
+    scopedProjects.filter((p) => p.assigned_cell === team);
+  const isCell = (team: unknown): team is Team =>
+    typeof team === "string" &&
+    (CELL_TEAMS as readonly string[]).includes(team as Team);
+
+  let changed = false;
+  const drafts = batch.drafts.map((d) => {
+    const work = Array.isArray(d.the_work) ? d.the_work : [];
+    const cellWork = work.filter((w) => isCell(w?.team));
+
+    // Draft with at least one the_work cell entry: fill each cell's project_id.
+    if (cellWork.length > 0) {
+      let workChanged = false;
+      const newWork = work.map((w) => {
+        if (
+          !w ||
+          !isCell(w.team) ||
+          (w.project_id && scoped.has(w.project_id))
+        ) {
+          return w;
+        }
+        // (1) the draft's top-level project_id (the agent's main assignment),
+        // but only when that repo actually belongs to this cell (a single-cell
+        // draft's repo should match its one cell; a mismatch is an agent error
+        // — fall through to auto-assign rather than land the wrong repo).
+        const top = d.project_id;
+        if (top && scoped.has(top)) {
+          const proj = scopedProjects.find((p) => p.id === top);
+          if (proj && proj.assigned_cell === w.team) {
+            workChanged = true;
+            return { ...w, project_id: top };
+          }
+        }
+        // (2) exactly one scoped repo belongs to this cell.
+        const matching = byCell(w.team);
+        if (matching.length === 1) {
+          workChanged = true;
+          return { ...w, project_id: matching[0].id };
+        }
+        return w;
+      });
+      if (!workChanged) return d;
+      changed = true;
+      return { ...d, the_work: newWork };
+    }
+
+    // Legacy draft (no cell the_work): one top-level project_id. Fill it only
+    // when exactly one scoped repo exists (truly unambiguous).
+    if (d.project_id && scoped.has(d.project_id)) return d;
+    if (scopedProjects.length === 1) {
+      changed = true;
+      return { ...d, project_id: scopedProjects[0].id };
+    }
+    return d;
+  });
+
+  return changed ? { ...batch, drafts } : batch;
+}
+
 // ---------------------------------------------------------------------------
 // Refresh durability
 //
@@ -271,6 +411,10 @@ export function usePrompter() {
   const [messages, setMessages] = useState<ChatMessage[]>([]);
   const [sessionId, setSessionId] = useState<string | null>(null);
   const [isSending, setIsSending] = useState(false);
+  // Every project (carries assigned_cell), for pre-filling a MegaTask batch's
+  // per-cell project_ids from the scoped repos (useProjects dedups with the
+  // batch card's own query by key).
+  const { data: allProjects = [] } = useProjects();
   const [isLaunching, setIsLaunching] = useState(false);
   /** The latest tool the agent is using — "watch it work" status line. */
   const [activity, setActivity] = useState<string | null>(null);
@@ -366,7 +510,15 @@ export function usePrompter() {
           if (evt.text) {
             setActivity(null); // first text clears the "preparing…" indicator
             appendDelta(evt.text);
-            setState("streaming");
+            // Trailing prose AFTER a draft/batch tool call (the agent keeps
+            // talking once the card is up) must still render in the bubble —
+            // but it must NOT clobber the preview state, or the card vanishes
+            // mid-turn and turn_end can't restore it (it only preserves these
+            // states if they're still set). Keep the card up; the delta lands
+            // in a fresh bubble (the batch/draft handler cleared streamingId).
+            setState((s) =>
+              s === "draft_preview" || s === "batch_preview" ? s : "streaming",
+            );
           }
           break;
         case "tool_use":
@@ -508,6 +660,17 @@ export function usePrompter() {
     }
   }, [sessionId, messages, state, editableDraft, batch, batchWaves]);
 
+  // Pre-fill each MegaTask draft's project from the scoped repos (the agent
+  // sets a top-level project_id; the backend falls back to it). Without this
+  // the review card shows empty Selects and blocks launch, forcing the human
+  // to re-pick the very projects they scoped at intake. Idempotent: once every
+  // cell has a project_id it returns the same batch reference (no loop).
+  useEffect(() => {
+    if (!batch) return;
+    const filled = fillBatchProjects(batch, projectIds, allProjects);
+    if (filled !== batch) setBatch(filled);
+  }, [batch, projectIds, allProjects]);
+
   // On mount, reconnect to a still-running session left behind by a reload.
   const didRestoreRef = useRef(false);
   useEffect(() => {
@@ -804,29 +967,33 @@ export function usePrompter() {
   // Confirm a MegaTask — create the umbrella + sequenced root-subtasks, reap
   // -----------------------------------------------------------------------
 
-  /** Reassign one cell of one task in the proposed MegaTask to a different
-   *  project. `entryIndex` is the the_work slot (the cell); pass -1 for a legacy
-   *  single-cell draft that has no per-cell map (sets the top-level project_id).
-   *  Project does not affect the wave plan (waves derive from collision surface),
-   *  so the previewed waves stay valid. */
-  const updateBatchDraftProject = useCallback(
-    (index: number, entryIndex: number, projectId: string) => {
+  /** Set the projects one MegaTask task targets. The review card exposes a
+   *  multi-select checkbox list per task (one task can span several repos), so
+   *  the human picks the whole set at once instead of one dropdown per cell.
+   *  A RoboCo project is per-cell, so the selection maps to one project per
+   *  cell in ``the_work[]`` (the backend's ``task_cell_projects`` is unique per
+   *  ``(task, team)`` — one repo per cell). Existing entries keep their
+   *  summary/items; a newly-selected cell gets a minimal entry; a deselected
+   *  cell's entry is dropped (that cell no longer participates). The top-level
+   *  ``project_id`` is cleared — a multi-repo task targets via its cell map.
+   *  Project choice does not affect the wave plan (waves derive from collision
+   *  surface), so the previewed waves stay valid. */
+  const setBatchDraftProjects = useCallback(
+    (index: number, ids: string[]) => {
       setBatch((prev) => {
         if (!prev) return prev;
         return {
           ...prev,
           drafts: prev.drafts.map((d, i) => {
             if (i !== index) return d;
-            if (entryIndex < 0) return { ...d, project_id: projectId };
-            const the_work = (d.the_work ?? []).map((w, wi) =>
-              wi === entryIndex ? { ...w, project_id: projectId } : w,
-            );
-            return { ...d, the_work };
+            const work = Array.isArray(d.the_work) ? d.the_work : [];
+            const rebuilt = rebuildCellWork(ids, allProjects, work);
+            return { ...d, the_work: rebuilt.the_work, project_id: null };
           }),
         };
       });
     },
-    [],
+    [allProjects],
   );
 
   const confirmBatch = useCallback(
@@ -891,10 +1058,24 @@ export function usePrompter() {
         setCreatedTaskId(result.umbrella_task_id);
         setCreatedTaskTitle(batch.title.trim() || "MegaTask");
         setCreatedTaskTeam(route === "board" ? Team.BOARD : Team.MAIN_PM);
-        toast.success(
-          `MegaTask launched — ${result.root_subtask_ids.length} tasks in ` +
-            `${result.waves.length} wave${result.waves.length === 1 ? "" : "s"}.`,
-        );
+        if (route === "board") {
+          // Board route: the umbrella + root-subtasks are created HELD for the
+          // PO + HoM to review — nothing is dispatched yet. The CEO releases the
+          // sequenced tasks with Approve & Start on the umbrella task once the
+          // board finishes (the existing CEO gate, not this chat). Say so, so
+          // "Board review & Start" doesn't read as "launched" the way the
+          // Main-PM route does.
+          toast.success(
+            "Sent to the Board for review — the Product Owner and Head of " +
+              "Marketing will review this MegaTask. Approve & Start it from the " +
+              "umbrella task once they're done.",
+          );
+        } else {
+          toast.success(
+            `MegaTask launched — ${result.root_subtask_ids.length} tasks in ` +
+              `${result.waves.length} wave${result.waves.length === 1 ? "" : "s"}.`,
+          );
+        }
         setState("success");
       } catch (err) {
         toast.error(`Failed to launch MegaTask: ${getErrorMessage(err)}`);
@@ -978,7 +1159,7 @@ export function usePrompter() {
     batch,
     batchWaves,
     batchResult,
-    updateBatchDraftProject,
+    setBatchDraftProjects,
     confirmBatch,
   };
 }
diff --git a/roboco/agent_sdk/intake_driver.py b/roboco/agent_sdk/intake_driver.py
index 24c78892..7d21785b 100644
--- a/roboco/agent_sdk/intake_driver.py
+++ b/roboco/agent_sdk/intake_driver.py
@@ -56,19 +56,69 @@ class StreamChunk:
     data: dict[str, Any] = field(default_factory=dict)
 
 
+def _coerce_to_list(value: Any) -> list[Any]:
+    """Wrap a lone scalar/dict into a one-element list; drop junk to ``[]``.
+
+    Mirrors ``content.validators.coerce_to_list`` but always returns a list
+    (never ``None``) and never passes a non-list, non-scalar/dict through — the
+    panel treats these fields as arrays and would crash on anything else. A
+    bare string/dict is the well-intentioned single-item case (wrap it); a
+    number/bool/None is not a work unit, so it is dropped.
+    """
+    if isinstance(value, list):
+        return value
+    if isinstance(value, str | dict):
+        return [value]
+    return []
+
+
+# Fields that are lists of plain strings (vs ``the_work``, a list of work-unit
+# dicts). The agent may emit these as XML-ish ``<item>…</item>`` elements, which
+# the SDK parses into ``[{"item": {"$text": "…"}}, …]`` — coerce to flat str
+# lists so they reach the panel and a VARCHAR[] column as strings, not dicts.
+_STR_LIST_FIELDS = ("acceptance_criteria", "what_this_builds", "notes")
+
+
 def _coerce_draft(data: Any) -> dict[str, Any] | None:
     """Return ``data`` as a draft dict (with a string ``title``), else ``None``.
 
