Sign in with a ChatGPT subscription for inference (#854)

Co-authored-by: Jonathan Singer <jonathansinger@Jonathans-MacBook-Pro.local>
Co-authored-by: Jonathan Singer <jonathansinger@Mac-3004.lan>
Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>

tests/test_auth_cli.py

@@ -0,0 +1,106 @@
+"""Tests for the `strix auth` CLI: subcommand routing and provider naming."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+import pytest
+
+from strix.config import codex
+from strix.interface import auth_cli
+
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+
+@pytest.fixture(autouse=True)
+def _tmp_store(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr(codex, "AUTH_PATH", tmp_path / "home" / ".strix" / "subscription-auth.json")
+
+
+def test_login_provider_is_chatgpt() -> None:
+    assert auth_cli.LOGIN_PROVIDER == "chatgpt"
+    assert codex.PROVIDER in auth_cli._ACCEPTED_PROVIDERS
+    assert "chatgpt" in auth_cli._ACCEPTED_PROVIDERS
+
+
+def test_unknown_subcommand_returns_usage_error() -> None:
+    assert auth_cli.run_auth(["bogus"]) == 2
+
+
+def test_help_returns_zero() -> None:
+    assert auth_cli.run_auth(["--help"]) == 0
+
+
+def test_status_not_signed_in() -> None:
+    assert auth_cli.run_auth(["status"]) == 1
+
+
+def test_login_rejects_unsupported_provider(monkeypatch: pytest.MonkeyPatch) -> None:
+    def _should_not_run(*_args: Any, **_kwargs: Any) -> dict[str, Any]:
+        msg = "OAuth flow must not start for an unsupported provider"
+        raise AssertionError(msg)
+
+    monkeypatch.setattr(auth_cli, "_run_oauth_flow", _should_not_run)
+    assert auth_cli.run_auth(["login", "gemini"]) == 2
+
+
+def test_finish_requires_state_on_loopback(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr(codex, "exchange_code", lambda *_: {"ok": True})
+
+    # Loopback (require_state=True): missing or mismatched state is rejected.
+    with pytest.raises(codex.CodexAuthError) as missing:
+        auth_cli._finish("code", None, "verifier", "expected", require_state=True)
+    assert missing.value.code == "state_mismatch"
+    with pytest.raises(codex.CodexAuthError) as mismatch:
+        auth_cli._finish("code", "wrong", "verifier", "expected", require_state=True)
+    assert mismatch.value.code == "state_mismatch"
+
+    # Matching state proceeds to the exchange.
+    assert auth_cli._finish("code", "expected", "verifier", "expected", require_state=True) == {
+        "ok": True
+    }
+
+
+def test_finish_manual_paste_allows_absent_state(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr(codex, "exchange_code", lambda *_: {"ok": True})
+    # Manual paste (require_state=False): a bare code with no state is accepted,
+    # but a present-and-wrong state is still rejected.
+    assert auth_cli._finish("code", None, "verifier", "expected", require_state=False) == {
+        "ok": True
+    }
+    with pytest.raises(codex.CodexAuthError):
+        auth_cli._finish("code", "wrong", "verifier", "expected", require_state=False)
+
+
+def test_finish_rejects_missing_code() -> None:
+    with pytest.raises(codex.CodexAuthError) as exc:
+        auth_cli._finish(None, "expected", "verifier", "expected", require_state=True)
+    assert exc.value.code == "no_code"
+
+
+def test_model_subcommand_removed() -> None:
+    assert auth_cli.run_auth(["model", "gpt-5.5"]) == 2
+
+
+@pytest.mark.parametrize("provider", ["chatgpt", "codex", "ChatGPT"])
+def test_login_accepts_provider_aliases(provider: str, monkeypatch: pytest.MonkeyPatch) -> None:
+    reached = {"flow": False}
+
+    def _fake_flow(*_args: Any, **_kwargs: Any) -> dict[str, Any]:
+        reached["flow"] = True
+        return {
+            "type": "oauth",
+            "provider": "codex",
+            "access": "a",
+            "refresh": "r",
+            "account_id": "acct",
+            "expires_at": 0,
+        }
+
+    monkeypatch.setattr(auth_cli, "_run_oauth_flow", _fake_flow)
+    monkeypatch.setattr(codex, "save_record", lambda _record: None)
+
+    assert auth_cli.run_auth(["login", provider]) == 0
+    assert reached["flow"] is True

