feat(models): STRIX_STREAM_MODE opt-in non-streaming for custom endpoints

Replace the api_base-based non-streaming heuristic with an explicit
STRIX_STREAM_MODE=auto|always|never setting. auto/always stream (unchanged
default); never routes through the non-streaming wrapper for endpoints whose
streamed responses drop tool calls.

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "strix-agent"
-version = "1.4.0"
+version = "1.4.1"
 description = "Open-source AI Hackers for your apps"
 readme = "README.md"
 license = "Apache-2.0"

strix/config/codex.py

@@ -18,12 +18,12 @@ import logging
 import secrets
 import threading
 import time
-import urllib.error
 import urllib.parse
-import urllib.request
 from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
+import requests
+
 
 if TYPE_CHECKING:
     from collections.abc import Iterator
@@ -221,26 +221,19 @@ def _first(query: dict[str, list[str]], key: str) -> str | None:
 
 
 def _post_form(payload: dict[str, str]) -> dict[str, Any]:
-    body = urllib.parse.urlencode(payload).encode("ascii")
-    request = urllib.request.Request(  # noqa: S310 - fixed https OAuth endpoint
-        TOKEN_URL,
-        data=body,
-        headers={
-            "Content-Type": "application/x-www-form-urlencoded",
-            "Accept": "application/json",
-        },
-        method="POST",
-    )
     try:
-        with urllib.request.urlopen(  # noqa: S310  # nosec B310 - fixed https endpoint
-            request, timeout=_TOKEN_TIMEOUT
-        ) as response:
-            data = json.loads(response.read() or b"{}")
-    except urllib.error.HTTPError as exc:
-        detail = exc.read().decode("utf-8", "replace")[:300]
-        raise CodexAuthError("token_http_error", f"HTTP {exc.code}: {detail}") from exc
-    except (urllib.error.URLError, TimeoutError, OSError) as exc:
+        response = requests.post(
+            TOKEN_URL,
+            data=payload,
+            headers={"Accept": "application/json"},
+            timeout=_TOKEN_TIMEOUT,
+        )
+    except requests.RequestException as exc:
         raise CodexAuthError("unavailable", str(exc)) from exc
+    if response.status_code >= 400:
+        detail = response.text[:300]
+        raise CodexAuthError("token_http_error", f"HTTP {response.status_code}: {detail}")
+    data = json.loads(response.content or b"{}")
     if not isinstance(data, dict):
         raise CodexAuthError("bad_response", "token endpoint returned non-object")
     return data

strix/config/models.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 import contextlib
 import inspect
 import os
+import time
 from typing import TYPE_CHECKING, Any
 
 from agents import (
@@ -13,6 +14,8 @@ from agents import (
     set_tracing_disabled,
 )
 from agents.model_settings import ModelSettings
+from agents.models.fake_id import FAKE_RESPONSES_ID
+from agents.models.interface import Model
 from agents.models.multi_provider import MultiProvider
 from agents.models.openai_responses import OpenAIResponsesModel
 from agents.retry import (
@@ -21,6 +24,12 @@ from agents.retry import (
     RetryPolicyContext,
     retry_policies,
 )
+from openai.types.responses import (
+    Response,
+    ResponseCompletedEvent,
+    ResponseOutputItemDoneEvent,
+    ResponseUsage,
+)
 from openai.types.shared import Reasoning
 
 from strix.config import codex
@@ -30,8 +39,14 @@ from strix.config.loader import load_settings
 if TYPE_CHECKING:
     from collections.abc import AsyncIterator
 
-    from agents.models.interface import Model, ModelProvider
+    from agents.agent_output import AgentOutputSchemaBase
+    from agents.handoffs import Handoff
+    from agents.items import ModelResponse, TResponseInputItem, TResponseStreamEvent
+    from agents.models.interface import ModelProvider, ModelTracing
+    from agents.tool import Tool
+    from agents.usage import Usage
     from openai import AsyncOpenAI
+    from openai.types.responses import ResponsePromptParam
 
     from strix.config.settings import ReasoningEffort, Settings
 
@@ -135,6 +150,99 @@ class _CodexResponsesModel(OpenAIResponsesModel):
                     await result
 
 
+def _to_response_usage(usage: Usage | None) -> ResponseUsage | None:
+    if usage is None:
+        return None
+    return ResponseUsage(
+        input_tokens=usage.input_tokens,
+        output_tokens=usage.output_tokens,
+        total_tokens=usage.total_tokens,
+        input_tokens_details=usage.input_tokens_details,
+        output_tokens_details=usage.output_tokens_details,
+    )
+
+
+class _NonStreamingModel(Model):
+    """Run a model non-streamed but expose the streaming interface the runner uses.
+
+    Some OpenAI-compatible endpoints (notably gateways serving reasoning models)
+    return valid ``tool_calls`` for a non-streamed completion but, when streamed,
+    emit the tool call as plain text or drop it and close the stream — leaving
+    Strix's tool-driven loop with nothing to execute. Selecting
+    ``STRIX_STREAM_MODE=never`` routes through this wrapper, which makes the real
+    request non-streamed (where tool calling works) and synthesizes the minimal
+    event sequence the runner consumes from a stream, so the rest of the pipeline
+    is unchanged. The only user-visible difference is no token-by-token output.
+    """
+
+    def __init__(self, inner: Model) -> None:
+        self._inner = inner
+
+    async def get_response(self, *args: Any, **kwargs: Any) -> ModelResponse:
+        return await self._inner.get_response(*args, **kwargs)
+
+    async def stream_response(
+        self,
+        system_instructions: str | None,
+        input: str | list[TResponseInputItem],  # noqa: A002
+        model_settings: ModelSettings,
+        tools: list[Tool],
+        output_schema: AgentOutputSchemaBase | None,
+        handoffs: list[Handoff],
+        tracing: ModelTracing,
+        *,
+        previous_response_id: str | None = None,
+        conversation_id: str | None = None,
+        prompt: ResponsePromptParam | None = None,
+    ) -> AsyncIterator[TResponseStreamEvent]:
+        model_response = await self._inner.get_response(
+            system_instructions,
+            input,
+            model_settings,
+            tools,
+            output_schema,
+            handoffs,
+            tracing,
+            previous_response_id=previous_response_id,
+            conversation_id=conversation_id,
+            prompt=prompt,
+        )
+
+        sequence = 0
+        for index, item in enumerate(model_response.output):
+            yield ResponseOutputItemDoneEvent(
+                item=item,
+                output_index=index,
+                type="response.output_item.done",
+                sequence_number=sequence,
+            )
+            sequence += 1
+
+        response = Response(
+            id=model_response.response_id or FAKE_RESPONSES_ID,
+            created_at=time.time(),
+            model=str(getattr(self._inner, "model", "")),
+            object="response",
+            output=model_response.output,
+            tool_choice="auto",
+            tools=[],
+            parallel_tool_calls=False,
+            usage=_to_response_usage(model_response.usage),
+        )
+        yield ResponseCompletedEvent(
+            response=response,
+            type="response.completed",
+            sequence_number=sequence,
+        )
+
+    def get_retry_advice(self, request: Any) -> Any:
+        return self._inner.get_retry_advice(request)
+
+
+def _should_run_non_streamed(settings: Settings) -> bool:
+    return settings.llm.stream_mode == "never"
+
+
 class StrixProvider(MultiProvider):
     """Route any non-OpenAI prefix through LiteLLM with the prefix preserved,
     so users type ``deepseek/deepseek-chat`` rather than
@@ -159,14 +267,18 @@ class StrixProvider(MultiProvider):
         return self._get_fallback_provider("litellm"), original_model_name
 