-    Accepts a dict, or a JSON string the agent may have passed.
+    Accepts a dict, or a JSON string the agent may have passed. Coerces the
+    list-shaped spec fields: ``the_work`` (a list of work-unit dicts) is wrapped
+    to a list, and each ``the_work`` entry's ``items`` plus the string-list
+    fields (``acceptance_criteria``, ``what_this_builds``, ``notes``) are
+    flattened to ``list[str]`` — the agent sometimes emits these as XML-ish
+    ``<item>…</item>`` elements that the SDK parses into dict wrappers, which
+    would crash a ``VARCHAR[]`` insert and dump ``str(dict)`` into the rendered
+    description. The panel renders ``(draft.the_work ?? []).map(...)`` and
+    throws if ``the_work`` is a bare object, so this is the single choke point
+    that keeps a non-array from ever reaching SSE.
     """
+    from roboco.foundation.policy.content.validators import coerce_str_list
+
     if isinstance(data, str):
         try:
             data = json.loads(data)
         except (ValueError, TypeError):
             return None
-    if isinstance(data, dict) and isinstance(data.get("title"), str):
-        return data
-    return None
+    if not (isinstance(data, dict) and isinstance(data.get("title"), str)):
+        return None
+    coerced = dict(data)
+    if "the_work" in coerced:
+        coerced["the_work"] = _coerce_to_list(coerced["the_work"])
+    for key in _STR_LIST_FIELDS:
+        if key in coerced:
+            coerced[key] = coerce_str_list(coerced[key])
+    work = coerced.get("the_work")
+    if isinstance(work, list):
+        coerced["the_work"] = [
+            {
+                **unit,
+                "items": coerce_str_list(unit.get("items")),
+            }
+            for unit in work
+            if isinstance(unit, dict)
+        ]
+    return coerced
 
 
 def _extract_draft(text: str) -> dict[str, Any] | None:
diff --git a/roboco/api/routes/v1/do.py b/roboco/api/routes/v1/do.py
index 549a2837..a7f038e7 100644
--- a/roboco/api/routes/v1/do.py
+++ b/roboco/api/routes/v1/do.py
@@ -77,6 +77,9 @@ async def do_note(
             "next_steps": body.next_steps,
         },
         section=body.section,
+        done=body.done,
+        next=body.next,
+        where_to_look=body.where_to_look,
     )
     return envelope_to_response(env, request)
 
diff --git a/roboco/api/schemas/v1/do.py b/roboco/api/schemas/v1/do.py
index 97052235..aa658b51 100644
--- a/roboco/api/schemas/v1/do.py
+++ b/roboco/api/schemas/v1/do.py
@@ -72,6 +72,18 @@ class NoteRequest(BaseModel):
     # (auditor). Validated by the content model server-side. Omit it to write a
     # developer summary straight from ``text``.
     section: dict[str, Any] | None = None
+    # handoff scope — the PM/coordinator RESUMPTION fields, promoted to
+    # top-level typed strings so the tool schema declares them machine-visible.
+    # ``section: dict[str, Any]`` renders a schema with no visible sub-fields,
+    # so a weak model (minimax-m3) emits ``section={}`` and the resumption gate
+    # rejects ``done — Field required`` — the 2026-06-27 PM respawn-loop
+    # meltdown. The same model fills top-level ``string`` decision fields fine
+    # (proven live), so ``done``/``next`` follow that precedent (typed
+    # ``str = ""`` → schema declares ``string`` not ``anyOf[string, null]``).
+    # Filled into ``section`` server-side; ignored for non-handoff scopes.
+    done: str = ""
+    next: str = ""
+    where_to_look: list[str] | None = None
 
     # List-typed fields tolerate a lone scalar: a single string (or, for
     # ``options``, a single dict) is wrapped into a one-element list before
diff --git a/roboco/api/schemas/v1/flow.py b/roboco/api/schemas/v1/flow.py
index 17d48824..6b4cbd32 100644
--- a/roboco/api/schemas/v1/flow.py
+++ b/roboco/api/schemas/v1/flow.py
@@ -1,9 +1,23 @@
 """Request schemas for /api/v1/flow/* intent verbs."""
 
-from typing import Any
+from typing import Annotated, Any
 from uuid import UUID
 
-from pydantic import BaseModel, Field, field_validator
+from pydantic import BaseModel, BeforeValidator, Field, field_validator
+
+from roboco.foundation.policy.content.validators import coerce_str_list
+
+# A ``list[str]`` field that tolerates the Claude SDK's XML-ish tool-input
+# parsing: an LLM emitting a bullet list as ``<item>…</item>`` elements arrives
+# nested (``[[["…"]]]`` / ``[{"item": {"$text": "…"}}, …]``), which a bare
+# ``list[str]`` hard-rejects at validation time (the live ``i_will_plan`` crash:
+# ``technical_considerations.1 Input should be a valid string``). The
+# ``BeforeValidator`` flattens it to a flat ``list[str]`` first — same
+# ``coerce_str_list`` used at the intake→DB boundary (Bug 3, MegaTask memory).
+# Applied to EVERY LLM-authored list-of-strings on the flow surface so the SDK
+# can't crash any of them (acceptance_criteria, issues, files, ac_verdicts,
+# technical_considerations, covers_parent_criteria).
+StrList = Annotated[list[str], BeforeValidator(coerce_str_list)]
 
 
 class GiveMeWorkRequest(BaseModel):
@@ -28,7 +42,7 @@ class IWillWorkOnRequest(BaseModel):
     # so re-entry/recovery calls that omit them still pass route validation;
     # depth + presence are enforced on FRESH dev claims by
     # choreographer._dev_plan_gate. The dev's `plan` doubles as the approach.
-    technical_considerations: list[str] = Field(default_factory=list)
+    technical_considerations: StrList = Field(default_factory=list)
     risks: list[dict[str, str]] = Field(default_factory=list)
     open_questions: list[dict[str, str | bool]] = Field(default_factory=list)
 
@@ -111,7 +125,7 @@ class ClaimReviewRequest(BaseModel):
 class PassReviewRequest(BaseModel):
     task_id: UUID
     notes: str = Field(..., min_length=1)
-    ac_verdicts: list[str] | None = Field(
+    ac_verdicts: StrList | None = Field(
         default=None,
         description=(
             "One verification entry per acceptance criterion (in criterion "
@@ -123,7 +137,7 @@ class PassReviewRequest(BaseModel):
 
 class FailReviewRequest(BaseModel):
     task_id: UUID
-    issues: list[str] = Field(..., min_length=1)
+    issues: StrList = Field(..., min_length=1)
 
 
 class ClaimPrReviewRequest(BaseModel):
@@ -162,7 +176,7 @@ class PrPassRequest(BaseModel):
 
 class PrFailRequest(BaseModel):
     task_id: UUID
-    issues: list[str] = Field(..., min_length=1)
+    issues: StrList = Field(..., min_length=1)
 
 
 class ClaimDocTaskRequest(BaseModel):
@@ -172,7 +186,7 @@ class ClaimDocTaskRequest(BaseModel):
 class IDocumentedRequest(BaseModel):
     task_id: UUID
     notes: str = Field(..., min_length=1)
-    files: list[str] = Field(..., min_length=1)
+    files: StrList = Field(..., min_length=1)
 
 
 class TriageRequest(BaseModel):
@@ -222,7 +236,7 @@ class IWillPlanRequest(BaseModel):
         default_factory=list,
         description="List of {title, description} — server assigns id + order",
     )
-    technical_considerations: list[str] = Field(default_factory=list)
+    technical_considerations: StrList = Field(default_factory=list)
     risks: list[dict[str, str]] = Field(default_factory=list)
     open_questions: list[dict[str, str | bool]] = Field(default_factory=list)
 