tests/test_codex_auth.py

@@ -0,0 +1,294 @@
+"""Tests for ChatGPT (Codex) subscription auth: PKCE, token handling, store."""
+
+from __future__ import annotations
+
+import base64
+import hashlib
+import json
+import time
+from typing import TYPE_CHECKING, Any
+
+import pytest
+
+from strix.config import codex
+
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+
+def _fake_jwt(account_id: str) -> str:
+    def seg(obj: dict[str, Any]) -> str:
+        return base64.urlsafe_b64encode(json.dumps(obj).encode()).rstrip(b"=").decode()
+
+    header = seg({"alg": "none"})
+    payload = seg({"https://api.openai.com/auth": {"chatgpt_account_id": account_id}})
+    return f"{header}.{payload}.sig"
+
+
+@pytest.fixture(autouse=True)
+def _tmp_store(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> Path:
+    path = tmp_path / "home" / ".strix" / "subscription-auth.json"
+    monkeypatch.setattr(codex, "AUTH_PATH", path)
+    return path
+
+
+def test_pkce_challenge_matches_verifier_and_is_unpadded() -> None:
+    verifier, challenge = codex.generate_pkce()
+    expected = (
+        base64.urlsafe_b64encode(hashlib.sha256(verifier.encode()).digest()).rstrip(b"=").decode()
+    )
+    assert challenge == expected
+    assert "=" not in verifier
+    assert "=" not in challenge
+
+
+def test_authorize_url_carries_pkce_and_client() -> None:
+    url = codex.build_authorize_url("chal", "st8")
+    assert codex.AUTHORIZE_URL in url
+    assert "code_challenge=chal" in url
+    assert "code_challenge_method=S256" in url
+    assert f"client_id={codex.CLIENT_ID}" in url
+    assert "state=st8" in url
+
+
+@pytest.mark.parametrize(
+    ("value", "expected"),
+    [
+        ("http://localhost:1455/auth/callback?code=AAA&state=BBB", ("AAA", "BBB")),
+        ("AAA#BBB", ("AAA", "BBB")),
+        ("code=AAA&state=BBB", ("AAA", "BBB")),
+        ("AAA", ("AAA", None)),
+        ("", (None, None)),
+    ],
+)
+def test_parse_redirect_input(value: str, expected: tuple[str | None, str | None]) -> None:
+    assert codex.parse_redirect_input(value) == expected
+
+
+@pytest.mark.parametrize(
+    ("model", "expected"),
+    [
+        ("chatgpt/gpt-5.4", "gpt-5.4"),
+        ("ChatGPT/GPT-5.5", "GPT-5.5"),
+        ("  chatgpt/gpt-5.4  ", "gpt-5.4"),
+        ("openai/gpt-5.4", None),  # metered API path
+        ("anthropic/claude-opus-4-8", None),
+        ("gpt-5.4", None),
+        ("chatgpt/", None),
+        ("", None),
+        (None, None),
+    ],
+)
+def test_subscription_model(model: str | None, expected: str | None) -> None:
+    assert codex.subscription_model(model) == expected
+
+
+def test_auth_mode() -> None:
+    assert codex.auth_mode("chatgpt/gpt-5.4") == "subscription"
+    assert codex.auth_mode("openai/gpt-5.4") == "api_key"
+    assert codex.auth_mode("anthropic/claude-opus-4-8") == "api_key"
+    assert codex.auth_mode(None) == "api_key"
+
+
+def test_is_content_guardrail_error() -> None:
+    # The backend's real wording (from a live gpt-5.6-sol block).
+    raw = RuntimeError(
+        "This content was flagged for possible cybersecurity risk. If this seems "
+        "wrong, try rephrasing. To get authorized, join the Trusted Access for Cyber program."
+    )
+    assert codex.is_content_guardrail_error(raw) is True
+    # The already-typed error is recognized regardless of its message wording.
+    assert codex.is_content_guardrail_error(codex.CodexContentGuardrailError("gpt-5.6-sol")) is True
+    # Unrelated errors are not misclassified.
+    assert codex.is_content_guardrail_error(RuntimeError("rate limit exceeded")) is False
+
+
+def test_content_guardrail_error_message() -> None:
+    err = codex.CodexContentGuardrailError("gpt-5.6-sol")
+    assert err.model == "gpt-5.6-sol"
+    assert "gpt-5.6-sol" in str(err)
+    assert "STRIX_LLM" in str(err)
+
+
+def test_account_id_from_jwt() -> None:
+    assert codex._account_id_from_jwt(_fake_jwt("acct-42")) == "acct-42"
+    assert codex._account_id_from_jwt("not-a-jwt") is None
+    assert codex._account_id_from_jwt("") is None
+
+
+def test_store_roundtrip_and_logout() -> None:
+    assert codex.read_record() is None
+    assert codex.is_authenticated() is False
+
+    codex.save_record(
+        {
+            "type": "oauth",
+            "provider": "codex",
+            "access": _fake_jwt("acct-42"),
+            "refresh": "r1",
+            "account_id": "acct-42",
+            "expires_at": time.time() + 3600,
+        }
+    )
+    record = codex.read_record()
+    assert record is not None
+    assert record["account_id"] == "acct-42"
+    assert codex.is_authenticated() is True
+
+    codex.logout()
+    assert codex.read_record() is None
+    codex.logout()  # no-op when already gone
+
+
+def test_read_record_rejects_incomplete_records() -> None:
+    codex.save_record({"type": "oauth", "access": "a"})  # missing refresh/account
+    assert codex.read_record() is None
+    assert codex.is_authenticated() is False
+
+
+def test_get_valid_token_returns_stored_when_fresh(monkeypatch: pytest.MonkeyPatch) -> None:
+    def _boom(_payload: dict[str, str]) -> dict[str, Any]:
+        msg = "should not refresh a fresh token"
+        raise AssertionError(msg)
+
+    monkeypatch.setattr(codex, "_post_form", _boom)
+    codex.save_record(
+        {
+            "type": "oauth",
+            "provider": "codex",
+            "access": "access-fresh",
+            "refresh": "r1",
+            "account_id": "acct-42",
+            "expires_at": time.time() + 3600,
+        }
+    )
+    assert codex.get_valid_token() == ("access-fresh", "acct-42")
+
+
+def test_get_valid_token_refreshes_and_persists_rotation(monkeypatch: pytest.MonkeyPatch) -> None:
+    calls = {"n": 0}
+
+    def _fake_post(payload: dict[str, str]) -> dict[str, Any]:
+        calls["n"] += 1
+        assert payload["grant_type"] == "refresh_token"
+        assert payload["refresh_token"] == "r1"
+        return {"access_token": _fake_jwt("acct-42"), "refresh_token": "r2", "expires_in": 3600}
+
+    monkeypatch.setattr(codex, "_post_form", _fake_post)
+    codex.save_record(
+        {
+            "type": "oauth",
+            "provider": "codex",
+            "access": "stale",
+            "refresh": "r1",
+            "account_id": "acct-42",
+            "expires_at": time.time() - 10,  # already expired
+        }
+    )
+    _access, account_id = codex.get_valid_token()
+    assert calls["n"] == 1
+    assert account_id == "acct-42"
+    # Rotated refresh token was written back to the store.
+    record = codex.read_record()
+    assert record is not None
+    assert record["refresh"] == "r2"
+
+
+def test_get_valid_token_uses_token_rotated_by_another_process(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    # Simulate a parallel Strix process rotating the token while we wait for the
+    # refresh guard: the pre-guard read sees the stale token, the in-guard read
+    # sees the winner's fresh one, so we must NOT exchange the now-dead refresh.
+    records = [
+        {
+            "type": "oauth",
+            "provider": "codex",
+            "access": "stale",
+            "refresh": "r1",
+            "account_id": "acct",
+            "expires_at": time.time() - 10,
+        },
+        {
+            "type": "oauth",
+            "provider": "codex",
+            "access": "fresh-from-other-process",
+            "refresh": "r2",
+            "account_id": "acct",
+            "expires_at": time.time() + 3600,
+        },
+    ]
+    calls = {"n": 0}
+
+    def _fake_read() -> dict[str, Any]:
+        record = records[min(calls["n"], len(records) - 1)]
+        calls["n"] += 1
+        return record
+
+    def _boom(_payload: dict[str, str]) -> dict[str, Any]:
+        msg = "must not refresh a token another process already rotated"
+        raise AssertionError(msg)
+
+    monkeypatch.setattr(codex, "read_record", _fake_read)
+    monkeypatch.setattr(codex, "_post_form", _boom)
+
+    access, account_id = codex.get_valid_token()
+    assert access == "fresh-from-other-process"
+    assert account_id == "acct"
+
+
+def _expired_record(refresh: str, access: str) -> dict[str, Any]:
+    return {
+        "type": "oauth",
+        "provider": "codex",
+        "access": access,
+        "refresh": refresh,
+        "account_id": "acct-42",
+        "expires_at": time.time() - 10,
+    }
+
+
+def test_get_valid_token_recovers_when_refresh_loses_race(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    # Lock failed open: our in-guard read still saw the stale token, so we tried to
+    # refresh and lost the race (invalid_grant). By then a peer has saved a fresh
+    # token — recover from it instead of failing the scan on the dead one.
+    codex.save_record(_expired_record("r1", "stale"))
+
+    def _fake_post(_payload: dict[str, str]) -> dict[str, Any]:
+        codex.save_record(
+            {
+                "type": "oauth",
+                "provider": "codex",
+                "access": "fresh-from-peer",
+                "refresh": "r2",
+                "account_id": "acct-42",
+                "expires_at": time.time() + 3600,
+            }
+        )
+        raise codex.CodexAuthError("token_http_error", "HTTP 400: invalid_grant")
+
+    monkeypatch.setattr(codex, "_post_form", _fake_post)
+    assert codex.get_valid_token() == ("fresh-from-peer", "acct-42")
+
+
+def test_get_valid_token_reraises_refresh_error_without_rotation(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    # Refresh fails and no peer rotated the token: surface the error, don't mask it.
+    codex.save_record(_expired_record("r1", "stale"))
+
+    def _fake_post(_payload: dict[str, str]) -> dict[str, Any]:
+        raise codex.CodexAuthError("token_http_error", "HTTP 400: invalid_grant")
+
+    monkeypatch.setattr(codex, "_post_form", _fake_post)
+    with pytest.raises(codex.CodexAuthError):
+        codex.get_valid_token()
+
+
+def test_get_valid_token_raises_when_not_signed_in() -> None:
+    with pytest.raises(codex.CodexAuthError) as exc:
+        codex.get_valid_token()
+    assert exc.value.code == "not_authenticated"