     def get_model(self, model_name: str | None) -> Model:
+        settings = load_settings()
         slug = codex.subscription_model(model_name)
         if slug:
             return _CodexResponsesModel(
                 slug,
                 codex.get_subscription_client(),
-                reasoning_effort=load_settings().llm.reasoning_effort,
+                reasoning_effort=settings.llm.reasoning_effort,
             )
-        return super().get_model(model_name)
+        model = super().get_model(model_name)
+        if _should_run_non_streamed(settings):
+            return _NonStreamingModel(model)
+        return model
 
 
 DEFAULT_MODEL_RETRY = ModelRetrySettings(
@@ -277,6 +389,51 @@ def _configure_litellm_compatibility() -> None:
     litellm.suppress_debug_info = True
 
     _register_litellm_cost_callback()
+    _install_openrouter_stream_cost_capture()
+
+
+def _install_openrouter_stream_cost_capture() -> None:
+    """Preserve OpenRouter's per-stream cost, which LiteLLM drops when streaming.
+
+    OpenRouter reports the real charge in ``usage.cost`` of the final stream
+    chunk, but LiteLLM rebuilds streamed responses from token-only fields and
+    discards it (its non-streamed path stashes the cost in hidden params; the
+    streaming path does not). Every scan streams, so without this the cost is
+    lost and Strix falls back to a cost-map estimate that is missing entirely
+    for new models (e.g. kimi-k3), reporting $0. Subclass the OpenRouter
+    streaming handler to record the cost keyed by response id so the cost
+    callback can recover the exact charge for the matching rebuilt response.
+    """
+    import litellm
+    from litellm.llms.openrouter.chat.transformation import (
+        OpenRouterChatCompletionStreamingHandler,
+        OpenrouterConfig,
+    )
+
+    from strix.report.state import streamed_openrouter_costs
+
+    class _StrixOpenRouterStreamingHandler(OpenRouterChatCompletionStreamingHandler):
+        def chunk_parser(self, chunk: dict[str, Any]) -> Any:
+            stream = super().chunk_parser(chunk)
+            streamed_openrouter_costs.remember(
+                chunk.get("id") or getattr(stream, "id", None), chunk.get("usage")
+            )
+            return stream
+
+    class _StrixOpenrouterConfig(OpenrouterConfig):
+        def get_model_response_iterator(
+            self, streaming_response: Any, sync_stream: bool, json_mode: bool | None = False
+        ) -> Any:
+            return _StrixOpenRouterStreamingHandler(
+                streaming_response=streaming_response,
+                sync_stream=sync_stream,
+                json_mode=json_mode,
+            )
+
+    # LiteLLM's provider-config factory reads litellm.OpenrouterConfig at call
+    # time, so overriding the attribute is enough for the subclass to take
+    # effect. (type: ignore — mypy rejects reassigning a class attribute.)
+    litellm.OpenrouterConfig = _StrixOpenrouterConfig  # type: ignore[misc]
 
 
 _OPENROUTER_ATTRIBUTION_HEADERS = {

strix/config/settings.py

@@ -9,6 +9,7 @@ from pydantic_settings import BaseSettings, SettingsConfigDict
 
 
 ReasoningEffort = Literal["none", "minimal", "low", "medium", "high", "xhigh"]
+StreamMode = Literal["auto", "always", "never"]
 
 _BASE_CONFIG = SettingsConfigDict(
     case_sensitive=False,
@@ -40,6 +41,12 @@ class LlmSettings(BaseSettings):
         default=False,
         alias="STRIX_FORCE_REQUIRED_TOOL_CHOICE",
     )
+    # auto/always stream; never runs non-streamed, for endpoints that stream tool
+    # calls incorrectly (some OpenAI-compatible gateways serving reasoning models).
+    stream_mode: StreamMode = Field(
+        default="auto",
+        alias="STRIX_STREAM_MODE",
+    )
     prompt_cache: bool = Field(
         default=True,
         alias="STRIX_PROMPT_CACHE",

strix/core/agents.py

@@ -200,7 +200,9 @@ class AgentCoordinator:
         logger.info("agent.status %s=%s", agent_id, status)
         await self._maybe_snapshot()
 
-    async def send(self, target_agent_id: str, message: dict[str, Any]) -> bool:
+    async def send(
+        self, target_agent_id: str, message: dict[str, Any], *, interrupt: bool = True
+    ) -> bool:
         """Deliver a user/peer message by appending it to the target SDK session."""
         if message.get("from") == "user" and self._budget_paused:
             await self.resume_from_budget_pause(exclude=target_agent_id)
@@ -211,7 +213,7 @@ class AgentCoordinator:
             runtime = self.runtimes.setdefault(target_agent_id, AgentRuntime())
             session = runtime.session
             stream = runtime.stream
-            interrupt = runtime.interrupt_on_message
+            interrupt_on_message = runtime.interrupt_on_message
         if session is None:
             logger.warning(
                 "agent.send dropped target=%s because its SDK session is not attached",
@@ -230,7 +232,7 @@ class AgentCoordinator:
         async with self._lock:
             self.pending_counts[target_agent_id] = self.pending_counts.get(target_agent_id, 0) + 1
             self.runtimes.setdefault(target_agent_id, AgentRuntime()).wake.set()
-        if stream is not None and interrupt:
+        if stream is not None and interrupt and interrupt_on_message:
             stream.cancel(mode="immediate")
         await self._maybe_snapshot()
         return True

strix/core/execution.py

@@ -21,6 +21,7 @@ from openai import (
     RateLimitError,
 )
 
+from strix.config import codex
 from strix.core.hooks import (
     BudgetExceededError,
     BudgetPausedError,
@@ -88,6 +89,11 @@ async def _compact_session(
     )
 