@@ -253,14 +267,14 @@ class DelegateRequest(BaseModel):
     estimated_complexity: str = Field(..., min_length=1)
     # acceptance_criteria is required and non-empty; downstream policy
     # also denylist-checks each item against placeholder phrases.
-    acceptance_criteria: list[str] = Field(..., min_length=1)
+    acceptance_criteria: StrList = Field(..., min_length=1)
     # Optional per-subtask project override. When omitted, the choreographer
     # resolves the project from the parent's Product map for this cell, then
     # falls back to the parent's project. Plain optional field — no validator.
     project_id: UUID | None = None
     # Parent acceptance-criterion ids this subtask is responsible for. Lets the
     # coverage + roll-up AC gates verify every parent AC is claimed and satisfied.
-    covers_parent_criteria: list[str] | None = None
+    covers_parent_criteria: StrList | None = None
 
     # Pre-gateway parity: cross-field validators that catch the most common
     # LLM-vs-schema confusions. Pre-gateway lived in
diff --git a/roboco/foundation/policy/content/validators.py b/roboco/foundation/policy/content/validators.py
index c18c7429..f47b2177 100644
--- a/roboco/foundation/policy/content/validators.py
+++ b/roboco/foundation/policy/content/validators.py
@@ -101,3 +101,55 @@ def coerce_to_list(value: Any) -> Any:
     if isinstance(value, str | dict):
         return [value]
     return value
+
+
+# Keys an LLM's tool input may wrap a piece of text under. The Claude SDK parses
+# XML-ish mixed content the model sometimes emits for a list-of-strings field
+# (e.g. ``<item>…</item>``) into ``{"item": {"$text": "…"}}``; ``$text`` is the
+# SDK's marker for element text content. Ordered: most specific first.
+_TEXT_KEYS: tuple[str, ...] = ("$text", "text", "item", "value", "content", "summary")
+
+
+def _extract_strs(item: Any) -> list[str]:
+    """Pull every string out of one list element, recursing through wrappers."""
+    if isinstance(item, str):
+        stripped = item.strip()
+        return [stripped] if stripped else []
+    if isinstance(item, dict):
+        for key in _TEXT_KEYS:
+            if key in item:
+                return _extract_strs(item[key])
+        # No recognized key: keep any bare string values the dict carries.
+        return [
+            str(v).strip() for v in item.values() if isinstance(v, str) and v.strip()
+        ]
+    if isinstance(item, list):
+        out: list[str] = []
+        for sub in item:
+            out.extend(_extract_strs(sub))
+        return out
+    return []
+
+
+def coerce_str_list(value: Any) -> list[str]:
+    """Coerce a list-of-strings field to a flat ``list[str]``.
+
+    An LLM may emit a ``list[str]`` field (``acceptance_criteria``, ``notes``,
+    ``what_this_builds``, a work unit's ``items``) as XML-ish ``<item>…</item>``
+    elements, which the Claude SDK parses into ``[{"item": {"$text": "…"}}, …]``.
+    A bare dict/scalar is the single-item case. This recurses through the
+    wrappers and returns only strings — never a dict — so the value is safe for a
+    ``VARCHAR[]`` column (asyncpg rejects non-str elements with a DataError) and
+    for markdown rendering (``str(dict)`` would otherwise dump ``"{'item': …}"``).
+    Anything that cannot be reduced to a string is dropped.
+    """
+    if value is None:
+        return []
+    if isinstance(value, str | dict):
+        value = [value]
+    if not isinstance(value, list):
+        return []
+    out: list[str] = []
+    for item in value:
+        out.extend(_extract_strs(item))
+    return out
diff --git a/roboco/mcp/do_server.py b/roboco/mcp/do_server.py
index d2e1c357..ce9ce1ae 100644
--- a/roboco/mcp/do_server.py
+++ b/roboco/mcp/do_server.py
@@ -196,6 +196,9 @@ def note(
     what_struggled: str = "",
     next_steps: list[str] | str | None = None,
     section: dict[str, Any] | None = None,
+    done: str = "",
+    next: str = "",
+    where_to_look: list[str] | None = None,
 ) -> dict[str, Any]:
     """Write a journal entry, or (scope='handoff') your note SECTION.
 
@@ -220,9 +223,13 @@ def note(
     Other journal scopes (note / learning / struggle) just need ``text``.
 
     scope='handoff' writes your dedicated SECTION (dev_notes / quick_context /
-    auditor_notes) instead of the journal: pass ``section={...}`` with the
-    section's fields (PM/resumption needs done+next; auditor needs
-    summary+severity), or just ``text`` for a developer summary.
+    auditor_notes) instead of the journal. For a PM/coordinator RESUMPTION
+    section pass the TOP-LEVEL fields ``done`` (what's been done) and ``next``
+    (the immediate next step) — these are the required fields and they show
+    up here as discrete string params; ``where_to_look`` is optional. (Do NOT
+    pass an empty ``section={}``; the ``section`` dict is the free-form path
+    for other content types — developer ``{summary, changes}``, auditor
+    ``{summary, severity}``.) Or just ``text`` for a developer summary.
     """
     return _post(
         "/api/v1/do/note",
@@ -241,6 +248,9 @@ def note(
             "what_struggled": what_struggled,
             "next_steps": next_steps,
             "section": section,
+            "done": done,
+            "next": next,
+            "where_to_look": where_to_look,
         },
     )
 
diff --git a/roboco/mcp/flow_server.py b/roboco/mcp/flow_server.py
index 166000bb..56879312 100644
--- a/roboco/mcp/flow_server.py
+++ b/roboco/mcp/flow_server.py
@@ -18,11 +18,25 @@ import json
 import os
 import uuid
 from pathlib import Path
-from typing import Any
+from typing import Annotated, Any
 
 import httpx
 import structlog
 from mcp.server.fastmcp import FastMCP
+from pydantic import BeforeValidator
+
+from roboco.foundation.policy.content.validators import coerce_str_list
+
+# A ``list[str]`` field that tolerates the Claude SDK's XML-ish tool-input
+# parsing: an LLM emitting a bullet list as ``<item>…</item>`` elements arrives
+# as ``[[["…"]]]`` / ``[{"item": {"$text": "…"}}, …]`` — nested arrays / dicts,
+# not strings. A bare ``list[str]`` annotation hard-rejects element 1 (a list,
+# not a str) at the MCP validation layer BEFORE the verb body runs, surfacing as
+# ``1 validation error for i_will_planArguments technical_considerations.1
+# Input should be a valid string``. The ``BeforeValidator`` flattens it to a
+# flat ``list[str]`` first (same ``coerce_str_list`` used at the intake→DB
+# boundary — see Bug 3 in the MegaTask memory).
+StrList = Annotated[list[str], BeforeValidator(coerce_str_list)]
 
 ORCHESTRATOR_URL = os.environ.get(
     "ROBOCO_ORCHESTRATOR_URL",
@@ -214,7 +228,7 @@ def i_will_work_on(
     task_id: str,
     plan: str | None = None,
     steps: list[dict[str, str]] | None = None,
-    technical_considerations: list[str] | None = None,
+    technical_considerations: StrList | None = None,
     risks: list[dict[str, str]] | None = None,
     open_questions: list[dict[str, str | bool]] | None = None,
 ) -> dict[str, Any]:
@@ -333,7 +347,7 @@ def claim_review(task_id: str) -> dict[str, Any]:
 
 
 def pass_review(
-    task_id: str, notes: str, ac_verdicts: list[str] | None = None
+    task_id: str, notes: str, ac_verdicts: StrList | None = None
 ) -> dict[str, Any]:
     """QA: accept the work. notes >= 80 chars; journal:learning required.
 
@@ -347,7 +361,7 @@ def pass_review(
     return _post(_role_path("pass"), payload)
 
 
-def fail_review(task_id: str, issues: list[str]) -> dict[str, Any]:
+def fail_review(task_id: str, issues: StrList) -> dict[str, Any]:
     """QA: reject the work with issues. Each issue should be concrete and actionable."""
     return _post(_role_path("fail"), {"task_id": task_id, "issues": issues})
 
@@ -401,7 +415,7 @@ def claim_doc_task(task_id: str) -> dict[str, Any]:
     return _post(_role_path("claim_doc_task"), {"task_id": task_id})
 
 
-def i_documented(task_id: str, notes: str, files: list[str]) -> dict[str, Any]:
+def i_documented(task_id: str, notes: str, files: StrList) -> dict[str, Any]:
     """Doc: mark documentation complete. files=['<doc-path>', ...]."""
     return _post(
         _role_path("i_documented"),
@@ -463,7 +477,7 @@ def i_will_plan(
     plan: str,
     approach: str = "",
     sub_tasks: list[dict[str, str]] | None = None,
-    technical_considerations: list[str] | None = None,
+    technical_considerations: StrList | None = None,
     risks: list[dict[str, str]] | None = None,
     open_questions: list[dict[str, str | bool]] | None = None,
 ) -> dict[str, Any]:
@@ -506,9 +520,9 @@ def delegate(
     team: str,
     task_type: str,
     nature: str,
-    acceptance_criteria: list[str],
+    acceptance_criteria: StrList,
     estimated_complexity: str = "medium",
-    covers_parent_criteria: list[str] | None = None,
+    covers_parent_criteria: StrList | None = None,
 ) -> dict[str, Any]:
     """PM: create a subtask of parent_task_id.
 