tests/test_codex_streaming.py

@@ -0,0 +1,225 @@
+"""Regression test for the ChatGPT Codex backend's streaming requirement.
+
+The backend rejects non-streamed requests with ``{"detail": "Stream must be set
+to true"}``. ``_CodexResponsesModel`` must therefore issue a streamed request
+even from the non-streaming ``get_response`` path and aggregate the events into
+a single response. A local server that mimics that behaviour proves the wrapper
+works where the stock responses model would fail.
+"""
+
+from __future__ import annotations
+
+import json
+import threading
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from typing import TYPE_CHECKING, Any
+
+import pytest
+from agents.model_settings import ModelSettings
+from agents.models.interface import ModelTracing
+from agents.models.openai_responses import OpenAIResponsesModel
+from openai import AsyncOpenAI, BadRequestError
+from openai.types.responses import ResponseOutputMessage, ResponseOutputText
+
+from strix.config import codex
+from strix.config.models import _CodexResponsesModel
+
+
+if TYPE_CHECKING:
+    from collections.abc import AsyncIterator, Iterator
+
+
+def _response_payload() -> dict[str, Any]:
+    return {
+        "id": "resp_1",
+        "object": "response",
+        "created_at": 0,
+        "status": "completed",
+        "model": "gpt-5.5",
+        "output": [
+            {
+                "type": "message",
+                "id": "m1",
+                "status": "completed",
+                "role": "assistant",
+                "content": [{"type": "output_text", "text": "OK", "annotations": []}],
+            }
+        ],
+        "usage": {
+            "input_tokens": 1,
+            "output_tokens": 1,
+            "total_tokens": 2,
+            "input_tokens_details": {"cached_tokens": 0},
+            "output_tokens_details": {"reasoning_tokens": 0},
+        },
+        "parallel_tool_calls": False,
+        "tool_choice": "auto",
+        "tools": [],
+        "metadata": {},
+        "temperature": 1.0,
+        "top_p": 1.0,
+        "error": None,
+        "incomplete_details": None,
+        "instructions": None,
+        "max_output_tokens": None,
+    }
+
+
+_CAPTURED: dict[str, Any] = {}
+
+
+class _Handler(BaseHTTPRequestHandler):
+    def log_message(self, *args: Any) -> None:
+        pass
+
+    def do_POST(self) -> None:
+        length = int(self.headers.get("Content-Length", 0))
+        body = json.loads(self.rfile.read(length) or b"{}")
+        _CAPTURED.clear()
+        _CAPTURED.update(body)
+        if not body.get("stream"):
+            payload = json.dumps({"detail": "Stream must be set to true"}).encode()
+            self.send_response(400)
+            self.send_header("Content-Type", "application/json")
+            self.send_header("Content-Length", str(len(payload)))
+            self.end_headers()
+            self.wfile.write(payload)
+            return
+        event = {
+            "type": "response.completed",
+            "sequence_number": 0,
+            "response": _response_payload(),
+        }
+        frame = f"event: response.completed\ndata: {json.dumps(event)}\n\n".encode()
+        self.send_response(200)
+        self.send_header("Content-Type", "text/event-stream")
+        self.end_headers()
+        self.wfile.write(frame)
+
+
+@pytest.fixture
+def backend_url() -> Iterator[str]:
+    server = HTTPServer(("127.0.0.1", 0), _Handler)
+    thread = threading.Thread(target=server.serve_forever, daemon=True)
+    thread.start()
+    try:
+        yield f"http://127.0.0.1:{server.server_address[1]}/backend-api/codex"
+    finally:
+        server.shutdown()
+        server.server_close()
+
+
+def _client(base_url: str) -> AsyncOpenAI:
+    return AsyncOpenAI(api_key="tok", base_url=base_url)
+
+
+def _call_kwargs() -> dict[str, Any]:
+    return {
+        "system_instructions": "s",
+        "input": "hi",
+        "model_settings": ModelSettings(
+            store=False, response_include=["reasoning.encrypted_content"]
+        ),
+        "tools": [],
