fix(proxy): recover compact bridge continuations

app/modules/proxy/_service/http_bridge/request_submit.py

@@ -106,6 +106,7 @@
     _prepare_websocket_request_state_for_auth_replay,
     _prepare_websocket_request_state_for_visible_output_replay,
     _prewarm_response_timeout_seconds,
+    _protected_agent_control_tool_call_ids,
     _release_websocket_response_create_gate,
     _response_create_client_metadata,
     _security_work_advisory_event,
@@ -605,6 +606,7 @@ def _prepare_response_bridge_request_state(
                 deduped_replayed_input_count = len(replayed_input_items)
                 deduped_replayed_input_fingerprint = _fingerprint_input_items(replayed_input_items)
                 payload = payload.model_copy(update={"input": deduped_input_items})
+        protected_tool_output_call_ids = _protected_agent_control_tool_call_ids(payload.input)
         upstream_payload = dict(payload.to_payload())
         upstream_payload.pop("stream", None)
         upstream_payload.pop("background", None)
@@ -684,6 +686,7 @@ def _prepare_response_bridge_request_state(
             slimmed_payload, slim_summary = _slim_response_create_payload_for_upstream(
                 upstream_payload,
                 max_bytes=max_bytes,
+                protected_tool_output_call_ids=protected_tool_output_call_ids,
             )
             if slim_summary is not None:
                 upstream_payload = slimmed_payload

app/modules/proxy/_service/http_bridge/service_stubs.py

@@ -24,6 +24,9 @@
 from app.core.types import JsonValue
 from app.core.utils.sse import CODEX_KEEPALIVE_FRAME
 from app.db.models import Account, DashboardSettings
+from app.modules.proxy._service.support import (
+    _protected_agent_control_tool_call_ids as _support_protected_agent_control_tool_call_ids,
+)
 from app.modules.proxy._service.support import _RequestLogFailureMetadata
 
 T = TypeVar("T")
@@ -336,6 +339,10 @@ def _slim_response_create_payload_for_upstream(*args: Any, **kwargs: Any) -> Any
     return _service_global("_slim_response_create_payload_for_upstream")(*args, **kwargs)
 
 
+def _protected_agent_control_tool_call_ids(*args: Any, **kwargs: Any) -> Any:
+    return _support_protected_agent_control_tool_call_ids(*args, **kwargs)
+
+
 def _enforce_response_create_size_limit(*args: Any, **kwargs: Any) -> Any:
     return _service_global("_enforce_response_create_size_limit")(*args, **kwargs)
 

app/modules/proxy/_service/response_create.py

@@ -33,6 +33,7 @@
     _PENDING_TOOL_CALL_ITEM_TYPES,
     _PENDING_TOOL_CALL_OUTPUT_ITEM_TYPE_BY_CALL_TYPE,
     _PENDING_TOOL_CALL_OUTPUT_ITEM_TYPES,
+    _protected_agent_control_tool_call_ids,
     _WebSocketRequestState,
 )
 
@@ -49,6 +50,7 @@
     "[codex-lb omitted historical tool output ({bytes} bytes) to fit upstream websocket budget]"
 )
 _RESPONSE_CREATE_IMAGE_OMISSION_NOTICE = "[codex-lb omitted historical inline image to fit upstream websocket budget]"
+_NO_PROTECTED_TOOL_OUTPUT_CALL_IDS: frozenset[str] = frozenset()
 _OVERSIZED_RESPONSE_CREATE_DUMP_DIR: Path | None = None
 _RESPONSE_CREATE_DUMP_SUFFIX = ".response-create.json.gz"
 _RESPONSE_CREATE_META_SUFFIX = ".meta.json"
@@ -212,6 +214,7 @@ def _response_create_text_with_size_guard(
     request_state: _WebSocketRequestState,
     transport: str,
 ) -> str | None:
+    protected_tool_output_call_ids = _protected_agent_control_tool_call_ids(payload.input)
     upstream_payload = dict(payload.to_payload())
     upstream_payload.pop("stream", None)
     upstream_payload.pop("background", None)
@@ -238,6 +241,7 @@ def _response_create_text_with_size_guard(
         slimmed_payload, slim_summary = slim_payload_for_upstream(
             upstream_payload,
             max_bytes=max_bytes,
+            protected_tool_output_call_ids=protected_tool_output_call_ids,
         )
         if slim_summary is not None:
             upstream_payload = slimmed_payload
@@ -302,6 +306,7 @@ def _slim_response_create_payload_for_upstream(
     payload: dict[str, JsonValue],
     *,
     max_bytes: int,
+    protected_tool_output_call_ids: set[str] | None = None,
 ) -> tuple[dict[str, JsonValue], dict[str, int] | None]:
     input_value = payload.get("input")
     if not isinstance(input_value, list) or not input_value:
@@ -315,13 +320,18 @@ def _slim_response_create_payload_for_upstream(
     tool_outputs_slimmed = 0
     images_slimmed = 0
 