@@ -566,7 +580,7 @@ def pr_pass(task_id: str, notes: str) -> dict[str, Any]:
     return _post(_role_path("pr_pass"), {"task_id": task_id, "notes": notes})
 
 
-def pr_fail(task_id: str, issues: list[str]) -> dict[str, Any]:
+def pr_fail(task_id: str, issues: StrList) -> dict[str, Any]:
     """PR reviewer: fail the assembled PR with concrete issues → needs_revision."""
     return _post(_role_path("pr_fail"), {"task_id": task_id, "issues": issues})
 
diff --git a/roboco/runtime/orchestrator.py b/roboco/runtime/orchestrator.py
index 44035b3f..6f2c04f0 100644
--- a/roboco/runtime/orchestrator.py
+++ b/roboco/runtime/orchestrator.py
@@ -4129,40 +4129,60 @@ class AgentOrchestrator:
     async def _persist_respawn_record(
         self, agent_slug: str, task_id: str, record: dict[str, Any]
     ) -> None:
-        """Write-through one PM-respawn counter row (delete-then-insert upsert).
+        """Write-through one PM-respawn counter row (atomic upsert).
 
         Best-effort, mirroring ``_persist_waiting_record``: a persistence failure
         must never gate or un-gate a spawn, so any error is logged and swallowed.
         The counter stays authoritative in memory regardless.
+
+        Unlike ``_persist_waiting_record`` (inline-awaited, one row per agent),
+        this is scheduled fire-and-forget per gate mutation, and a respawn loop
+        fires several persists for the same ``(agent_slug, task_id)`` in quick
+        succession. A delete-then-insert raced under that concurrency: two
+        transactions for the same key overlapped, the loser's INSERT hit
+        ``pk_respawn_tracker`` UniqueViolation, the durable count stuck at the
+        first INSERT's value, and a restart re-burned the strike threshold — the
+        exact re-burn this feature was built to stop (2026-06-27 live meltdown).
+        The single ``ON CONFLICT DO UPDATE`` upsert is race-free: concurrent
+        upserts on the same key serialize at row level.
         """
         try:
             from uuid import UUID as _UUID
 
-            from sqlalchemy import delete
+            from sqlalchemy.dialects.postgresql import insert as pg_insert
 
             from roboco.db.base import get_session_factory
             from roboco.db.tables import RespawnTrackerTable
 
             tid = _UUID(task_id)
+            now = datetime.now(UTC)
+            stmt = pg_insert(RespawnTrackerTable).values(
+                agent_slug=agent_slug,
+                task_id=tid,
+                count=int(record["count"]),
+                last_status=record.get("last_status"),
+                last_check=record["last_check"],
+                tracing_resets=int(record.get("tracing_resets", 0)),
+                notified=bool(record.get("notified", False)),
+                updated_at=now,
+            )
+            stmt = stmt.on_conflict_do_update(
+                index_elements=[
+                    RespawnTrackerTable.agent_slug,
+                    RespawnTrackerTable.task_id,
+                ],
+                set_={
+                    "count": stmt.excluded.count,
+                    "last_status": stmt.excluded.last_status,
+                    "last_check": stmt.excluded.last_check,
+                    "tracing_resets": stmt.excluded.tracing_resets,
+                    "notified": stmt.excluded.notified,
+                    "updated_at": stmt.excluded.updated_at,
+                },
+            )
             session_factory = get_session_factory()
             async with session_factory() as db:
-                await db.execute(
-                    delete(RespawnTrackerTable).where(
-                        RespawnTrackerTable.agent_slug == agent_slug,
-                        RespawnTrackerTable.task_id == tid,
-                    )
-                )
-                db.add(
-                    RespawnTrackerTable(
-                        agent_slug=agent_slug,
-                        task_id=tid,
-                        count=int(record["count"]),
-                        last_status=record.get("last_status"),
-                        last_check=record["last_check"],
-                        tracing_resets=int(record.get("tracing_resets", 0)),
-                        notified=bool(record.get("notified", False)),
-                    )
-                )
+                await db.execute(stmt)
                 await db.commit()
         except Exception as e:
             logger.error(
diff --git a/roboco/services/gateway/choreographer/_impl.py b/roboco/services/gateway/choreographer/_impl.py
index 83b04f06..81470810 100644
--- a/roboco/services/gateway/choreographer/_impl.py
+++ b/roboco/services/gateway/choreographer/_impl.py
@@ -4541,6 +4541,7 @@ class Choreographer:
         have gotten from the upfront completeness check.
         """
         from roboco.foundation.policy.task_completeness import TaskCompletenessError
+        from roboco.services.base import ValidationError
 
         parent_task_id = parent.id
         try:
@@ -4563,6 +4564,24 @@ class Choreographer:
                 task_id=parent_task_id,
                 verb="delegate",
             )
+        except ValidationError as exc:
+            # A user-input error from task creation (e.g. delegating past
+            # MAX_TASK_DEPTH — ``_validate_parent_depth`` raises this with a
+            # remediation message telling the PM to create a sibling instead of
+            # a further nested subtask). Without this translator it escaped
+            # uncaught as a 500 ExceptionGroup; the agent never saw the fix.
+            # ``invalid_state`` carries the message as the remediation so the
+            # PM gets a clean, actionable rejection it can act on.
+            return await self._emit_rejection(
+                Envelope.invalid_state(
+                    message=exc.message,
+                    remediate=exc.message,
+                    context_briefing=briefing,
+                ).with_introspection(task=parent, role=role_str),
+                agent_id=pm_agent_id,
+                task_id=parent_task_id,
+                verb="delegate",
+            )
         await self._wire_ux_frontend_dependency(new_task, parent)
         hint = self._sizing_hint(inputs)
         if hint is not None:
diff --git a/roboco/services/gateway/content_actions.py b/roboco/services/gateway/content_actions.py
index 6312f54d..0b0e6cfc 100644
--- a/roboco/services/gateway/content_actions.py
+++ b/roboco/services/gateway/content_actions.py
@@ -36,6 +36,36 @@ if TYPE_CHECKING:
 logger = structlog.get_logger()
 
 
+def _merge_resumption_fields(
+    section: dict[str, Any] | None,
+    *,
+    done: str,
+    next: str,
+    where_to_look: list[str] | None,
+) -> dict[str, Any] | None:
+    """Fold the top-level resumption fields into the handoff ``section``.
+
+    ``section: dict[str, Any]`` renders a tool schema with no visible
+    sub-fields, so a weak model (minimax-m3) emits ``section={}`` and the
+    resumption gate rejects ``done — Field required`` (the 2026-06-27 PM
+    respawn-loop meltdown). The top-level ``done`` / ``next`` /
+    ``where_to_look`` string fields are the LLM-facing contract — they show
+    up in the tool schema as discrete fields the same model fills fine. Here
+    they fill any keys the explicit ``section`` omits without overwriting
+    keys the agent already supplied, so a capable model passing ``section``
+    directly is unaffected. Returns ``None`` when nothing was supplied so the
+    downstream ``{'summary': text}`` fallback + gate remediation still fire.
+    """
+    merged: dict[str, Any] = dict(section) if section else {}
+    if done and "done" not in merged:
+        merged["done"] = done
+    if next and "next" not in merged:
+        merged["next"] = next
+    if where_to_look and "where_to_look" not in merged:
+        merged["where_to_look"] = where_to_look
+    return merged or None
+
+
 # Scope catalog is canonical in foundation.policy.journaling.
 # Derived here as a string frozenset for the existing call sites that
 # compare strings rather than the Scope enum.
@@ -541,6 +571,9 @@ class ContentActions:
         task_id: UUID | None = None,
         structured: dict[str, Any] | None = None,
         section: dict[str, Any] | None = None,
+        done: str = "",
+        next: str = "",
+        where_to_look: list[str] | None = None,
     ) -> Envelope:
         """Write a journal entry, or (scope='handoff') the role's note section.
 