 
+_GUARDRAIL_PARK_ERROR = (
+    "Blocked by the model's content guardrail (flagged as a possible cybersecurity risk). "
+    "Set STRIX_LLM to a model that isn't blocked and resume the scan to continue."
+)
+
 _TRANSIENT_MODEL_STATUS_CODES = frozenset({408, 500, 502, 503, 504})
 _MAX_TRANSIENT_MODEL_RETRIES = 4
 _TRANSIENT_MODEL_RETRY_BASE_DELAY_S = 2.0
@@ -304,6 +310,7 @@ async def respawn_subagents(
             if coordinator.parent_of.get(aid) is None or aid == root_id:
                 continue
             md["_restored_status"] = status
+            md["_restored_error"] = coordinator.errors.get(aid)
             candidates.append(
                 (
                     aid,
@@ -316,7 +323,8 @@ async def respawn_subagents(
     for child_id, name, parent_id, md in candidates:
         try:
             restored_status = str(md.get("_restored_status") or "running")
-            start_parked = interactive and restored_status != "running"
+            recoverable_park = restored_status == "waiting" and bool(md.get("_restored_error"))
+            start_parked = interactive and restored_status != "running" and not recoverable_park
 
             if start_parked:
                 logger.warning(
@@ -572,6 +580,10 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
                 if session is not None:
                     input_data = []
                 continue
+            if codex.is_content_guardrail_error(exc):
+                return await _handle_content_guardrail(
+                    coordinator, agent_id, exc, interactive=interactive
+                )
             if not interactive:
                 raise
             if isinstance(exc, MaxTurnsExceeded):
@@ -589,6 +601,22 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
             return stream
 
 
+async def _handle_content_guardrail(
+    coordinator: AgentCoordinator,
+    agent_id: str,
+    exc: BaseException,
+    *,
+    interactive: bool,
+) -> RunResultBase | None:
+    logger.warning("agent %s blocked by the model's content guardrail: %s", agent_id, exc)
+    if interactive:
+        await coordinator.set_status(agent_id, "waiting", error=_GUARDRAIL_PARK_ERROR)
+        return None
+    await coordinator.set_status(agent_id, "failed", error=_GUARDRAIL_PARK_ERROR)
+    await _notify_parent_on_terminal(coordinator, agent_id, "failed")
+    return None
+
+
 async def _settle_run_result(
     coordinator: AgentCoordinator,
     agent_id: str,
@@ -685,6 +713,7 @@ async def _notify_parent_on_terminal(
             "priority": "high",
             "content": template.format(name=name, agent_id=agent_id),
         },
+        interrupt=False,
     )
 
 

strix/core/runner.py

@@ -376,6 +376,12 @@ async def run_strix_scan(
 
         async with coordinator._lock:
             root_status = coordinator.statuses.get(root_id)
+            root_error = coordinator.errors.get(root_id)
+
+        root_recoverable_park = root_status == "waiting" and bool(root_error)
+        root_start_parked = bool(
+            interactive and is_resume and root_status != "running" and not root_recoverable_park
+        )
 
         result = await run_agent_loop(
             agent=root_agent,
@@ -387,7 +393,7 @@ async def run_strix_scan(
             agent_id=root_id,
             interactive=interactive,
             session=root_session,
-            start_parked=bool(interactive and is_resume and root_status != "running"),
+            start_parked=root_start_parked,
             event_sink=event_sink,
             hooks=hooks,
         )

strix/interface/main.py

@@ -884,7 +884,7 @@ def display_completion_message(args: argparse.Namespace, results_path: Path) ->
     view_text = Text()
     view_text.append("\n")
     view_text.append("View", style="dim")
-    view_text.append("         ")
+    view_text.append("    ")
     view_text.append(f"strix view {args.run_name}", style="#22c55e")
     panel_parts.extend(["\n", view_text])
 

strix/interface/tui/live_view.py

@@ -20,7 +20,7 @@ class TuiLiveView:
         self.events: list[dict[str, Any]] = []
         self._next_event_id = 1
         self._open_assistant_event_by_agent: dict[str, dict[str, Any]] = {}
-        self._tool_event_by_call_id: dict[str, dict[str, Any]] = {}
+        self._tool_event_by_agent_and_call_id: dict[tuple[str, str], dict[str, Any]] = {}
 
     def hydrate_from_run_dir(self, run_dir: Path) -> None:
         state_dir = runtime_state_dir(run_dir)
@@ -223,7 +223,8 @@ class TuiLiveView:
         timestamp: str | None = None,
     ) -> None:
         call_id = call["call_id"]
-        existing = self._tool_event_by_call_id.get(call_id)
+        event_key = (agent_id, call_id)
+        existing = self._tool_event_by_agent_and_call_id.get(event_key)
         tool_data = {
             "tool_name": call["tool_name"],
             "args": call["args"],
@@ -233,7 +234,7 @@ class TuiLiveView:
         }
         if existing is None:
             event = self._append_event(agent_id, "tool", tool_data, timestamp=timestamp)
-            self._tool_event_by_call_id[call_id] = event
+            self._tool_event_by_agent_and_call_id[event_key] = event
         else:
             existing["data"].update(tool_data)
             self._bump_event(existing, timestamp=timestamp)
@@ -249,7 +250,8 @@ class TuiLiveView:
         timestamp: str | None = None,
     ) -> None:
         call_id = output["call_id"]
-        event = self._tool_event_by_call_id.get(call_id)
+        event_key = (agent_id, call_id)
+        event = self._tool_event_by_agent_and_call_id.get(event_key)
         if event is None:
             event = self._append_event(
                 agent_id,
@@ -263,7 +265,7 @@ class TuiLiveView:
                 },
                 timestamp=timestamp,
             )
-            self._tool_event_by_call_id[call_id] = event
+            self._tool_event_by_agent_and_call_id[event_key] = event
 
         result = _parse_json_value(output["output"])
         event["data"]["result"] = result

strix/interface/utils.py

@@ -11,11 +11,10 @@ import tempfile
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any
-from urllib.error import HTTPError, URLError
 from urllib.parse import urlparse
-from urllib.request import Request, urlopen
 
 import docker
+import requests
 from docker.errors import DockerException, ImageNotFound
 from rich.console import Console
 from rich.panel import Panel
@@ -1088,13 +1087,12 @@ def resolve_diff_scope_context(
 def _is_http_git_repo(url: str) -> bool:
     check_url = f"{url.rstrip('/')}/info/refs?service=git-upload-pack"
     try:
-        req = Request(check_url, headers={"User-Agent": "git/strix"})  # noqa: S310
-        with urlopen(req, timeout=10) as resp:  # noqa: S310  # nosec B310
-            return "x-git-upload-pack-advertisement" in resp.headers.get("Content-Type", "")
-    except HTTPError as e:
-        return e.code == 401
-    except (URLError, OSError, ValueError):
+        resp = requests.get(check_url, headers={"User-Agent": "git/strix"}, timeout=10)
+    except (requests.RequestException, ValueError):
         return False
+    if resp.status_code >= 400:
+        return resp.status_code == 401
+    return "x-git-upload-pack-advertisement" in resp.headers.get("Content-Type", "")
 
 
 def infer_target_type(target: str) -> tuple[str, dict[str, str]]:  # noqa: PLR0911

strix/interface/viewer/auth.py

@@ -15,12 +15,12 @@ import base64
 import contextlib
 import json
 import logging
-import urllib.error
-import urllib.request
 from datetime import UTC, datetime
 from pathlib import Path
 from typing import Any
 
