[aidan] fix/agent-errors: surface provider rate limits

backend/apps/agents/agent_manager.py

@@ -1520,6 +1520,7 @@ class AgentManager:
             # exit code 1 / Check stderr output for details", which masks
             # transient capacity issues.
             _stderr_buffer: list[str] = []
+            _system_event_buffer: list[str] = []
 
             def _stderr_cb(line: str) -> None:
                 _stderr_buffer.append(line)
@@ -2376,6 +2377,12 @@ class AgentManager:
                     # Log system messages (MCP server status, errors, etc.)
                     if isinstance(message, SystemMessage):
                         raw = message.__dict__ if hasattr(message, '__dict__') else str(message)
+                        try:
+                            _system_event_buffer.append(json.dumps(raw, default=str))
+                            if len(_system_event_buffer) > 200:
+                                del _system_event_buffer[:100]
+                        except Exception:
+                            pass
                         logger.info(f"[MCP-DEBUG] SystemMessage: {raw}")
 
                     if isinstance(message, StreamEvent):
@@ -2949,6 +2956,7 @@ class AgentManager:
                         _current_turn_emitted = False
                         await asyncio.sleep(wait)
                         _stderr_buffer.clear()
+                        _system_event_buffer.clear()
                         if session.sdk_session_id:
                             options_kwargs["resume"] = session.sdk_session_id
                             options = ClaudeAgentOptions(**options_kwargs)
@@ -2991,7 +2999,10 @@ class AgentManager:
             # user can't recover by waiting, this is a tier-gate, not a rate
             # limit, so the UX matters.
             try:
-                _stderr_tail = "\n".join(_stderr_buffer[-50:])
+                _stderr_tail = "\n".join([
+                    "\n".join(_stderr_buffer[-50:]),
+                    "\n".join(_system_event_buffer[-50:]),
+                ])
             except Exception:
                 _stderr_tail = ""
             # If we already streamed a substantive assistant response this
@@ -3169,6 +3180,31 @@ class AgentManager:
                     "session_id": session_id,
                     "message": error_msg.model_dump(mode="json"),
                 })
+            elif _is_transient_capacity_error(e, extra_text=_stderr_tail):
+                friendly_msg = (
+                    "provider_rate_limit: This model hit your account or "
+                    "session rate limit. Wait until the reset time shown by "
+                    "your provider, then send your message again, or switch "
+                    "to a different model."
+                )
+                try:
+                    from backend.apps.service.client import submit_diagnostic
+                    submit_diagnostic({
+                        "kind": "model_error",
+                        "subkind": "rate_limit",
+                        "model": session.model,
+                        "provider": session.provider,
+                        "connection_mode": getattr(load_settings(), "connection_mode", "own_key"),
+                        "error_preview": (f"{e!s}\n{_stderr_tail}")[:600],
+                    })
+                except Exception:
+                    logger.debug("submit_diagnostic transient_capacity failed", exc_info=True)
+                error_msg = Message(role="assistant", content=friendly_msg, branch_id=session.active_branch_id)
+                session.messages.append(error_msg)
+                await ws_manager.send_to_session(session_id, "agent:message", {
+                    "session_id": session_id,
+                    "message": error_msg.model_dump(mode="json"),
+                })
             else:
                 # Track unclassified agent failures too so we stop flying blind on them.
                 try:

backend/tests/test_free_trial.py

@@ -48,6 +48,17 @@ def test_exhaustion_is_classified_and_not_retried():
     assert not _is_transient_capacity_error(Exception("free_trial_exhausted"))
 
 
+def test_generic_cli_failure_uses_sdk_system_events_for_rate_limits():
+    system_event_tail = (
+        '{"subtype":"api_retry","data":{"error_status":429,'
+        '"error":"rate_limit","max_retries":10}}'
+    )
+    assert _is_transient_capacity_error(
+        Exception("Command failed with exit code 1"),
+        extra_text=system_event_tail,
+    )
+
+
 def test_has_own_model_never_shadows_a_real_provider():
     assert not _has_own_model(AppSettings(connection_mode="free-trial", free_trial_token="x"))
     assert not _has_own_model(AppSettings())

frontend/src/app/pages/AgentChat/bubbles/MessageBubble.tsx

@@ -111,7 +111,17 @@ function parseOpenSwarmError(text: string, ctx?: OverflowContext): OpenSwarmErro
       ctaAction: 'upgrade',
     };
   }
-  // Transient throttle: Anthropic's upstream 429/overload or our own pool-shed. Not the user's
+  if (/provider_rate_limit|account'?s rate limit|session rate limit|This request would exceed your account'?s rate limit/i.test(text)) {
+    const reset = text.match(/reset after ([^)\\.]+)/i)?.[1];
+    return {
+      kind: 'network',
+      title: "You've hit this model's rate limit",
+      detail: reset
+        ? `This model can send more requests after ${reset}. Wait for that reset window, or switch to another model.`
+        : 'Wait for the reset window shown by your provider, or switch to another model.',
+    };
+  }
+  // Transient throttle: Anthropic's upstream overload or our own pool-shed. Not the user's
   // fault and not a plan cap, so don't say "upgrade", just tell them it's busy. claude.ai-style.
   if (/rate_limit_error|free_pool_busy|overloaded_error|too many requests/i.test(text)) {
     return {