@@ -554,6 +587,17 @@ class ContentActions:
         - decision: context, options[], chosen, rationale, consequences
         - reflect: what_done, what_learned, what_struggled, next_steps
 
+        For ``scope='handoff'`` the resumption section (PM / coordinator
+        roles) can be authored two ways: the nested ``section`` dict, OR the
+        top-level ``done`` / ``next`` / ``where_to_look`` string fields. The
+        top-level path is the LLM-facing contract — ``section: dict[str, Any]``
+        renders a tool schema with no visible sub-fields, so a weak model
+        (minimax-m3) emits ``section={}`` and the resumption gate rejects
+        ``done — Field required``; the top-level typed strings show up in the
+        tool schema as discrete fields the same model fills fine (proven on
+        the decision scope). Top-level fields fill any keys the explicit
+        ``section`` omits without overwriting supplied ones.
+
         The note is always recorded. List-typed fields tolerate a
         lone scalar (wrapped into a one-element list) and missing decision/
         reflect narrative fields default to a visible placeholder, so a
@@ -571,7 +615,12 @@ class ContentActions:
             # a journal entry. Content quality is enforced by the content model
             # (apply_structured_note), so skip the journal-text soup check.
             return await self._record_section_handoff(
-                agent_id=agent_id, text=text, task_id=task_id, structured=section
+                agent_id=agent_id,
+                text=text,
+                task_id=task_id,
+                structured=_merge_resumption_fields(
+                    section, done=done, next=next, where_to_look=where_to_look
+                ),
             )
         return await self._write_journal_note(
             agent_id=agent_id,
diff --git a/roboco/services/gateway/remediation.py b/roboco/services/gateway/remediation.py
index 75911eba..7df37df9 100644
--- a/roboco/services/gateway/remediation.py
+++ b/roboco/services/gateway/remediation.py
@@ -94,6 +94,7 @@ def hint_for_short_quick_context(*, min_chars: int, task_id: str) -> str:
     return (
         f"your quick_context section is empty or under {min_chars} chars. Before "
         f"delegate, call note(scope='handoff', task_id='{task_id}', "
-        "section={'done': '<state so far>', 'next': '<what the cell should do>'})"
-        " to leave a resumption handoff, then retry."
+        "done='<state so far>', next='<what the cell should do>') to leave a "
+        "resumption handoff (pass done and next as top-level string args, not "
+        "nested in section), then retry."
     )
diff --git a/roboco/services/prompter.py b/roboco/services/prompter.py
index a64886f7..2d8d5ff0 100644
--- a/roboco/services/prompter.py
+++ b/roboco/services/prompter.py
@@ -22,6 +22,7 @@ from sqlalchemy import select
 from roboco.db.tables import AgentTable, TaskTable
 from roboco.foundation.identity import CELL_TEAMS
 from roboco.foundation.policy.batch import is_batch_umbrella
+from roboco.foundation.policy.content.validators import coerce_str_list
 from roboco.foundation.policy.sequencing.models import DraftSurface, SequencePlan
 from roboco.models.base import (
     AgentRole,
@@ -235,6 +236,27 @@ class PrompterService:
                 message="This task draft is missing acceptance criteria.",
                 field="acceptance_criteria",
             )
+        # Flatten the string-list fields to list[str]. The agent sometimes emits
+        # these as XML-ish <item>…</item> elements the SDK parses into dict
+        # wrappers ({"item": {"$text": "…"}}); left as-is they crash the
+        # VARCHAR[] insert (asyncpg: "expected str, got dict") and dump str(dict)
+        # into the rendered description. Coerce here too — a draft can arrive
+        # via redraft/localStorage, not only the intake choke point.
+        draft_data["acceptance_criteria"] = coerce_str_list(
+            draft_data.get("acceptance_criteria")
+        )
+        draft_data["what_this_builds"] = coerce_str_list(
+            draft_data.get("what_this_builds")
+        )
+        draft_data["notes"] = coerce_str_list(draft_data.get("notes"))
+        for unit in draft_data.get("the_work") or []:
+            if isinstance(unit, dict):
+                unit["items"] = coerce_str_list(unit.get("items"))
+        if not draft_data["acceptance_criteria"]:
+            raise ValidationError(
+                message="This task draft is missing acceptance criteria.",
+                field="acceptance_criteria",
+            )
         # Recompose the description from the (possibly edited) structured fields —
         # the task always carries a freshly-composed, consistent description.
         draft_data["description"] = compose_description(draft_data)
@@ -833,8 +855,12 @@ def derive_scale(the_work: list[Any]) -> str:
 
 
 def _clean_list(value: Any) -> list[str]:
-    """Trimmed, non-empty string items from a possibly-missing list field."""
-    return [str(i).strip() for i in (value or []) if str(i).strip()]
+    """Trimmed, non-empty string items from a possibly-missing list field.
+
+    Uses :func:`coerce_str_list` so an LLM's dict-wrapped items (``{"$text": …}``
+    etc.) are extracted to text rather than dumped as ``str(dict)``.
+    """
+    return coerce_str_list(value)
 
 
 def _text(value: Any) -> str:
diff --git a/roboco/services/task.py b/roboco/services/task.py
index 91fb1d93..53e771e2 100644
--- a/roboco/services/task.py
+++ b/roboco/services/task.py
@@ -713,10 +713,14 @@ class TaskService(BaseService):
     async def _validate_parent_depth(self, parent_task_id: UUID) -> None:
         """Enforce MAX_TASK_DEPTH at creation time.
 
-        Walks up the parent chain counting ancestors. Raises ValueError if
-        adding a child under this parent would exceed MAX_TASK_DEPTH.
-        Previously this was only enforced at branch-name generation time,
-        so invalid hierarchies could be created and only fail later at claim.
+        Walks up the parent chain counting ancestors. Raises ValidationError
+        (a ServiceError the API/gateway translate to a clean 400 / remediation
+        envelope) if adding a child under this parent would exceed
+        MAX_TASK_DEPTH. Previously this raised a bare ValueError that escaped
+        uncaught as a 500 (the message told the agent to create a sibling, but
+        it never reached the agent as a handled error), and before that it was
+        only enforced at branch-name generation time, so invalid hierarchies
+        could be created and only fail later at claim.
         """
         from roboco.templates.git.constants import MAX_TASK_DEPTH
 
@@ -726,19 +730,24 @@ class TaskService(BaseService):
         while current_id is not None:
             key = str(current_id)
             if key in visited:
-                raise ValueError(
-                    f"Circular reference detected at {key} while validating depth"
+                raise ValidationError(
+                    f"Circular reference detected at {key} while validating depth",
+                    field="parent_task_id",
                 )
             visited.add(key)
             parent = await self.get(current_id)
             if parent is None:
-                raise ValueError(f"Parent task {current_id} not found")
+                raise ValidationError(
+                    f"Parent task {current_id} not found",
+                    field="parent_task_id",
+                )
             depth += 1
             if depth >= MAX_TASK_DEPTH:
-                raise ValueError(
+                raise ValidationError(
                     f"Task hierarchy would exceed MAX_TASK_DEPTH={MAX_TASK_DEPTH}. "
                     "Create this work as a sibling of the deepest task instead "
-                    "of a further nested subtask."
+                    "of a further nested subtask.",
+                    field="parent_task_id",
                 )
             parent_parent = parent.parent_task_id
             current_id = UUID(str(parent_parent)) if parent_parent else None
diff --git a/roboco/templates/git/branch.py b/roboco/templates/git/branch.py
index 1661a547..fe283e14 100644
--- a/roboco/templates/git/branch.py
+++ b/roboco/templates/git/branch.py
@@ -13,7 +13,8 @@ Uses '--' separator for task hierarchy to avoid git ref conflicts.
 Git cannot have both 'foo' as a branch AND 'foo/bar' as another branch,
 so we use '--' instead of '/' for the task hierarchy portion.
 
-Max 3 levels deep (root → subtask → sub-subtask).
+Max 4 levels deep — MegaTask adds an umbrella Main-PM layer on top of the
+normal flow, so the full path is umbrella → root → cell → dev.
 """
 
 from typing import TYPE_CHECKING
@@ -39,7 +40,7 @@ async def build_branch_name(
     team: str,
     task_service: "TaskService",
 ) -> str:
-    """Build branch name with ancestor path (max 3 levels, full UUIDs).
+    """Build branch name with ancestor path (max 4 levels, full UUIDs).
 
     Args:
         task_id: The task to create branch for
@@ -48,7 +49,9 @@ async def build_branch_name(
         task_service: TaskService instance for fetching task hierarchy
 
     Returns:
-        Branch name in format: {type}/{team}/{root}--{sub}--{subsub}
+        Branch name in format: {type}/{team}/{root}--{sub}--{subsub}--{subsubsub}
+        (MegaTask: umbrella--root--cell--dev; shorter hierarchies omit the
+        trailing segments).
 