+import requests
+
 from strix.config.loader import load_settings
 
 
@@ -147,21 +147,17 @@ def _post_json(path: str, payload: dict[str, Any], *, timeout: int) -> tuple[int
     map, not raised.
     """
     url = f"{_app_url()}{path}"
-    body = json.dumps(payload).encode("utf-8")
-    request = urllib.request.Request(  # noqa: S310 - fixed https relay URL
-        url,
-        data=body,
-        headers={"Content-Type": "application/json", "Accept": "application/json"},
-        method="POST",
-    )
     try:
-        with urllib.request.urlopen(request, timeout=timeout) as response:  # noqa: S310  # nosec B310
-            return response.status, _parse_body(response.read())
-    except urllib.error.HTTPError as exc:
-        return exc.code, _parse_body(exc.read())
-    except (urllib.error.URLError, TimeoutError, OSError) as exc:
+        response = requests.post(
+            url,
+            json=payload,
+            headers={"Accept": "application/json"},
+            timeout=timeout,
+        )
+    except requests.RequestException as exc:
         logger.warning("relay request to %s failed: %s", path, exc)
         raise RelayError("unavailable") from exc
+    return response.status_code, _parse_body(response.content)
 
 
 def _parse_body(raw: bytes) -> dict[str, Any]:

strix/interface/viewer/server.py

@@ -107,8 +107,11 @@ def resolve_run_dir(base_dir: Path, run_param: str | None, default_run_dir: Path
     return candidate
 
 
-# Name of the cookie carrying the per-process session capability.
-SESSION_COOKIE = "strix_viewer_session"
+# Prefix of the cookie carrying the per-process session capability. The bound
+# port is appended (``strix_viewer_session_<port>``) because browsers scope
+# cookies by host only, never by port: concurrent viewers on 127.0.0.1 would
+# otherwise share one cookie slot and clobber each other's session.
+SESSION_COOKIE_PREFIX = "strix_viewer_session"
 
 
 class _ViewerState:
@@ -135,6 +138,9 @@ class _ViewerState:
         # enough to steer a live scan, trigger a report, or browse history --
         # the token is never handed to a caller who merely reaches ``/``.
         self.session_token = secrets.token_urlsafe(32)
+        # Finalized in ``serve()`` once the port is known (the server binds
+        # after this state is constructed); see SESSION_COOKIE_PREFIX.
+        self.cookie_name = SESSION_COOKIE_PREFIX
 
 
 def _make_handler(state: _ViewerState) -> type[BaseHTTPRequestHandler]:
@@ -476,7 +482,7 @@ def _make_handler(state: _ViewerState) -> type[BaseHTTPRequestHandler]:
             the browser this process handed the page to can pass. A direct
             caller on an exposed port has no cookie and is rejected.
             """
-            supplied = self._cookies().get(SESSION_COOKIE, "")
+            supplied = self._cookies().get(state.cookie_name, "")
             return bool(supplied) and secrets.compare_digest(supplied, state.session_token)
 
         def _token_presented(self, query: dict[str, list[str]]) -> bool:
@@ -512,7 +518,7 @@ def _make_handler(state: _ViewerState) -> type[BaseHTTPRequestHandler]:
                 # SameSite=Strict (never sent from a cross-site context).
                 self.send_header(
                     "Set-Cookie",
-                    f"{SESSION_COOKIE}={state.session_token}; Path=/; HttpOnly; SameSite=Strict",
+                    f"{state.cookie_name}={state.session_token}; Path=/; HttpOnly; SameSite=Strict",
                 )
             self.end_headers()
             self.wfile.write(content)
@@ -586,6 +592,7 @@ def serve(
 
     httpd.daemon_threads = True
     bound_port = int(httpd.server_address[1])
+    state.cookie_name = f"{SESSION_COOKIE_PREFIX}_{bound_port}"
     url = f"http://{host}:{bound_port}"
 
     thread = threading.Thread(target=httpd.serve_forever, name="strix-viewer", daemon=True)

strix/report/state.py

@@ -1,6 +1,7 @@
 import json
 import logging
 import subprocess
+import threading
 from collections.abc import Callable
 from datetime import UTC, datetime
 from importlib.metadata import PackageNotFoundError, version
@@ -95,6 +96,8 @@ def get_global_report_state() -> Optional["ReportState"]:
 def set_global_report_state(report_state: "ReportState") -> None:
     global _global_report_state  # noqa: PLW0603
     _global_report_state = report_state
+    # New run: drop any streamed-cost entries a prior run left unconsumed.
+    streamed_openrouter_costs.clear()
 
 
 class ReportState:
@@ -507,6 +510,72 @@ class ReportState:
         self._sync_llm_usage_record()
 
 
+def openrouter_stream_cost(usage: Any) -> float | None:
+    """Total OpenRouter-reported cost from a raw stream ``usage`` block, or None.
+
+    Non-BYOK responses bill everything to ``usage.cost``. BYOK responses put the
+    OpenRouter fee in ``usage.cost`` (often 0) and the provider charge in
+    ``usage.cost_details.upstream_inference_cost``, so BYOK totals sum the two.
+    """
+    if not isinstance(usage, dict):
+        return None
+    total = 0.0
+    cost = usage.get("cost")
+    if isinstance(cost, int | float) and cost > 0:
+        total += float(cost)
+    if bool(usage.get("is_byok")):
+        details = usage.get("cost_details")
+        upstream = details.get("upstream_inference_cost") if isinstance(details, dict) else None
+        if isinstance(upstream, int | float) and upstream > 0:
+            total += float(upstream)
+    return total if total > 0 else None
+
+
+def _response_id(completion_response: Any) -> str | None:
+    response_id = getattr(completion_response, "id", None)
+    if response_id is None and isinstance(completion_response, dict):
+        response_id = cast("dict[str, Any]", completion_response).get("id")
+    return response_id if isinstance(response_id, str) and response_id else None
+
+
+class StreamedOpenRouterCosts:
+    """Correlates OpenRouter's per-stream cost from the parser to the cost callback.
+
+    LiteLLM rebuilds streamed responses from token-only chunks and drops the
+    ``usage.cost`` OpenRouter reports in its final stream chunk (its non-streamed
+    path preserves it; streaming snapshots hidden params at stream start). Every
+    scan streams, so the OpenRouter streaming handler (see strix.config.models)
+    records the cost here keyed by response id, and the callback takes it back out
+    for the matching rebuilt response. Entries are removed on read; ``clear()``
+    runs per scan so nothing accumulates across runs.
+    """
+
+    def __init__(self) -> None:
+        self._costs: dict[str, float] = {}
+        self._lock = threading.Lock()
+
+    def remember(self, response_id: Any, usage: Any) -> None:
+        cost = openrouter_stream_cost(usage)
+        if cost is None or not (isinstance(response_id, str) and response_id):
+            return
+        with self._lock:
+            self._costs[response_id] = cost
+
+    def take(self, completion_response: Any) -> float | None:
+        response_id = _response_id(completion_response)
+        if response_id is None:
+            return None
+        with self._lock:
+            return self._costs.pop(response_id, None)
+
+    def clear(self) -> None:
+        with self._lock:
+            self._costs.clear()
+
+
+streamed_openrouter_costs = StreamedOpenRouterCosts()
+
+
 def litellm_cost_callback(
     kwargs: Any,
     completion_response: Any,
@@ -541,6 +610,11 @@ def litellm_cost_callback(
     if cost is None:
         cost = _usage_reported_cost(completion_response)
 
