diff --git a/src/daemon/providers/claude/index.ts b/src/daemon/providers/claude/index.ts
index 1a5a8ec..35b9232 100644
--- a/src/daemon/providers/claude/index.ts
+++ b/src/daemon/providers/claude/index.ts
@@ -162,6 +162,10 @@ export class ClaudeProvider implements SessionProvider {
   /** Last TurnOpts — the retry after a skill approval rebuilds the loop with
    * these, then re-pushes #lastPushedContent into the SAME turn queue (#233). */
   #lastTurnOpts: TurnOpts | null = null;
+  /** One re-send per turn for an UNEXPLAINED zero-turn result. Reset in
+   * runTurn() so every new prompt gets its own single attempt, and in
+   * resetToNewSession() with the rest of the backing-session state. */
+  #zeroTurnRecoveryAttempted = false;
   /** Rendered transcript from seedFromHistory() — prepended to the next prompt. */
   #pendingHistorySeed: string | null = null;
 
@@ -239,6 +243,7 @@ export class ClaudeProvider implements SessionProvider {
     this.#claudeCodeSessionId = newBackingId;
     this.#hasQueried = false;
     this.#backingRecoveryAttempted = false;
+    this.#zeroTurnRecoveryAttempted = false;
     this.#lastPushedContent = null;
   }
 
@@ -253,6 +258,8 @@ export class ClaudeProvider implements SessionProvider {
     this.#currentRequestUserInput = opts.requestUserInput ?? null;
     this.#currentSender = opts.sender ?? null;
     this.#lastTurnOpts = opts;
+    // A genuinely new prompt earns its own single recovery attempt.
+    this.#zeroTurnRecoveryAttempted = false;
 
     this.#ensureQueryLoop(opts);
 
@@ -955,6 +962,64 @@ export class ClaudeProvider implements SessionProvider {
     this.#pushSDKMessage(this.#lastPushedContent, "later", true);
   }
 