     Raises:
         BranchNameError: If branch_type invalid, task not found, or hierarchy too deep
diff --git a/roboco/templates/git/constants.py b/roboco/templates/git/constants.py
index 0fcef34b..9fb189e0 100644
--- a/roboco/templates/git/constants.py
+++ b/roboco/templates/git/constants.py
@@ -33,8 +33,16 @@ COMMIT_TYPES: Final[frozenset[str]] = frozenset(
     }
 )
 
-# Maximum task hierarchy depth (root → subtask → sub-subtask)
-MAX_TASK_DEPTH: Final[int] = 3
+# Maximum task hierarchy depth. MegaTask adds one Main-PM layer on top of the
+# normal flow, so the full hierarchy is 4 layers: umbrella (Main PM, depth 0)
+# → root-subtask (Main PM, depth 1) → cell task (cell PM, depth 2) → dev
+# subtask (depth 3). This was sized at 3 for the pre-MegaTask 3-layer flow
+# (root→cell→dev = depths 0,1,2) and never raised when MegaTask shipped, so
+# cell-PM delegation of dev subtasks was rejected with MAX_TASK_DEPTH=3 and
+# deadlocked the cell PM into a respawn loop (2026-06-27 live meltdown).
+# The validator rejects a child whose depth would reach MAX_TASK_DEPTH, so 4
+# permits the dev subtask at depth 3 while still capping a 5th layer.
+MAX_TASK_DEPTH: Final[int] = 4
 
 # Git branch name character limit
 GIT_BRANCH_MAX_LENGTH: Final[int] = 255
diff --git a/tests/integration/test_branch_name_builder.py b/tests/integration/test_branch_name_builder.py
index 629b12bf..4450f035 100644
--- a/tests/integration/test_branch_name_builder.py
+++ b/tests/integration/test_branch_name_builder.py
@@ -174,3 +174,50 @@ async def test_get_root_task_id_too_deep_raises() -> None:
     )
     with pytest.raises(BranchNameError, match="too deep"):
         await get_root_task_id(uuid4(), fake_svc)
+
+
+@pytest.mark.asyncio
+async def test_build_branch_name_allows_four_level_megatask_hierarchy() -> None:
+    """The MegaTask hierarchy is 4 layers — umbrella (Main PM, depth 0) →
+    root-subtask (Main PM, depth 1) → cell task (cell PM, depth 2) → dev
+    subtask (depth 3). MAX_TASK_DEPTH was sized for the 3-layer normal flow
+    (root→cell→dev = depths 0,1,2) and never raised when MegaTask added the
+    umbrella layer, so cell-PM delegation of dev subtasks was rejected with
+    MAX_TASK_DEPTH=3 and the cell PM deadlocked into a respawn loop
+    (2026-06-27 live meltdown — be-pm on task 9980d0a0). The cap must
+    accommodate the 4-level MegaTask: a dev subtask's branch name builds as a
+    4-segment path (umbrella--root--cell--dev), not raise 'too deep'."""
+    umbrella, root, cell, dev = uuid4(), uuid4(), uuid4(), uuid4()
+    tasks = {
+        dev: SimpleNamespace(id=dev, parent_task_id=cell),
+        cell: SimpleNamespace(id=cell, parent_task_id=root),
+        root: SimpleNamespace(id=root, parent_task_id=umbrella),
+        umbrella: SimpleNamespace(id=umbrella, parent_task_id=None),
+    }
+    fake_svc = AsyncMock()
+    fake_svc.get.side_effect = lambda task_id: tasks[task_id]
+
+    name = await build_branch_name(dev, "feature", "backend", fake_svc)
+
+    assert name == (
+        f"feature/backend/{str(umbrella)[:8]}--{str(root)[:8]}"
+        f"--{str(cell)[:8]}--{str(dev)[:8]}"
+    )
+
+
+@pytest.mark.asyncio
+async def test_get_root_task_id_returns_umbrella_for_four_level_hierarchy() -> None:
+    """Companion to the branch-name test: the root of a 4-level MegaTask chain
+    is the umbrella, and it must be reachable (not truncated as 'too deep')
+    under the raised cap."""
+    umbrella, root, cell, dev = uuid4(), uuid4(), uuid4(), uuid4()
+    tasks = {
+        dev: SimpleNamespace(id=dev, parent_task_id=cell),
+        cell: SimpleNamespace(id=cell, parent_task_id=root),
+        root: SimpleNamespace(id=root, parent_task_id=umbrella),
+        umbrella: SimpleNamespace(id=umbrella, parent_task_id=None),
+    }
+    fake_svc = AsyncMock()
+    fake_svc.get.side_effect = lambda task_id: tasks[task_id]
+
+    assert await get_root_task_id(dev, fake_svc) == umbrella
diff --git a/tests/integration/test_task_service_misc.py b/tests/integration/test_task_service_misc.py
index f1303ad4..625caf72 100644
--- a/tests/integration/test_task_service_misc.py
+++ b/tests/integration/test_task_service_misc.py
@@ -181,7 +181,7 @@ async def test_validate_parent_depth_missing_parent_raises(
     task_setup: dict,
 ) -> None:
     svc = task_setup["svc"]
-    with pytest.raises(ValueError, match="not found"):
+    with pytest.raises(ValidationError, match="not found"):
         await svc._validate_parent_depth(uuid4())
 
 
@@ -190,14 +190,14 @@ async def test_validate_parent_depth_circular_reference(
     task_setup: dict,
     db_session: AsyncSession,
 ) -> None:
-    """A self-referential parent loop raises ValueError."""
+    """A self-referential parent loop raises ValidationError."""
     svc = task_setup["svc"]
     a = await svc.create(_req(task_setup))
     b = await svc.create(_req(task_setup, parent_task_id=a.id))
     # Force circular: a.parent_task_id = b.id
     a.parent_task_id = b.id
     await db_session.flush()
-    with pytest.raises(ValueError, match="Circular reference"):
+    with pytest.raises(ValidationError, match="Circular reference"):
         await svc._validate_parent_depth(a.id)
 
 
@@ -205,7 +205,7 @@ async def test_validate_parent_depth_circular_reference(
 async def test_validate_parent_depth_exceeds_max(
     task_setup: dict,
 ) -> None:
-    """Adding a child past MAX_TASK_DEPTH raises ValueError."""
+    """Adding a child past MAX_TASK_DEPTH raises ValidationError."""
     svc = task_setup["svc"]
     # Build a chain of MAX_TASK_DEPTH+1 tasks
     parent = None
@@ -215,7 +215,7 @@ async def test_validate_parent_depth_exceeds_max(
         )
         parent = new
     assert parent is not None
-    with pytest.raises(ValueError, match="MAX_TASK_DEPTH"):
+    with pytest.raises(ValidationError, match="MAX_TASK_DEPTH"):
         await svc._validate_parent_depth(parent.id)
 
 
diff --git a/tests/unit/agent_sdk/test_intake_driver.py b/tests/unit/agent_sdk/test_intake_driver.py
index d51864c8..41ea82a8 100644
--- a/tests/unit/agent_sdk/test_intake_driver.py
+++ b/tests/unit/agent_sdk/test_intake_driver.py
@@ -224,6 +224,54 @@ def test_normalize_propose_batch_reports_dropped_count() -> None:
     assert [d["title"] for d in chunks[0].data["drafts"]] == ["Good"]
 
 
+def test_normalize_propose_batch_coerces_the_work_scalar_to_list() -> None:
+    # The agent may emit `the_work` as a lone object (single-cell task) or its
+    # `items` as a lone string instead of a list. The panel treats `the_work`
+    # as an array (`(draft.the_work ?? []).map`) and crashes if it isn't — so
+    # the backend coerces these list-shaped fields before they reach SSE, the
+    # same graceful single-item wrap the content models apply at confirm time.
+    msg = AssistantMessage(
+        [
+            ToolUseBlock(
+                "propose_batch",
+                {
+                    "drafts": [
+                        {
+                            "title": "Single-cell with bare the_work",
+                            "acceptance_criteria": "one string not a list",
+                            "the_work": {
+                                "team": "backend",
+                                "summary": "do the thing",
+                                "items": "lone item string",
+                            },
+                        }
+                    ]
+                },
+            )
+        ]
+    )
+    chunks = normalize(msg)
+    assert [c.kind for c in chunks] == ["batch"]
+    draft = chunks[0].data["drafts"][0]
+    assert draft["acceptance_criteria"] == ["one string not a list"]
+    assert isinstance(draft["the_work"], list) and len(draft["the_work"]) == 1
+    unit = draft["the_work"][0]
+    assert unit["team"] == "backend"
+    assert unit["items"] == ["lone item string"]
+    # A non-list, non-dict the_work (e.g. a stray string) is dropped so the
+    # panel never sees a value it cannot .map over.
+    msg2 = AssistantMessage(
+        [
+            ToolUseBlock(
+                "propose_batch",
+                {"drafts": [{"title": "Bad work", "the_work": "not a list"}]},
+            )
+        ]
+    )
+    draft2 = normalize(msg2)[0].data["drafts"][0]
+    assert draft2["the_work"] == []
+
+
 def test_normalize_propose_draft_without_title_is_ignored() -> None:
     msg = AssistantMessage(
         [ToolUseBlock("propose_draft", {"draft": {"acceptance_criteria": []}})]
diff --git a/tests/unit/api/test_schemas_v1_flow.py b/tests/unit/api/test_schemas_v1_flow.py
index 60dcbf02..54a0fe8d 100644
--- a/tests/unit/api/test_schemas_v1_flow.py
+++ b/tests/unit/api/test_schemas_v1_flow.py
@@ -6,7 +6,11 @@ from uuid import uuid4
 