+    if protected_tool_output_call_ids is None:
+        protected_tool_output_call_ids = _protected_agent_control_tool_call_ids(historical)
     slimmed_historical: list[JsonValue] = []
     for item in historical:
         (
             slimmed_item,
             item_tool_outputs_slimmed,
             item_images_slimmed,
-        ) = _slim_historical_response_input_item(item)
+        ) = _slim_historical_response_input_item(
+            item,
+            protected_tool_output_call_ids=protected_tool_output_call_ids,
+        )
         tool_outputs_slimmed += item_tool_outputs_slimmed
         images_slimmed += item_images_slimmed
         slimmed_historical.append(slimmed_item)
@@ -449,7 +459,11 @@ def _response_create_recent_suffix_start(input_items: list[JsonValue]) -> int:
     return 0
 
 
-def _slim_historical_response_input_item(item: JsonValue) -> tuple[JsonValue, int, int]:
+def _slim_historical_response_input_item(
+    item: JsonValue,
+    *,
+    protected_tool_output_call_ids: set[str] | None = None,
+) -> tuple[JsonValue, int, int]:
     if not is_json_mapping(item):
         return item, 0, 0
 
@@ -459,6 +473,10 @@ def _slim_historical_response_input_item(item: JsonValue) -> tuple[JsonValue, in
 
     item_type = item_mapping.get("type")
     if isinstance(item_type, str) and item_type in _PENDING_TOOL_CALL_OUTPUT_ITEM_TYPES:
+        call_id = item_mapping.get("call_id")
+        protected_call_ids = protected_tool_output_call_ids or _NO_PROTECTED_TOOL_OUTPUT_CALL_IDS
+        if isinstance(call_id, str) and call_id in protected_call_ids:
+            return item_mapping, tool_outputs_slimmed, images_slimmed
         output = item_mapping.get("output")
         if isinstance(output, str):
             if _should_slim_historical_tool_output(output):

app/modules/proxy/_service/support.py

@@ -66,6 +66,8 @@
 _PENDING_TOOL_CALL_ITEM_TYPES = frozenset(_PENDING_TOOL_CALL_OUTPUT_ITEM_TYPE_BY_CALL_TYPE)
 _PENDING_TOOL_CALL_OUTPUT_ITEM_TYPES = frozenset(_PENDING_TOOL_CALL_OUTPUT_ITEM_TYPE_BY_CALL_TYPE.values())
 _TTFT_OUTPUT_ITEM_TYPES = _PENDING_TOOL_CALL_ITEM_TYPES - {"function_call"}
+_AGENT_CONTROL_TOOL_NAMESPACES = frozenset({"collaboration", "multi_agent_v1"})
+_AGENT_CONTROL_TOOL_NAMES = frozenset({"close_agent", "resume_agent", "send_input", "spawn_agent", "wait_agent"})
 _WEBSOCKET_FULL_REPLAY_WAIT_MIN_ITEMS = 20
 _WEBSOCKET_FULL_REPLAY_WAIT_POLL_SECONDS = 0.05
 _HARD_HTTP_BRIDGE_AFFINITY_KINDS = frozenset(
@@ -78,6 +80,34 @@
     }
 )
 _ACCOUNT_SELECTION_RECOVERY_MIN_SLEEP_SECONDS = 1.0
+
+
+def _protected_agent_control_tool_call_ids(input_value: JsonValue) -> set[str]:
+    if not isinstance(input_value, list):
+        return set()
+    call_ids: set[str] = set()
+    for item in input_value:
+        if not isinstance(item, Mapping):
+            continue
+        item_type = item.get("type")
+        if not isinstance(item_type, str) or item_type not in _PENDING_TOOL_CALL_ITEM_TYPES:
+            continue
+        namespace = item.get("namespace")
+        name = item.get("name")
+        is_agent_control = (
+            isinstance(namespace, str)
+            and namespace in _AGENT_CONTROL_TOOL_NAMESPACES
+            or isinstance(name, str)
+            and name in _AGENT_CONTROL_TOOL_NAMES
+        )
+        if not is_agent_control:
+            continue
+        call_id = item.get("call_id")
+        if isinstance(call_id, str) and call_id:
+            call_ids.add(call_id)
+    return call_ids
+
+
 _HARD_AFFINITY_RECOVERY_SLEEP_SECONDS = 2.0
 _ACCOUNT_SELECTION_RECOVERY_DEFAULT_SLEEP_SECONDS = 30.0
 _ACCOUNT_SELECTION_RECOVERY_MAX_SLEEP_SECONDS = 300.0

app/modules/proxy/_service/websocket/helpers.py

@@ -1609,6 +1609,21 @@ def _matching_websocket_request_states_for_previous_response_error(
         ]
         if matching_requests:
             return matching_requests
+    visible_unresolved_followups = [
+        request_state
+        for request_state in followup_requests
+        if request_state.response_id is None and _http_bridge_request_counts_against_queue(request_state)
+    ]
+    if len(visible_unresolved_followups) == 1:
+        return visible_unresolved_followups
+    if len(visible_unresolved_followups) > 1:
+        unique_previous_response_ids = {
+            request_state.previous_response_id
+            for request_state in visible_unresolved_followups
+            if request_state.previous_response_id
+        }
+        if len(unique_previous_response_ids) == 1:
+            return visible_unresolved_followups
     unresolved_followups = [request_state for request_state in followup_requests if request_state.response_id is None]
     if len(unresolved_followups) == 1:
         return unresolved_followups

openspec/changes/archive/2026-08-13-fix-replayed-namespaced-function-call/specs/responses-api-compat/spec.md