+        "output_schema": None,
+        "handoffs": [],
+        "tracing": ModelTracing.DISABLED,
+        "previous_response_id": None,
+        "conversation_id": None,
+        "prompt": None,
+    }
+
+
+@pytest.mark.asyncio
+async def test_stock_model_fails_on_non_streamed_backend(backend_url: str) -> None:
+    model = OpenAIResponsesModel(model="gpt-5.5", openai_client=_client(backend_url))
+    with pytest.raises(BadRequestError, match="Stream must be set to true"):
+        await model.get_response(**_call_kwargs())
+
+
+@pytest.mark.asyncio
+async def test_codex_model_streams_and_aggregates(backend_url: str) -> None:
+    model = _CodexResponsesModel(model="gpt-5.5", openai_client=_client(backend_url))
+    response = await model.get_response(**_call_kwargs())
+    message = response.output[0]
+    assert isinstance(message, ResponseOutputMessage)
+    text = message.content[0]
+    assert isinstance(text, ResponseOutputText)
+    assert text.text == "OK"
+    assert response.usage.total_tokens == 2
+
+
+class _TrackingStream:
+    """An async iterator that yields, then raises, and records if it was closed."""
+
+    def __init__(self, events: list[Any], error: Exception | None) -> None:
+        self._events = iter(events)
+        self._error = error
+        self.closed = False
+
+    def __aiter__(self) -> _TrackingStream:
+        return self
+
+    async def __anext__(self) -> Any:
+        try:
+            return next(self._events)
+        except StopIteration:
+            if self._error is not None:
+                raise self._error from None
+            raise StopAsyncIteration from None
+
+    async def aclose(self) -> None:
+        self.closed = True
+
+
+async def _drain(gen: AsyncIterator[Any]) -> list[Any]:
+    return [event async for event in gen]
+
+
+@pytest.mark.asyncio
+async def test_guarded_converts_guardrail_error() -> None:
+    # A mid-stream backend rejection becomes a typed, model-tagged error.
+    model = _CodexResponsesModel(model="gpt-5.6-sol", openai_client=_client("http://x/backend-api"))
+    guardrail = RuntimeError("This content was flagged for possible cybersecurity risk.")
+    stream = _TrackingStream(["a", "b"], guardrail)
+    with pytest.raises(codex.CodexContentGuardrailError) as exc_info:
+        await _drain(model._guarded(stream))
+    assert exc_info.value.model == "gpt-5.6-sol"
+    assert stream.closed is True  # underlying stream is released
+
+
+@pytest.mark.asyncio
+async def test_guarded_passes_through_other_errors() -> None:
+    # A non-guardrail error propagates unchanged (still not swallowed).
+    model = _CodexResponsesModel(model="gpt-5.5", openai_client=_client("http://x/backend-api"))
+    boom = RuntimeError("some unrelated failure")
+    stream = _TrackingStream(["a"], boom)
+    with pytest.raises(RuntimeError, match="some unrelated failure"):
+        await _drain(model._guarded(stream))
+    assert stream.closed is True
+
+
+@pytest.mark.asyncio
+async def test_guarded_yields_all_events_when_clean() -> None:
+    model = _CodexResponsesModel(model="gpt-5.4", openai_client=_client("http://x/backend-api"))
+    stream = _TrackingStream(["a", "b", "c"], None)
+    assert await _drain(model._guarded(stream)) == ["a", "b", "c"]
+    assert stream.closed is True
+
+
+@pytest.mark.asyncio
+async def test_codex_model_self_enforces_backend_requirements(backend_url: str) -> None:
+    # The caller passes ordinary settings; the model must impose the backend's
+    # requirements (stream, store=false, encrypted reasoning) and the configured
+    # reasoning effort itself.
+    model = _CodexResponsesModel(
+        model="gpt-5.4", openai_client=_client(backend_url), reasoning_effort="high"
+    )
+    kwargs = _call_kwargs()
+    kwargs["model_settings"] = ModelSettings()  # nothing special from the caller
+    await model.get_response(**kwargs)
+
+    assert _CAPTURED["stream"] is True
+    assert _CAPTURED["store"] is False
+    assert _CAPTURED["include"] == ["reasoning.encrypted_content"]
+    assert _CAPTURED["reasoning"] == {"effort": "high"}