 import pytest
 from pydantic import ValidationError
-from roboco.api.schemas.v1.flow import DelegateRequest
+from roboco.api.schemas.v1.flow import (
+    DelegateRequest,
+    IWillPlanRequest,
+    IWillWorkOnRequest,
+)
 
 
 def test_delegate_request_requires_task_type() -> None:
@@ -48,3 +52,85 @@ def test_delegate_request_accepts_explicit_task_type() -> None:
         acceptance_criteria=["returns 200"],
     )
     assert req.task_type == "code"
+
+
+# ---------------------------------------------------------------------------
+# StrList — SDK-nested list-of-strings coercion (Bug A)
+# ---------------------------------------------------------------------------
+
+
+def test_i_will_plan_request_flattens_sdk_nested_technical_considerations() -> None:
+    """The Claude SDK parses XML-ish ``<item>…</item>`` list-of-strings tool
+    input into nested arrays (``[[["…"]]]``). A bare ``list[str]`` field
+    hard-rejects element 1 (a list, not a str) at validation time — the live
+    ``i_will_plan`` crash: ``technical_considerations.1 Input should be a
+    valid string``. The ``StrList`` BeforeValidator must flatten it to a flat
+    ``list[str]`` so the verb body receives clean strings.
+    """
+    req = IWillPlanRequest(
+        task_id=uuid4(),
+        plan="Plan narrative describing the approach in full sentences.",
+        approach=(
+            "Approach text long enough to clear the 150-character minimum "
+            "enforced on the plan's Approach field so the Plan tab is fully "
+            "populated for audit and tracing instead of rendering an empty view."
+        ),
+        technical_considerations=[
+            [[["Empty state distinct from loaded state, coverage target 80%"]]],
+            [{"item": {"$text": "Use asyncpg prepared statements"}}],
+        ],
+    )
+    assert req.technical_considerations == [
+        "Empty state distinct from loaded state, coverage target 80%",
+        "Use asyncpg prepared statements",
+    ]
+
+
+def test_i_will_work_on_request_flattens_dict_wrapped_technical_considerations() -> (
+    None
+):
+    """Same coercion on the developer planning verb — a dict-wrapped string
+    (``{"item": {"$text": "…"}}``, the SDK's element-text marker) must reduce
+    to the bare string, not ``str(dict)``."""
+    req = IWillWorkOnRequest(
+        task_id=uuid4(),
+        technical_considerations=[{"item": {"$text": "Cache the lookup result"}}],
+    )
+    assert req.technical_considerations == ["Cache the lookup result"]
+
+
+def test_delegate_request_flattens_sdk_nested_acceptance_criteria() -> None:
+    """``delegate``'s ``acceptance_criteria`` is the same list-of-strings shape
+    the SDK can nest (this is the ``delegate``-verb analogue of the MegaTask
+    Bug 3 crash). The ``StrList`` field must flatten the nested input so the
+    VARCHAR[] insert downstream never sees a dict/list element."""
+    req = DelegateRequest(
+        parent_task_id=uuid4(),
+        title="t",
+        description="add the new endpoint plus tests",
+        assigned_to="be-dev-1",
+        team="backend",
+        task_type="code",
+        nature="technical",
+        estimated_complexity="medium",
+        acceptance_criteria=[
+            [[["returns 200 for valid input"]]],
+            [{"item": {"$text": "rejects malformed input with 400"}}],
+        ],
+    )
+    assert req.acceptance_criteria == [
+        "returns 200 for valid input",
+        "rejects malformed input with 400",
+    ]
+
+
+def test_strlist_drops_non_string_junk_instead_of_crashing() -> None:
+    """Non-string junk (a bare int, a dict with no string values, whitespace)
+    is dropped — the field never raises on garbage the SDK might emit; only
+    real strings survive. An all-junk payload yields an empty list (the
+    delegate min_length=1 gate then rejects it cleanly, not a 500)."""
+    req = IWillWorkOnRequest(
+        task_id=uuid4(),
+        technical_considerations=[42, {"foo": 123}, [["  "]], "real note"],
+    )
+    assert req.technical_considerations == ["real note"]
diff --git a/tests/unit/foundation/policy/content/test_validators.py b/tests/unit/foundation/policy/content/test_validators.py
index 3ab740ef..d77c9649 100644
--- a/tests/unit/foundation/policy/content/test_validators.py
+++ b/tests/unit/foundation/policy/content/test_validators.py
@@ -9,7 +9,7 @@ every whitespace token is a placeholder (``wip wip``, ``tbd / na``).
 from __future__ import annotations
 
 import pytest
-from roboco.foundation.policy.content.validators import reject_trivial
+from roboco.foundation.policy.content.validators import coerce_str_list, reject_trivial
 
 
 def test_returns_trimmed_value_when_substantive() -> None:
@@ -57,3 +57,47 @@ def test_rejects_all_filler_token_string(soup: str) -> None:
 )
 def test_accepts_real_text_containing_a_filler_word(ok: str) -> None:
     assert reject_trivial(ok, field="reason", min_chars=3) == ok.strip()
+
+
+# ---------------------------------------------------------------------------
+# coerce_str_list — flatten an LLM's dict-wrapped list-of-strings to list[str].
+# The agent sometimes emits a list[str] field as XML-ish <item>…</item> elements
+# the Claude SDK parses into {"item": {"$text": "…"}}; left as dicts they crash
+# a VARCHAR[] insert ("expected str, got dict") and dump str(dict) into prose.
+# ---------------------------------------------------------------------------
+
+
+def test_coerce_str_list_passes_plain_strings_through() -> None:
+    assert coerce_str_list(["one", "two"]) == ["one", "two"]
+
+
+def test_coerce_str_list_extracts_sdk_item_text_wrapper() -> None:
+    # The exact shape from the live crash: [{"item": {"$text": "…"}}, ...].
+    assert coerce_str_list(
+        [{"item": {"$text": "UX-UI delivers a sketch"}}, {"item": "Spec lands first"}]
+    ) == ["UX-UI delivers a sketch", "Spec lands first"]
+
+
+def test_coerce_str_list_wraps_a_bare_dict() -> None:
+    assert coerce_str_list({"$text": "only one"}) == ["only one"]
+
+
+def test_coerce_str_list_wraps_a_bare_string() -> None:
+    assert coerce_str_list("only one") == ["only one"]
+
+
+def test_coerce_str_list_drops_non_string_junk() -> None:
+    # A non-str, non-dict element is dropped — never passed to a VARCHAR[] column.
+    assert coerce_str_list(["keep", 7, None, {"text": "nested"}]) == [
+        "keep",
+        "nested",
+    ]
+
+
+def test_coerce_str_list_recurses_into_nested_lists() -> None:
+    assert coerce_str_list([["a", {"item": "b"}], "c"]) == ["a", "b", "c"]
+
+
+def test_coerce_str_list_none_and_empty() -> None:
+    assert coerce_str_list(None) == []
+    assert coerce_str_list([]) == []
diff --git a/tests/unit/gateway/test_choreographer_delegate_guards.py b/tests/unit/gateway/test_choreographer_delegate_guards.py
index 9d1a83b8..8ef88063 100644
--- a/tests/unit/gateway/test_choreographer_delegate_guards.py
+++ b/tests/unit/gateway/test_choreographer_delegate_guards.py
@@ -16,6 +16,7 @@ from unittest.mock import AsyncMock, MagicMock
 from uuid import uuid4
 