+    # Recover the exact OpenRouter cost the streaming handler stashed for this
+    # response — LiteLLM drops it from streamed usage, so nothing above sees it.
+    if cost is None:
+        cost = streamed_openrouter_costs.take(completion_response)
+
     if cost is None:
         cost = _estimate_response_cost(kwargs, completion_response)
 

strix/telemetry/posthog.py

@@ -1,9 +1,9 @@
-import json
 import logging
-import urllib.request
 from datetime import datetime
 from typing import TYPE_CHECKING, Any
 
+import requests
+
 from strix.config import load_settings
 from strix.telemetry._common import (
     SESSION_ID,
@@ -37,13 +37,7 @@ def _send(event: str, properties: dict[str, Any]) -> bool:
             "distinct_id": SESSION_ID,
             "properties": properties,
         }
-        req = urllib.request.Request(  # noqa: S310
-            f"{_POSTHOG_HOST}/capture/",
-            data=json.dumps(payload).encode(),
-            headers={"Content-Type": "application/json"},
-        )
-        with urllib.request.urlopen(req, timeout=10):  # noqa: S310  # nosec B310
-            pass
+        requests.post(f"{_POSTHOG_HOST}/capture/", json=payload, timeout=10)
     except Exception:  # noqa: BLE001
         logger.debug("posthog send failed for event %s", event, exc_info=True)
         return False

strix/telemetry/scarf.py

@@ -2,10 +2,11 @@ from __future__ import annotations
 
 import logging
 import urllib.parse
-import urllib.request
 from datetime import datetime
 from typing import TYPE_CHECKING, Any
 
+import requests
+
 from strix.config import load_settings
 from strix.telemetry._common import (
     SESSION_ID,
@@ -42,9 +43,7 @@ def _send(event: str, properties: dict[str, Any]) -> bool:
         url = f"{_SCARF_ENDPOINT}{path}"
         if query:
             url = f"{url}?{query}"
-        req = urllib.request.Request(url, method="POST")  # noqa: S310
-        with urllib.request.urlopen(req, timeout=10):  # noqa: S310  # nosec B310
-            pass
+        requests.post(url, timeout=10)
     except Exception:  # noqa: BLE001
         logger.debug("scarf send failed for event %s", event, exc_info=True)
         return False

tests/test_codex_auth.py

@@ -7,8 +7,10 @@ import hashlib
 import json
 import time
 from typing import TYPE_CHECKING, Any
+from unittest import mock
 
 import pytest
+import requests
 
 from strix.config import codex
 
@@ -52,6 +54,18 @@ def test_authorize_url_carries_pkce_and_client() -> None:
     assert "state=st8" in url
 
 
+def test_post_form_returns_parsed_body() -> None:
+    resp = mock.MagicMock()
+    resp.status_code = 200
+    resp.content = b'{"access_token": "tok"}'
+
+    with mock.patch.object(requests, "post", return_value=resp) as post:
+        data = codex._post_form({"grant_type": "refresh_token"})
+
+    assert data == {"access_token": "tok"}
+    assert post.call_args.kwargs["timeout"] == codex._TOKEN_TIMEOUT
+
+
 @pytest.mark.parametrize(
     ("value", "expected"),
     [

tests/test_cost_tracking.py

@@ -7,9 +7,25 @@ from unittest.mock import MagicMock, patch
 
 import litellm
 import pytest
+from litellm.types.utils import LlmProviders
+from litellm.utils import ProviderConfigManager
 
-from strix.config.models import _configure_litellm_compatibility
-from strix.report.state import litellm_cost_callback
+from strix.config.models import (
+    _configure_litellm_compatibility,
+    _install_openrouter_stream_cost_capture,
+)
+from strix.report.state import (
+    ReportState,
+    litellm_cost_callback,
+    openrouter_stream_cost,
+    set_global_report_state,
+    streamed_openrouter_costs,
+)
+
+
+@pytest.fixture(autouse=True)
+def _clear_streamed_costs() -> None:
+    streamed_openrouter_costs.clear()
 
 
 def test_streaming_logging_stays_enabled_for_cost_callback() -> None:
@@ -151,3 +167,90 @@ def test_cost_callback_records_nothing_when_no_cost_available() -> None:
         litellm_cost_callback({"response_cost": None, "model": "x/y"}, response)
 