@@ -1,13 +1,15 @@
 ## ADDED Requirements
 
-### Requirement: Replayed tool-call namespace metadata is local-only on upstream input
+### Requirement: Replayed tool-call namespace metadata survives live Responses egress
 
-For standard and compact Responses requests, the proxy MUST omit `namespace` from every replayed `input` item whose `type` is `function_call`, `custom_tool_call`, or `apply_patch_call` before forwarding the request upstream. The proxy MUST preserve all other fields on that item, MUST retain the original namespace metadata for local call-identity and replay-deduplication processing, and MUST NOT alter client-provided top-level tool entries as part of this normalization.
+For standard live Responses requests, including WebSocket `response.create` and configured Responses model-source egress, the proxy MUST preserve `namespace` on replayed `input` items whose `type` is `function_call`, `custom_tool_call`, or `apply_patch_call`. Tool namespaces are part of the live tool-routing identity. Compact Responses egress MAY omit those namespaces only as part of the compact-specific upstream compatibility serializer. The proxy MUST preserve all other fields on each tool-call item and MUST NOT alter client-provided top-level tool entries as part of this normalization.
 
-#### Scenario: Standard Responses replay omits tool-call namespaces upstream
+Historical response.create slimming MUST preserve outputs for namespaced agent-control calls, including `collaboration` and `multi_agent_v1` calls, even when those outputs are large. Such outputs carry the completed spawn/wait state needed by the next model turn and MUST NOT be replaced with a generic historical tool-output omission notice. Unrelated historical tool outputs MAY still be slimmed under the normal payload-budget policy.
+
+#### Scenario: Standard Responses replay preserves tool-call namespaces upstream
 
 - **WHEN** a standard Responses request replays `function_call` and `custom_tool_call` input items with `namespace`
-- **THEN** the upstream payload omits only those items' `namespace`
+- **THEN** the upstream payload preserves those items' `namespace`
 - **AND** preserves their remaining call fields
 - **AND** the local request input retains the namespace metadata
 
@@ -17,16 +19,16 @@ For standard and compact Responses requests, the proxy MUST omit `namespace` fro
 - **THEN** its upstream payload omits the input item's `namespace`
 - **AND** preserves the remaining tool-call fields
 
-#### Scenario: WebSocket response.create omits tool-call namespaces upstream
+#### Scenario: WebSocket response.create preserves tool-call namespaces upstream
 
 - **WHEN** a Responses WebSocket request replays namespaced `function_call` and `custom_tool_call` input items
-- **THEN** the upstream `response.create` frame omits only those items' `namespace`
+- **THEN** the upstream `response.create` frame preserves those items' `namespace`
 - **AND** preserves their remaining call fields
 
-#### Scenario: Configured Responses model source omits tool-call namespaces upstream
+#### Scenario: Configured Responses model source preserves tool-call namespaces upstream
 
 - **WHEN** `/v1/responses` routes a replayed namespaced tool call to a configured OpenAI-compatible Responses model source
-- **THEN** the source payload omits only the call item's `namespace`
+- **THEN** the source payload preserves the call item's `namespace`
 - **AND** preserves source-compatible request fields that the Codex upstream path does not support
 
 #### Scenario: Account-neutral replay classification retains namespace identity
@@ -45,3 +47,10 @@ For standard and compact Responses requests, the proxy MUST omit `namespace` fro
 
 - **WHEN** the client includes a top-level tool entry whose `type` is `namespace`
 - **THEN** standard Responses serialization forwards that tool entry byte-identically
+
+#### Scenario: Historical agent wait output remains visible after slimming
+
+- **WHEN** a response.create payload contains a historical `multi_agent_v1.wait_agent` call with a large matching output
+- **AND** also contains an unrelated large historical shell output
+- **THEN** the agent wait output remains byte-preserved in the upstream input
+- **AND** the unrelated shell output MAY be replaced with the historical tool-output omission notice

openspec/specs/responses-api-compat/spec.md

@@ -3263,6 +3263,8 @@ For a namespaced side-effect function or custom-tool call, the service MUST use
 
 Flat legacy side-effect calls MAY continue to use argument-based replay identity so reconnects that change only a call ID do not repeat shell, patch, or terminal side effects.
 
+Historical response.create slimming MUST preserve outputs for namespaced agent-control calls, including `collaboration` and `multi_agent_v1` calls, even when those outputs are large. Such outputs carry the completed spawn/wait state needed by the next model turn and MUST NOT be replaced with a generic historical tool-output omission notice. Unrelated historical tool outputs MAY still be slimmed under the normal payload-budget policy.
+
 #### Scenario: Distinct namespaced spawns use identical arguments
 
 - **WHEN** two `collaboration.spawn_agent` calls have identical arguments and different call IDs