 import pytest
+from roboco.services.base import ValidationError
 from roboco.services.gateway.choreographer import (
     Choreographer,
     ChoreographerDeps,
@@ -284,3 +285,45 @@ async def test_delegate_blocks_when_parent_quick_context_empty() -> None:
     assert body["error"] == "tracing_gap"
     assert "quick_context>=min" in body["missing"]
     task_svc.create_subtask.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_delegate_past_max_depth_returns_invalid_state_not_500() -> None:
+    """Bug B: delegating past MAX_TASK_DEPTH raised a bare ValueError that
+    escaped uncaught as a 500 ExceptionGroup. ``_validate_parent_depth`` now
+    raises ``ValidationError`` (a ServiceError), and the choreographer
+    translates it to an ``invalid_state`` envelope whose remediate carries the
+    "create as a sibling" instruction — so the PM gets a clean, actionable
+    rejection instead of a crash, and the agent loop doesn't respawn-blind.
+    """
+    pm_id = uuid4()
+    parent_id = uuid4()
+    parent = MagicMock(
+        id=parent_id,
+        project_id=uuid4(),
+        status="in_progress",
+        assigned_to=pm_id,
+        quick_context="Decomposition planned; cells implement their slice next.",
+    )
+    depth_msg = (
+        "Task hierarchy would exceed MAX_TASK_DEPTH=4. Create this work as a "
+        "sibling of the deepest task instead of a further nested subtask."
+    )
+    task_svc = AsyncMock()
+    task_svc.get.return_value = parent
+    task_svc.agent_for.return_value = MagicMock(role="cell_pm", team="backend")
+    task_svc.get_subtasks.return_value = []
+    # create_subtask -> create -> _validate_parent_depth raises ValidationError.
+    task_svc.create_subtask.side_effect = ValidationError(
+        depth_msg, field="parent_task_id"
+    )
+    deps = _make_deps(task=task_svc)
+    c = Choreographer(deps)
+
+    env = await c.delegate(pm_id, parent_id, _delegate_inputs())
+    body = env.as_dict()
+    assert body["error"] == "invalid_state"
+    assert "MAX_TASK_DEPTH" in body["message"]
+    # The remediation must reach the agent — the whole point of the fix.
+    assert "sibling" in body["remediate"]
+    task_svc.create_subtask.assert_awaited_once()
diff --git a/tests/unit/gateway/test_note_handoff.py b/tests/unit/gateway/test_note_handoff.py
index 26075314..9cb601c8 100644
--- a/tests/unit/gateway/test_note_handoff.py
+++ b/tests/unit/gateway/test_note_handoff.py
@@ -130,6 +130,96 @@ async def test_handoff_validation_error_returns_remediation_not_422() -> None:
     assert "auditor" in body["remediate"]
 
 
+@pytest.mark.asyncio
+async def test_handoff_pm_resumption_from_top_level_done_next() -> None:
+    """A PM authors the resumption section from TOP-LEVEL ``done``/``next``
+    fields, not the nested ``section`` dict.
+
+    Background (2026-06-27 live meltdown): minimax-m3:cloud running the PM
+    roles emitted ``section={}`` (an empty object) for the handoff note 3
+    times identically, ignoring the prose remediate. ``ResumptionNote`` requires
+    ``done``+``next`` so the gate rejected ``done — Field required``, the
+    do-server circuit breaker tripped after 3-4 rejections, and the tracing
+    gate (which obligates a handoff note before ``delegate``) deadlocked the
+    PM into a respawn loop. Root cause was structural: ``section: dict[str,
+    Any]`` renders a tool schema with NO visible sub-fields, so a weak model
+    emits ``{}`` while the SAME model fills the top-level ``string`` decision
+    fields fine (proven live — those succeeded). The fix follows the
+    precedent already in ``NoteRequest`` (the decision fields typed ``str =
+    ""`` so the schema declares ``string``): promote ``done``/``next`` to
+    top-level typed string params so the tool schema shows them
+    machine-visible. Passing ``done``+``next`` with no ``section`` must write
+    the resumption section successfully.
+    """
+    agent_id, task_id = uuid4(), uuid4()
+    svc = _dev_task_svc(task_id, role="cell_pm")
+    ca = ContentActions(_make_deps(task=svc))
+
+    env = await ca.note(
+        agent_id=agent_id,
+        text="handoff",
+        scope="handoff",
+        done="Planned the decomposition into 3 cell tasks.",
+        next="Cells implement their slices in wave order.",
+        where_to_look=["panel/tasks — wave column", "the decomposition note"],
+    )
+
+    body = env.as_dict()
+    assert body["error"] is None
+    _tid, content_type, payload = svc.record_section_note.call_args.args
+    assert content_type == "resumption"
+    assert payload["done"] == "Planned the decomposition into 3 cell tasks."
+    assert payload["next"] == "Cells implement their slices in wave order."
+    assert payload["where_to_look"] == [
+        "panel/tasks — wave column",
+        "the decomposition note",
+    ]
+
+
+@pytest.mark.asyncio
+async def test_handoff_top_level_done_next_merge_into_explicit_section() -> None:
+    """Top-level ``done``/``next`` fill any keys the explicit ``section`` omits
+    without overwriting keys the agent already supplied — backward compatible
+    with a capable model that passes ``section`` directly."""
+    agent_id, task_id = uuid4(), uuid4()
+    svc = _dev_task_svc(task_id, role="cell_pm")
+    ca = ContentActions(_make_deps(task=svc))
+
+    env = await ca.note(
+        agent_id=agent_id,
+        text="handoff",
+        scope="handoff",
+        section={"done": "Already-specified done."},
+        next="Top-level next fills the omitted next.",
+    )
+
+    assert env.as_dict()["error"] is None
+    _tid, _ctype, payload = svc.record_section_note.call_args.args
+    assert payload["done"] == "Already-specified done."
+    assert payload["next"] == "Top-level next fills the omitted next."
+
+
+@pytest.mark.asyncio
+async def test_handoff_empty_done_next_empty_section_still_remediates() -> None:
+    """The backstop holds: with no top-level ``done``/``next`` AND an empty
+    ``section``, the resumption gate still rejects with the remediation
+    envelope (never a raw 422) — so a model that ignores both paths still
+    gets the actionable message rather than tripping a silent crash."""
+    agent_id, task_id = uuid4(), uuid4()
+    svc = _dev_task_svc(task_id, role="cell_pm")
+    svc.record_section_note.side_effect = ContentValidationError(
+        "done", "field required"
+    )
+    ca = ContentActions(_make_deps(task=svc))
+
+    env = await ca.note(agent_id=agent_id, text="handoff", scope="handoff", section={})
+    body = env.as_dict()
+
+    assert body["error"] == "invalid_state"
+    assert "done" in body["message"]
+    assert "resumption" in body["remediate"]
+
+
 @pytest.mark.asyncio
 async def test_handoff_no_task_to_attach_is_rejected() -> None:
     """With no active/context task and no explicit task_id, handoff refuses."""
diff --git a/tests/unit/runtime/test_respawn_persistence.py b/tests/unit/runtime/test_respawn_persistence.py
index b1831f1a..f65d1459 100644
--- a/tests/unit/runtime/test_respawn_persistence.py
+++ b/tests/unit/runtime/test_respawn_persistence.py
@@ -20,6 +20,7 @@ from uuid import uuid4
 
 import pytest
 from roboco.runtime.orchestrator import AgentOrchestrator
+from sqlalchemy.dialects import postgresql
 
 _SEEDED_COUNT = 3  # a persisted strike count, one below the trip threshold
 _STRIKE_COUNT = 2
@@ -239,6 +240,47 @@ async def test_persist_record_swallows_db_failure() -> None:
         )
 
 
+@pytest.mark.asyncio
+async def test_persist_record_uses_atomic_upsert_no_delete_then_insert() -> None:
+    """Wedge #3 (2026-06-27 live meltdown): the persist did delete-then-insert in
+    its own transaction. A respawn loop fires count 1->2->3->4 in quick
+    succession, scheduling a fire-and-forget persist per increment; two of those
+    for the same (agent_slug, task_id) race in separate transactions and the
+    loser's INSERT hit pk_respawn_tracker UniqueViolation, so the durable count
+    stuck at the first INSERT's value and a restart re-burned the strike
+    threshold — exactly the re-burn this feature was built to stop. The persist
+    must be a single atomic ON CONFLICT DO UPDATE upsert (no DELETE, no db.add):
+    concurrent upserts on the same key serialize at row level and never collide.
+    """
+    orch = _new_orchestrator()
+    db = AsyncMock()
+    db.execute = AsyncMock(return_value=MagicMock())
+    ctx = MagicMock()
+    ctx.__aenter__ = AsyncMock(return_value=db)
+    ctx.__aexit__ = AsyncMock(return_value=False)
+    factory = MagicMock(return_value=ctx)
+    with patch("roboco.db.base.get_session_factory", return_value=factory):
+        await orch._persist_respawn_record(
+            "be-pm",
+            str(uuid4()),
+            {
+                "count": 4,
+                "last_status": "pending",
+                "last_check": datetime.now(UTC),
+                "tracing_resets": 0,
+                "notified": True,
+            },
+        )
+    # One atomic statement, no ORM add, no delete.
+    assert db.execute.await_count == 1
+    db.add.assert_not_called()
+    stmt = db.execute.await_args.args[0]
+    sql = str(stmt.compile(dialect=postgresql.dialect()))
+    assert "ON CONFLICT" in sql
+    assert "DO UPDATE" in sql
+    assert "DELETE" not in sql.upper()
+
+
 # --------------------------------------------------------------------------- #
 # Safety regression + transparency
 # --------------------------------------------------------------------------- #