     report_state.record_observed_llm_cost.assert_not_called()
+
+
+def test_openrouter_stream_cost_extracts_plain_and_byok_totals() -> None:
+    assert openrouter_stream_cost({"cost": 0.003168}) == pytest.approx(0.003168)
+    assert openrouter_stream_cost(
+        {"cost": 0.01, "is_byok": True, "cost_details": {"upstream_inference_cost": 0.2}}
+    ) == pytest.approx(0.21)
+    # Upstream cost is only added for BYOK responses.
+    assert openrouter_stream_cost(
+        {"cost": 0.05, "is_byok": False, "cost_details": {"upstream_inference_cost": 0.04}}
+    ) == pytest.approx(0.05)
+    assert openrouter_stream_cost({"prompt_tokens": 10}) is None
+    assert openrouter_stream_cost(None) is None
+
+
+def test_cost_callback_recovers_streamed_openrouter_cost_by_response_id() -> None:
+    report_state = MagicMock()
+    streamed_openrouter_costs.remember("gen-abc", {"cost": 0.42})
+    # LiteLLM strips cost from the rebuilt streamed usage; only the id survives.
+    response = SimpleNamespace(id="gen-abc", usage=SimpleNamespace(cost=None), _hidden_params={})
+
+    with (
+        patch("strix.report.state.get_global_report_state", return_value=report_state),
+        patch("litellm.completion_cost", side_effect=ValueError("unknown model")),
+    ):
+        litellm_cost_callback({"response_cost": None, "model": "moonshotai/kimi-k3"}, response)
+
+    report_state.record_observed_llm_cost.assert_called_once_with(0.42)
+    # The entry is consumed so a later response cannot double-count it.
+    assert streamed_openrouter_costs.take(response) is None
+
+
+def test_streamed_openrouter_cost_prefers_provider_report_over_estimate() -> None:
+    report_state = MagicMock()
+    streamed_openrouter_costs.remember("gen-xyz", {"cost": 0.9})
+    response = SimpleNamespace(
+        id="gen-xyz",
+        usage=SimpleNamespace(prompt_tokens=10, completion_tokens=5, total_tokens=15),
+        _hidden_params={},
+    )
+
+    with (
+        patch("strix.report.state.get_global_report_state", return_value=report_state),
+        patch("litellm.completion_cost", return_value=0.1) as estimate,
+    ):
+        litellm_cost_callback({"response_cost": None, "model": "moonshotai/kimi-k3"}, response)
+
+    report_state.record_observed_llm_cost.assert_called_once_with(0.9)
+    estimate.assert_not_called()
+
+
+def test_streamed_openrouter_costs_ignores_entries_without_cost() -> None:
+    streamed_openrouter_costs.remember("gen-none", {"prompt_tokens": 10})
+    streamed_openrouter_costs.remember("", {"cost": 0.5})
+    assert streamed_openrouter_costs.take(SimpleNamespace(id="gen-none")) is None
+
+
+def test_streamed_openrouter_costs_cleared_on_new_run() -> None:
+    streamed_openrouter_costs.remember("gen-stale", {"cost": 0.7})
+    set_global_report_state(ReportState.__new__(ReportState))
+    assert streamed_openrouter_costs.take(SimpleNamespace(id="gen-stale")) is None
+
+
+def test_openrouter_stream_handler_records_cost() -> None:
+    _install_openrouter_stream_cost_capture()
+    # Resolve the config the way LiteLLM does in production so we prove the
+    # override is actually reachable through provider resolution, not just as a
+    # directly-constructed class.
+    config = ProviderConfigManager.get_provider_chat_config(
+        model="moonshotai/kimi-k3", provider=LlmProviders.OPENROUTER
+    )
+    assert config is not None
+    assert type(config).__name__ == "_StrixOpenrouterConfig"
+    handler = config.get_model_response_iterator(streaming_response=iter([]), sync_stream=True)
+
+    chunk = {
+        "id": "gen-stream",
+        "created": 1,
+        "model": "moonshotai/kimi-k3",
+        "choices": [{"index": 0, "delta": {"content": None}}],
+        "usage": {"prompt_tokens": 89, "completion_tokens": 138, "cost": 0.0035055},
+    }
+    handler.chunk_parser(chunk)
+
+    assert streamed_openrouter_costs.take(SimpleNamespace(id="gen-stream")) == pytest.approx(
+        0.0035055
+    )

tests/test_execution.py

@@ -6,13 +6,21 @@ import asyncio
 import contextlib
 import json
 from typing import Any
+from unittest.mock import MagicMock
 
 import pytest
 from agents.memory import SQLiteSession
 from agents.tool_context import ToolContext
 
+from strix.config import codex
+from strix.core import execution
 from strix.core.agents import AgentCoordinator
-from strix.core.execution import _notify_parent_on_terminal, _notify_root_on_budget_reserve
+from strix.core.execution import (
+    _handle_content_guardrail,
+    _notify_parent_on_terminal,
+    _notify_root_on_budget_reserve,
+    respawn_subagents,
+)
 from strix.tools.finish.tool import finish_scan
 
 
@@ -465,3 +473,106 @@ async def test_notify_parent_on_terminal_ignores_non_terminal_status(tmp_path: A
 
     assert coordinator.pending_counts.get("root", 0) == 0
     session.close()
+
+
+class _RecordingStream:
+    def __init__(self) -> None:
+        self.cancelled = False
+        self.cancel_mode: str | None = None
+
+    def cancel(self, mode: str = "immediate") -> None:
+        self.cancelled = True
+        self.cancel_mode = mode
+
+
+@pytest.mark.asyncio
+async def test_terminal_notice_does_not_cancel_parent_stream(tmp_path: Any) -> None:
+    coordinator = AgentCoordinator()
+    await coordinator.register("root", "strix", parent_id=None)
+    await coordinator.register("child", "recon", parent_id="root")
+    session = SQLiteSession("root", tmp_path / "agents.db")
+    stream = _RecordingStream()
+    await coordinator.attach_runtime("root", session=session, interrupt_on_message=True)
+    await coordinator.attach_stream("root", stream)
+
+    await _notify_parent_on_terminal(coordinator, "child", "crashed")
+
+    assert stream.cancelled is False
+    assert coordinator.pending_counts.get("root", 0) > 0
+    session.close()
+
+
+@pytest.mark.asyncio
+async def test_guardrail_interactive_parks_agent_wakeable(tmp_path: Any) -> None:
+    coordinator = AgentCoordinator()
+    await coordinator.register("root", "strix", parent_id=None)
+    await coordinator.register("child", "recon", parent_id="root")
+    exc = codex.CodexContentGuardrailError("chatgpt/gpt-5.6-sol")
+
+    result = await _handle_content_guardrail(coordinator, "child", exc, interactive=True)
+
+    assert result is None
+    assert coordinator.statuses["child"] == "waiting"
+    assert "STRIX_LLM" in coordinator.errors["child"]
+
+    waiter = asyncio.create_task(coordinator.wait_for_message("child"))
+    await asyncio.sleep(0)
+    assert not waiter.done()
+    session = SQLiteSession("child", tmp_path / "agents.db")
+    await coordinator.attach_runtime("child", session=session)
+    await coordinator.send("child", {"from": "user", "content": "switched model, resume"})
+    await asyncio.wait_for(waiter, timeout=1.0)
+    session.close()
+
+
+@pytest.mark.asyncio
+async def test_guardrail_noninteractive_fails_only_blocked_agent(tmp_path: Any) -> None:
+    coordinator = AgentCoordinator()
+    await coordinator.register("root", "strix", parent_id=None)
+    await coordinator.register("child", "recon", parent_id="root")
+    session = SQLiteSession("root", tmp_path / "agents.db")
+    await coordinator.attach_runtime("root", session=session)
+    exc = codex.CodexContentGuardrailError("chatgpt/gpt-5.6-sol")
+
+    result = await _handle_content_guardrail(coordinator, "child", exc, interactive=False)
+
+    assert result is None
+    assert coordinator.statuses["child"] == "failed"
+    assert "STRIX_LLM" in coordinator.errors["child"]
+    assert coordinator.statuses["root"] == "running"
+    assert coordinator.pending_counts.get("root", 0) > 0
+    session.close()
+
+
+@pytest.mark.asyncio
+async def test_resume_revives_guardrail_parked_child_but_not_plain_waiting(
+    tmp_path: Any, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    coordinator = AgentCoordinator()
+    await coordinator.register("root", "strix", parent_id=None)
+    await coordinator.register("blocked", "recon", parent_id="root")
+    await coordinator.register("peer_waiter", "recon", parent_id="root")
+    await coordinator.set_status("blocked", "waiting", error="STRIX_LLM guardrail")
+    await coordinator.set_status("peer_waiter", "waiting")
+
+    parked: dict[str, bool] = {}
+
+    async def _fake_start_child_runner(**kwargs: Any) -> None:
+        parked[kwargs["child_id"]] = bool(kwargs["start_parked"])
+
+    monkeypatch.setattr(execution, "_start_child_runner", _fake_start_child_runner)
+
+    await respawn_subagents(
+        coordinator=coordinator,
+        factory=lambda **_kwargs: object(),
+        agents_db_path=tmp_path / "agents.db",
+        sessions_to_close=[],
+        run_config=MagicMock(),
+        max_turns=10,
+        interactive=True,
+        parent_ctx={"agent_id": "root", "parent_id": None},
+        root_id="root",
+    )
+
+    assert parked["blocked"] is False
+    assert parked["peer_waiter"] is True

