diff --git a/src/agent_env/agents/agents/claude-code/a2a_server.py b/src/agent_env/agents/agents/claude-code/a2a_server.py
index 7dc27625e..5df5029ca 100644
--- a/src/agent_env/agents/agents/claude-code/a2a_server.py
+++ b/src/agent_env/agents/agents/claude-code/a2a_server.py
@@ -137,6 +137,9 @@ class ClaudeCodeConfig(AgentConfig):
     max_thinking_tokens: int | None = None
     # Which of the CLI's own tools it may use. With any selection, its MCP servers are only the ones agent-env gives it.
     builtin_tools: ClaudeCodeTools | None = None
+    # What a turn cut short by a transient API error is resumed with; unset, the bridge asks the model to finish the
+    # request it was given, which suits a task but not a simulated user.
+    resume_prompt: str | None = None
 
 
 def _rewrite_session_ids(jsonl_text: str, session_id: str) -> tuple[list[str], int]:
@@ -492,6 +495,7 @@ async def _turn(self, request: TaskRequest[ClaudeCodeConfig], cfg: ClaudeCodeCon
                 max_turns=cfg.max_turns,
                 max_thinking_tokens=cfg.max_thinking_tokens,
                 builtin_tools=cfg.builtin_tools,
+                resume_prompt=cfg.resume_prompt,
                 event_callback=request.trajectory.append,
             )
         except Exception as e:
diff --git a/src/agent_env/agents/agents/claude-code/claude_bridge.py b/src/agent_env/agents/agents/claude-code/claude_bridge.py
index cecd8286f..a17e3cc70 100644
--- a/src/agent_env/agents/agents/claude-code/claude_bridge.py
+++ b/src/agent_env/agents/agents/claude-code/claude_bridge.py
@@ -415,10 +415,12 @@ async def run_claude(
     max_thinking_tokens: int | None = None,
     builtin_tools: ClaudeCodeTools | None = None,
     event_callback: Callable[[dict], None] | None = None,
+    resume_prompt: str | None = None,
 ) -> ClaudeResult:
     """Run one turn, resuming the CLI session when a transient upstream failure cuts it short; a resume spends what
     is left of ``timeout_s``, so the turn never outlives its caller's deadline. ``CLAUDE_API_ERROR_RETRIES=0`` turns
-    resumes off. ``event_callback`` gets each stdout event as the CLI prints it; an attempt's trailing `result` is
+    resumes off. The resume is prompted with ``resume_prompt``, else ``_RESUME_PROMPT``, which asks the model to
+    finish the request it was given: a caller whose model isn't doing a task, such as a simulated user, sets its own. ``event_callback`` gets each stdout event as the CLI prints it; an attempt's trailing `result` is
     held until it's known not to be resumed."""
     deadline = time.monotonic() + timeout_s
     retries_left = _configured_api_error_retries()
@@ -468,7 +470,7 @@ async def run_claude(
         trajectory.extend(_without_terminal_result(result.trajectory))
         first_failure = first_failure or result.error_detail
         retries_left -= 1
-        attempt_prompt, attempt_resume, attempt_session = _RESUME_PROMPT, True, result.session_id
+        attempt_prompt, attempt_resume, attempt_session = resume_prompt or _RESUME_PROMPT, True, result.session_id
         backoff = _RETRY_BACKOFF_S[min(attempts - 1, len(_RETRY_BACKOFF_S) - 1)]
         logger.warning(f"claude CLI turn cut short ({_redact(failure)!r}); resuming session {attempt_session} "
                        f"in {backoff}s with {int(budget_left)}s of the original {timeout_s}s budget left")
diff --git a/src/agent_env/task_step/task_steps/prompt_agent.py b/src/agent_env/task_step/task_steps/prompt_agent.py
index c11fb716b..9e8026628 100644
--- a/src/agent_env/task_step/task_steps/prompt_agent.py
+++ b/src/agent_env/task_step/task_steps/prompt_agent.py
@@ -115,6 +115,15 @@ def _file_uris(parts: list[dict]) -> list[str]:
 # The decide body that asks a user agent to record nothing when no agent trigger fires.
 _QUIET_DECIDE = ("turn", "solver_message", "context_id", "env_triggers", "record_unfired")
 
+# What a user-sim turn cut short by a transient API error is resumed with, for an agent that takes `resume_prompt`.