+  /**
+   * Re-send a prompt the backend consumed without running it.
+   *
+   * A zero-turn result with NO `<local-command-stderr>` is the unexplained
+   * variant: the SDK reports `subtype: "success"`, the model never ran, and
+   * nothing says why. Observed after a long session idle, where the result
+   * lands ~2s after the prompt and immediately behind a fresh `system:init` —
+   * i.e. the loop had gone cold and the first prompt into the rebuilt one was
+   * swallowed. Before this, the prompt was simply lost and the user retyped it.
+   *
+   * Re-sending is safe *because* it is a zero-turn: `num_turns === 0` means no
+   * assistant turn ran, so no tool executed and there is no side effect to
+   * duplicate. That is what separates this from a normal failed turn, which we
+   * must never silently repeat.
+   *
+   * Deliberately keyed on `num_turns` — a typed field carrying correct data —
+   * rather than on any cause, so it recovers whatever produced the zero turn
+   * and does not depend on the upstream reporting defect being fixed
+   * (anthropics/claude-code#80223, still open).
+   *
+   * Scoped tightly: only when NO stderr explains it (a real skill-command
+   * denial is #233's parked-approval path, and re-sending it would just be
+   * denied again), and only once per turn.
+   *
+   * Returns true when the turn was re-sent — the caller then suppresses the
+   * translation, so no terminal `turn_done` is emitted and the Session keeps
+   * the turn open, exactly as the #233 park does.
+   */
+  #tryRecoverConsumedPrompt(): boolean {
+    const tag = `[claude-provider ${this.#init.sessionId.slice(0, 8)}]`;
+    // Falling through emits the real turn_done error, so the USER sees the
+    // failure either way. These lines are for the operator: they separate "the
+    // retry ran and also came back empty" from "never retried", which is the
+    // difference between a backend that is wedged and a one-off swallow.
+    if (this.#zeroTurnRecoveryAttempted) {
+      console.error(
+        `${tag} zero-turn again after the re-send — giving up, surfacing as a turn error`,
+      );
+      return false;
+    }
+    if (!this.#lastTurnOpts || this.#lastPushedContent === null) {
+      console.error(`${tag} zero-turn with no reported cause and no prompt to re-send`);
+      return false;
+    }
+    this.#zeroTurnRecoveryAttempted = true;
+    console.error(
+      `${tag} zero-turn with no reported cause — re-sending the consumed prompt (${this.#lastPushedContent.length}B)`,
+    );
+    this.#emit({
+      type: "custom_message",
+      role: "info",
+      content: "The backend consumed the prompt without running it — re-sending…",
+    });
+    this.#ensureQueryLoop(this.#lastTurnOpts);
+    this.#pushSDKMessage(this.#lastPushedContent, "later", true);
+    return true;
+  }
+
   /** End a parked skill turn as a clean error (denied / dismissed / broken). */
   #failSkillTurn(command: string, reason: string): void {
     const detail = command ? `: ${command}` : "";
@@ -994,6 +1059,12 @@ export class ClaudeProvider implements SessionProvider {
       this.#translateState.lastLocalCommandStderr = null;
       return;
     }
+    // The UNEXPLAINED zero-turn: same "prompt consumed, model never ran", but
+    // with no stderr naming a cause. Re-send it once rather than losing it.
+    // Same suppression contract as the skill-block branch above.
+    if (m.type === "result" && m.num_turns === 0 && !stderr && this.#tryRecoverConsumedPrompt()) {
+      return;
+    }
     translateSDKMessage(msg, this.#emit.bind(this), this.id, this.#translateState, (e) => this.onSessionEvent?.(e));
   }
 }
diff --git a/src/tests/provider-claude.test.ts b/src/tests/provider-claude.test.ts
index b97160a..1e343af 100644
--- a/src/tests/provider-claude.test.ts
+++ b/src/tests/provider-claude.test.ts
@@ -507,6 +507,92 @@ describe("skillCommandAllowRules", () => {
 // Park-and-retry: a blocked skill command must NOT fail the turn (that killed the
 // pipeline — #233). It parks the turn (approval_pending), and on approval the
 // loop rebuilds with the grant and retries the SAME prompt in place.
+/**
+ * The UNEXPLAINED zero-turn: `num_turns: 0` with no `<local-command-stderr>`
+ * naming a cause. The prompt was consumed and the model never ran, so before
+ * this the prompt was simply lost and the user retyped it.
+ *
+ * Distinct from #233 (a blocked skill command, which HAS a stderr cause and
+ * parks for approval) and from the upstream reporting defect
+ * (anthropics/claude-code#80223) — this recovers whatever produced the zero
+ * turn, because it keys on `num_turns` rather than on any cause.
+ */
+describe("ClaudeProvider – unexplained zero-turn re-send", () => {
+  beforeEach(() => { sdkMessages = []; sdkThrowError = null; capturedQueryOpts = null; sdkGate = null; });
+
+  const ZERO_TURN = { type: "result", subtype: "success", is_error: false, num_turns: 0, result: "" };
+  const REAL_TURN = { type: "result", subtype: "success", num_turns: 2, result: "done", modelUsage: {} };
+  const info = (events: ProviderEvent[]) =>
+    events.filter(
+      (e) => e.type === "custom_message" && String((e as { content?: string }).content).includes("re-sending"),
+    );
+  const zeroTurnErrors = (events: ProviderEvent[]) =>
+    events.filter(
+      (e) =>
+        e.type === "turn_done" &&
+        String((e as { result?: { errorMessage?: string } }).result?.errorMessage ?? "").includes(
+          "produced no turn",
+        ),
+    );
+
+  it("re-sends the consumed prompt instead of surfacing an error", async () => {
+    sdkMessages = [ZERO_TURN, REAL_TURN];
+    const events = await collectTurnEvents(makeProvider());
+    // The user is told, once, that the turn is being retried...
+    expect(info(events)).toHaveLength(1);
+    // ...and never sees the zero-turn error, because the turn recovered.
+    expect(zeroTurnErrors(events)).toHaveLength(0);
+    // The turn still ends normally, on the real result.
+    expect(events.some((e) => e.type === "turn_done")).toBe(true);
+  });
+
+  it("gives up after ONE attempt — a second zero-turn is a real error", async () => {
+    sdkMessages = [ZERO_TURN, ZERO_TURN];
+    const events = await collectTurnEvents(makeProvider());
+    expect(info(events)).toHaveLength(1); // not two
+    // The retry failed the same way, so the turn ends honestly rather than
+    // looping on a backend that will never run this prompt.
+    expect(zeroTurnErrors(events)).toHaveLength(1);
+  });
+
+  it("leaves a normal multi-turn result alone", async () => {
+    sdkMessages = [REAL_TURN];
+    const events = await collectTurnEvents(makeProvider());
+    expect(info(events)).toHaveLength(0);
+    expect(zeroTurnErrors(events)).toHaveLength(0);
+  });
+
+  it("does NOT touch a zero-turn that HAS a reported cause", async () => {
+    // A stderr cause that is not a skill-command pattern, so #233's parked
+    // approval declines it too — it must fall through to the honest error with
+    // the real reason, NOT be silently re-sent as if unexplained.
+    sdkMessages = [
+      {
+        type: "user",
+        message: { role: "user", content: "<local-command-stderr>Error: disk on fire</local-command-stderr>" },
+      },
+      ZERO_TURN,
+    ];
+    const events = await collectTurnEvents(makeProvider());
+    expect(info(events)).toHaveLength(0);
+    const errs = zeroTurnErrors(events);
+    expect(errs).toHaveLength(1);
+    expect(String((errs[0] as { result: { errorMessage?: string } }).result.errorMessage)).toContain(
+      "disk on fire",
+    );
+  });
+
+  it("gives each NEW prompt its own single attempt", async () => {
+    const provider = makeProvider();
+    sdkMessages = [ZERO_TURN, REAL_TURN];
+    expect(info(await collectTurnEvents(provider, "first"))).toHaveLength(1);
+    // The one-shot guard is per-turn, not per-session: a later prompt that hits
+    // the same cold-loop failure must still be recoverable.
+    sdkMessages = [ZERO_TURN, REAL_TURN];
+    expect(info(await collectTurnEvents(provider, "second"))).toHaveLength(1);
+  });
+});
+
 describe("ClaudeProvider – skill-command approval (#233)", () => {
   // A Store backed by a real Map, so a persisted grant is visible to the retry's
   // query rebuild (a fixed-map stub could never satisfy the retry).