tests/test_nonstreaming_model.py

@@ -0,0 +1,173 @@
+"""Tests for the non-streaming wrapper used on custom OpenAI-compatible endpoints."""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+from typing import TYPE_CHECKING, cast
+from unittest.mock import patch
+
+import pytest
+from agents.items import ModelResponse
+from agents.model_settings import ModelSettings
+from agents.models.interface import Model, ModelTracing
+from agents.usage import Usage
+from openai.types.responses import (
+    ResponseCompletedEvent,
+    ResponseFunctionToolCall,
+    ResponseOutputItemDoneEvent,
+    ResponseStreamEvent,
+)
+
+from strix.config.models import StrixProvider, _NonStreamingModel, _to_response_usage
+
+
+if TYPE_CHECKING:
+    from collections.abc import AsyncIterator
+
+
+def _tool_call() -> ResponseFunctionToolCall:
+    return ResponseFunctionToolCall(
+        arguments='{"command": "ls"}',
+        call_id="call_1",
+        name="terminal_execute",
+        type="function_call",
+    )
+
+
+class _FakeModel(Model):
+    def __init__(self, response: ModelResponse) -> None:
+        self.model = "fake-model"
+        self._response = response
+        self.get_response_calls = 0
+
+    async def get_response(self, *_args: object, **_kwargs: object) -> ModelResponse:
+        self.get_response_calls += 1
+        return self._response
+
+    async def stream_response(  # pragma: no cover
+        self, *_args: object, **_kwargs: object
+    ) -> AsyncIterator[ResponseStreamEvent]:
+        for _ in range(0):
+            yield cast("ResponseStreamEvent", None)
+        raise AssertionError("inner stream_response must never be called")
+
+
+def _settings(*, api_base: str | None, stream_mode: str = "auto") -> SimpleNamespace:
+    return SimpleNamespace(
+        llm=SimpleNamespace(
+            model="openai/glm",
+            api_base=api_base,
+            stream_mode=stream_mode,
+            reasoning_effort="high",
+        )
+    )
+
+
+def test_to_response_usage_maps_token_details() -> None:
+    usage = Usage(requests=1, input_tokens=10, output_tokens=5, total_tokens=15)
+    usage.input_tokens_details.cached_tokens = 4
+    usage.output_tokens_details.reasoning_tokens = 3
+    mapped = _to_response_usage(usage)
+    assert mapped is not None
+    assert (mapped.input_tokens, mapped.output_tokens, mapped.total_tokens) == (10, 5, 15)
+    assert mapped.input_tokens_details.cached_tokens == 4
+    assert mapped.output_tokens_details.reasoning_tokens == 3
+
+
+def test_to_response_usage_none() -> None:
+    assert _to_response_usage(None) is None
+
+
+@pytest.mark.asyncio
+async def test_stream_response_synthesizes_tool_call_from_non_streamed() -> None:
+    tool_call = _tool_call()
+    inner = _FakeModel(
+        ModelResponse(
+            output=[tool_call],
+            usage=Usage(requests=1, input_tokens=10, output_tokens=5, total_tokens=15),
+            response_id="resp_123",
+        )
+    )
+    wrapper = _NonStreamingModel(inner)
+
+    events = [
+        event
+        async for event in wrapper.stream_response(
+            "sys",
+            "hi",
+            ModelSettings(),
+            [],
+            None,
+            [],
+            ModelTracing.DISABLED,
+        )
+    ]
+
+    assert inner.get_response_calls == 1
+    item_done = [e for e in events if isinstance(e, ResponseOutputItemDoneEvent)]
+    completed = [e for e in events if isinstance(e, ResponseCompletedEvent)]
+    assert len(item_done) == 1
+    assert item_done[0].item == tool_call
+    assert len(completed) == 1
+    final = completed[0].response
+    assert final.output == [tool_call]
+    assert final.id == "resp_123"
+    assert final.usage is not None
+    assert final.usage.total_tokens == 15
+    # sequence numbers are strictly increasing
+    assert [e.sequence_number for e in events] == list(range(len(events)))
+
+
+@pytest.mark.asyncio
+async def test_stream_response_delegates_get_response() -> None:
+    inner = _FakeModel(
+        ModelResponse(output=[], usage=Usage(), response_id=None),
+    )
+    wrapper = _NonStreamingModel(inner)
+    result = await wrapper.get_response(
+        "sys", "hi", ModelSettings(), [], None, [], ModelTracing.DISABLED
+    )
+    assert result is inner._response
+    assert inner.get_response_calls == 1
+
+
+def test_get_model_auto_streams_custom_endpoint() -> None:
+    sentinel = _FakeModel(ModelResponse(output=[], usage=Usage(), response_id=None))
+    with (
+        patch("strix.config.models.load_settings", return_value=_settings(api_base="http://x/v1")),
+        patch(
+            "agents.models.multi_provider.MultiProvider.get_model",
+            return_value=sentinel,
+        ),
+    ):
+        model = StrixProvider().get_model("openai/glm")
+    assert model is sentinel
+
+
+def test_get_model_auto_streams_hosted() -> None:
+    sentinel = _FakeModel(ModelResponse(output=[], usage=Usage(), response_id=None))
+    with (
+        patch("strix.config.models.load_settings", return_value=_settings(api_base="")),
+        patch(
+            "agents.models.multi_provider.MultiProvider.get_model",
+            return_value=sentinel,
+        ),
+    ):
+        model = StrixProvider().get_model("openai/gpt-4o")
+    assert model is sentinel
+
+
+def test_get_model_stream_mode_never_wraps() -> None:
+    sentinel = _FakeModel(ModelResponse(output=[], usage=Usage(), response_id=None))
+    with (
+        patch(
+            "strix.config.models.load_settings",
+            return_value=_settings(api_base="http://x/v1", stream_mode="never"),
+        ),
+        patch(
+            "agents.models.multi_provider.MultiProvider.get_model",
+            return_value=sentinel,
+        ),
+    ):
+        model = StrixProvider().get_model("openai/glm")
+    assert isinstance(model, _NonStreamingModel)

tests/test_viewer.py

@@ -4,9 +4,11 @@ from __future__ import annotations
 
 import json
 import os
+import sqlite3
 import urllib.error
 import urllib.request
 from typing import TYPE_CHECKING
+from urllib.parse import urlsplit
 
 from strix.core.paths import latest_run_dir, runs_base_dir
 from strix.interface.viewer.server import serve
@@ -86,6 +88,75 @@ def test_build_run_state_from_agents_json(tmp_path: Path) -> None:
     assert state["events"] == []
 