+# An agent's own resume prompt asks it to finish the request it was given; a user-sim's request is the agent's reply,
+# so that prompt has it carry on as the agent.
+USER_SIM_RESUME_PROMPT = (
+    "Your previous turn was interrupted by a transient API error before you sent your message. You are still the "
+    "user. Write the user's next message, as you were about to; the assistant's last reply above is what you are "
+    "answering."
+)
+
 _DEFAULT_USER_SIM_OUTPUT_FORMAT: dict[str, Any] = {
     "type": "json_schema",
     "schema": {
@@ -503,7 +512,8 @@ async def _converse(
         if is_user_sim:
             user_output_format = self.user_output_format or _DEFAULT_USER_SIM_OUTPUT_FORMAT
             user_model = self._resolve_user_model(context)
-            user_config = {k: v for k, v in {"output_format": user_output_format, "model": user_model}.items() if v is not None}
+            user_config = {k: v for k, v in {"output_format": user_output_format, "model": user_model,
+                                             "resume_prompt": USER_SIM_RESUME_PROMPT}.items() if v is not None}
             negotiated = A2AAgent.negotiate_agent_config(user_sim.a2a_card or {}, user_config)
             if negotiated:
                 await protocol.post_agent_config(user_url + negotiated.endpoint, negotiated.fields)
diff --git a/tst/unit/agents/claude_code/test_card_contract.py b/tst/unit/agents/claude_code/test_card_contract.py
index 0ace58b4f..95b15f08e 100644
--- a/tst/unit/agents/claude_code/test_card_contract.py
+++ b/tst/unit/agents/claude_code/test_card_contract.py
@@ -68,7 +68,7 @@ def test_the_config_takes_every_field_prompt_agent_sends_that_the_cli_has(app):
     config = _extensions(app)["urn:agentenv:agent-config/v1"]["params"]["methods"]
     assert config["set"]["request"]["supported"] == [
         "allowed_tools", "builtin_tools", "description", "effort", "max_thinking_tokens", "max_turns", "model", "name",
-        "output_format", "project_id", "role", "system_prompt", "task_id", "timeout_seconds",
+        "output_format", "project_id", "resume_prompt", "role", "system_prompt", "task_id", "timeout_seconds",
     ]
     assert set(config["set"]["request"]["schema"]["properties"]) == set(config["set"]["request"]["supported"])
     assert config["get"] == {"method": "GET", "response": {"required": ["config"]}}
diff --git a/tst/unit/agents/claude_code/test_claude_bridge_retry.py b/tst/unit/agents/claude_code/test_claude_bridge_retry.py
index 35e1999ea..4ff6210f3 100644
--- a/tst/unit/agents/claude_code/test_claude_bridge_retry.py
+++ b/tst/unit/agents/claude_code/test_claude_bridge_retry.py
@@ -61,6 +61,18 @@ def _write_session(scripted, session_id="sess-1"):
     (scripted["session_dir"] / f"{session_id}.jsonl").write_text('{"type":"assistant"}\n')
 
 
+def test_a_resume_prompt_given_replaces_the_default_one(scripted):
+    _write_session(scripted)
+    scripted["queue"].append((_api_error_stdout("API Error: The operation timed out.", tool_uses=1), 1))
+    scripted["queue"].append((_ok_stdout("resumed"), 0))
+
+    result = asyncio.run(claude_bridge.run_claude(prompt="p", session_id="sess-1", resume=False,
+                                                  resume_prompt="stay the user"))
+
+    assert result.attempts == 2
+    assert scripted["procs"][1].stdin.data == b"stay the user"
+
+
 def test_success_makes_exactly_one_invocation(scripted):
     scripted["queue"].append((_ok_stdout(), 0))
     result = asyncio.run(claude_bridge.run_claude(prompt="p", session_id="sess-1", resume=False))
diff --git a/tst/unit/task_step/test_prompt_agent_user_sim_resume.py b/tst/unit/task_step/test_prompt_agent_user_sim_resume.py
new file mode 100644
index 000000000..df1862746
--- /dev/null
+++ b/tst/unit/task_step/test_prompt_agent_user_sim_resume.py
@@ -0,0 +1,74 @@
+"""A user simulator is sent the resume prompt that keeps it the user when its card takes one: an agent resumes a
+turn cut short by asking its model to finish the request, and a user simulator's request is the agent's reply."""