 
+def test_build_run_state_keeps_same_call_id_separate_per_agent(tmp_path: Path) -> None:
+    run_dir = _make_run(tmp_path, "tools", status="completed", end_time=None)
+    agents_db = run_dir / ".state" / "agents.db"
+    rows = [
+        (
+            "root",
+            {
+                "type": "function_call",
+                "call_id": "exec_command_0",
+                "name": "exec_command",
+                "arguments": json.dumps({"cmd": "echo root"}),
+            },
+        ),
+        (
+            "root",
+            {
+                "type": "function_call_output",
+                "call_id": "exec_command_0",
+                "output": json.dumps({"success": True, "output": "root"}),
+            },
+        ),
+        (
+            "child",
+            {
+                "type": "function_call",
+                "call_id": "exec_command_0",
+                "name": "exec_command",
+                "arguments": json.dumps({"cmd": "echo child"}),
+            },
+        ),
+        (
+            "child",
+            {
+                "type": "function_call_output",
+                "call_id": "exec_command_0",
+                "output": json.dumps({"success": True, "output": "child"}),
+            },
+        ),
+    ]
+    with sqlite3.connect(agents_db) as conn:
+        conn.execute(
+            """
+            create table agent_messages (
+                id integer primary key,
+                session_id text not null,
+                message_data text not null,
+                created_at text not null
+            )
+            """
+        )
+        conn.executemany(
+            """
+            insert into agent_messages (session_id, message_data, created_at)
+            values (?, ?, '2026-01-01T00:00:00+00:00')
+            """,
+            [(agent_id, json.dumps(message)) for agent_id, message in rows],
+        )
+
+    state = build_run_state(run_dir)
+    tools = [event for event in state["events"] if event["type"] == "tool"]
+
+    assert len(tools) == 2
+    by_agent = {event["agent_id"]: event for event in tools}
+    assert by_agent["root"]["data"]["args"] == {"cmd": "echo root"}
+    assert by_agent["root"]["data"]["result"]["output"] == "root"
+    assert by_agent["child"]["data"]["args"] == {"cmd": "echo child"}
+    assert by_agent["child"]["data"]["result"]["output"] == "child"
+
+
 def _get(url: str, *, cookie: str | None = None) -> tuple[int, str, bytes]:
     headers = {"Cookie": cookie} if cookie else {}
     req = urllib.request.Request(url, headers=headers)  # noqa: S310 - localhost test server
@@ -271,6 +342,11 @@ def _session_cookie(url: str, token: str) -> str:
     return raw.split(";", 1)[0]
 
 
+def _cookie_name(url: str) -> str:
+    """The per-server session cookie name, derived from the bound port."""
+    return f"strix_viewer_session_{urlsplit(url).port}"
+
+
 def _get_status(url: str, *, cookie: str | None = None) -> int:
     headers = {"Cookie": cookie} if cookie else {}
     req = urllib.request.Request(url, headers=headers)  # noqa: S310 - localhost test server
@@ -311,7 +387,7 @@ def test_capability_issued_only_for_tokened_bootstrap(
         # Only the correct bootstrap token mints the session cookie.
         with urllib.request.urlopen(f"{url}/?token={token}") as resp:  # noqa: S310  # nosec B310
             cookie = str(resp.headers.get("Set-Cookie", ""))
-        assert "strix_viewer_session=" in cookie
+        assert f"{_cookie_name(url)}=" in cookie
         assert "HttpOnly" in cookie and "SameSite=Strict" in cookie
 
         # Static assets never carry it.
@@ -344,7 +420,7 @@ def test_unauthorized_client_cannot_acquire_capability(
             url,
             "/api/agents/steer",
             {"agent_id": "root", "message": "pwn"},
-            cookie="strix_viewer_session=",
+            cookie=f"{_cookie_name(url)}=",
         )
         assert status == 403
         assert delivered == []
@@ -541,6 +617,53 @@ def test_runs_list_requires_session_and_verification(
         httpd.server_close()
 
 
+def test_concurrent_servers_use_distinct_cookies(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """Cookies are host-scoped, not port-scoped: two viewers on 127.0.0.1 must
+    not share a cookie slot, and one server's cookie must not pass the other's
+    session gate."""
+    run_a = _make_run(tmp_path / "a", "run-a", status="running", end_time=None)
+    run_b = _make_run(tmp_path / "b", "run-b", status="running", end_time=None)
+    _bundle(tmp_path, monkeypatch)
+    monkeypatch.setattr(
+        "strix.interface.viewer.auth.read_auth", lambda: {"email": "a@b.com", "token": "t"}
+    )
+    monkeypatch.setattr("strix.interface.viewer.auth.is_verified", lambda: True)
+
+    httpd_a, url_a, token_a = serve(run_a, open_browser=False)
+    httpd_b, url_b, token_b = serve(run_b, open_browser=False)
+    try:
+        cookie_a = _session_cookie(url_a, token_a)
+        cookie_b = _session_cookie(url_b, token_b)
+
+        # The two servers mint differently named cookies, so a browser stores both.
+        assert cookie_a.split("=", 1)[0] == _cookie_name(url_a)
+        assert cookie_b.split("=", 1)[0] == _cookie_name(url_b)
+        assert cookie_a.split("=", 1)[0] != cookie_b.split("=", 1)[0]
+
+        def _status(url: str, cookie: str) -> dict[str, object]:
+            _, _, body = _get(f"{url}/api/auth/status", cookie=cookie)
+            return dict(json.loads(body))
+
+        # Each server honors its own cookie...
+        assert _status(url_a, cookie_a)["verified"] is True
+        assert _status(url_b, cookie_b)["verified"] is True
+        # ...but treats the other server's cookie as session-less.
+        assert _status(url_a, cookie_b)["verified"] is False
+        assert _status(url_b, cookie_a)["verified"] is False
+        # Even both cookies together (what a real browser would send) only
+        # match the token minted by the receiving server.
+        both = f"{cookie_a}; {cookie_b}"
+        assert _status(url_a, both)["verified"] is True
+        assert _status(url_b, both)["verified"] is True
+    finally:
+        httpd_a.shutdown()
+        httpd_a.server_close()
+        httpd_b.shutdown()
+        httpd_b.server_close()
+
+
 def test_server_rejects_path_traversal(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
     run_dir = _make_run(tmp_path, "guard", status="completed", end_time="2026-01-01T00:00:00Z")
     secret = tmp_path / "secret.txt"

uv.lock

@@ -2411,7 +2411,7 @@ wheels = [
 
 [[package]]
 name = "strix-agent"
-version = "1.4.0"
+version = "1.4.1"
 source = { editable = "." }
 dependencies = [
     { name = "caido-sdk-client" },