+
+from unittest.mock import MagicMock
+
+import pytest
+from a2a.types import TaskState
+
+from agent_env.a2a_agent import A2AAgent, protocol
+from agent_env.task_step.context import DeployedAgent, TaskStepContext
+from agent_env.task_step.task_steps import prompt_agent as pa_mod
+from agent_env.task_step.task_steps.prompt_agent import USER_SIM_RESUME_PROMPT, PromptAgentTaskStep
+
+
+def _card(*supported: str) -> dict:
+    return {"capabilities": {"extensions": [{"uri": A2AAgent.EXT_AGENT_CONFIG, "params": {
+        "endpoint": "/ext/agent-config", "methods": {"set": {"request": {"supported": list(supported)}}}}}]}}
+
+
+@pytest.fixture
+def posted(monkeypatch):
+    """Mocks the conversation's HTTP and store boundary, and collects each agent config posted, by URL."""
+    cfg = MagicMock()
+    cfg.get_model_params.return_value = {}
+    cfg.get_model_for_role.return_value = None
+    monkeypatch.setattr(pa_mod, "get_config", lambda: cfg)
+    for fn in ("create_conversation", "add_a2a_task", "complete_a2a_task", "mark_closed", "get_conversation"):
+        monkeypatch.setattr(pa_mod.conversation_store, fn, lambda *a, **kw: None)
+    configs = {}
+
+    async def post_agent_config(url, fields):
+        configs[url] = fields
+
+    async def send(url, parts, message_id, context_id, timeout):
+        return f"task-{message_id}", None
+
+    async def poll(url, task_id, timeout, interval, **_):
+        text = '{"message": "thanks", "done": true}' if "user-sim" in url else "done"
+        return {"status": {"state": TaskState.completed, "message": {"parts": [{"kind": "text", "text": text}]}}}
+
+    monkeypatch.setattr(protocol, "post_agent_config", post_agent_config)
+    monkeypatch.setattr(pa_mod.protocol, "send_a2a_message", send)
+    monkeypatch.setattr(pa_mod.protocol, "poll_a2a_task", poll)
+    return configs
+
+
+def _conversation(solver_card: dict, user_card: dict) -> TaskStepContext:
+    context = TaskStepContext(instance_id="ti-1")
+    context.deployed_agents.append(DeployedAgent(agent_name="solver", api_url="http://solver",
+                                                 a2a_url="http://solver", a2a_card=solver_card))
+    context.deployed_agents.append(DeployedAgent(agent_name="usersim", api_url="http://user-sim",
+                                                 a2a_url="http://user-sim", a2a_card=user_card))
+    return context
+
+
+def _step() -> PromptAgentTaskStep:
+    return PromptAgentTaskStep(id="solve", version=None, prompt="hi", agent_name="solver", model="task-solver",
+                               max_conversation_turns=2, user_agent_name="usersim", user_model="task-user")
+
+
+@pytest.mark.asyncio
+async def test_a_user_simulator_whose_card_takes_a_resume_prompt_is_sent_the_one_that_keeps_it_the_user(posted):
+    card = _card("model", "output_format", "resume_prompt")
+    await _step().execute(_conversation(card, card))
+
+    assert posted["http://user-sim/ext/agent-config"]["resume_prompt"] == USER_SIM_RESUME_PROMPT
+    assert "resume_prompt" not in posted["http://solver/ext/agent-config"]
+
+
+@pytest.mark.asyncio
+async def test_a_user_simulator_whose_card_takes_no_resume_prompt_is_sent_none(posted):
+    await _step().execute(_conversation(_card("model"), _card("model", "output_format")))
+
+    assert "resume_prompt" not in posted["http://user-sim/ext/agent-config"]