tests/test_model_retry.py

@@ -13,12 +13,15 @@ import asyncio
 
 from agents.retry import ModelRetryNormalizedError, RetryPolicyContext
 
+from strix.config import codex
 from strix.config.models import DEFAULT_MODEL_RETRY, _retry_statusless_provider_errors
 
 
-def _context(normalized: ModelRetryNormalizedError) -> RetryPolicyContext:
+def _context(
+    normalized: ModelRetryNormalizedError, error: Exception | None = None
+) -> RetryPolicyContext:
     return RetryPolicyContext(
-        error=RuntimeError("boom"),
+        error=error or RuntimeError("boom"),
         attempt=1,
         max_retries=5,
         stream=True,
@@ -27,11 +30,11 @@ def _context(normalized: ModelRetryNormalizedError) -> RetryPolicyContext:
     )
 
 
-def _retries(normalized: ModelRetryNormalizedError) -> bool:
+def _retries(normalized: ModelRetryNormalizedError, error: Exception | None = None) -> bool:
     """Evaluate the composed DEFAULT_MODEL_RETRY policy for a normalized error."""
     policy = DEFAULT_MODEL_RETRY.policy
     assert policy is not None
-    decision = asyncio.run(policy(_context(normalized)))
+    decision = asyncio.run(policy(_context(normalized, error)))
     return bool(getattr(decision, "retry", decision))
 
 
@@ -63,6 +66,16 @@ def test_timeout_error_is_retried() -> None:
     assert _retries(ModelRetryNormalizedError(is_network_error=True)) is True
 
 
+def test_content_guardrail_error_is_not_retried() -> None:
+    # A guardrail block is status-less, so it would match the statusless policy;
+    # the guard must keep it from being retried (retrying never clears it).
+    guardrail = codex.CodexContentGuardrailError("gpt-5.6-sol")
+    assert _retries(ModelRetryNormalizedError(status_code=None), guardrail) is False
+    # A raw provider error carrying the backend's wording is excluded too.
+    raw = RuntimeError("This content was flagged for possible cybersecurity risk.")
+    assert _retries(ModelRetryNormalizedError(status_code=None), raw) is False
+
+
 def test_policy_helper_matches_statusless_only() -> None:
     assert _retry_statusless_provider_errors(_context(ModelRetryNormalizedError())) is True
     assert (

tests/test_models.py

@@ -42,9 +42,11 @@ def test_recommended_models_are_matched_case_insensitively() -> None:
     "model_name",
     [
         "gpt-5.5",
+        "chatgpt/gpt-5.4",
         "litellm/openai/gpt-5.4-pro",
         "azure_ai/gpt-5.5-pro",
         "bedrock_mantle/openai.gpt-5.5",
+        "anthropic/claude-opus-5",
         "anthropic/claude-opus-4-8",
         "anthropic.claude-opus-4-8",
         "anthropic/claude-opus-4-7",
@@ -60,8 +62,10 @@ def test_recommended_models_are_matched_case_insensitively() -> None:
         "deepseek/deepseek-reasoner",
         "dashscope/qwen3-max-2026-01-23",
         "qwen3.7-max",
+        "dashscope/qwen3.8-max",
         "moonshot/kimi-k2.6",
         "kimi-k2.7-code",
+        "moonshot/kimi-k3",
     ],
 )
 def test_frontier_model_families_are_accepted(model_name: str) -> None:

tests/test_runner_root_prompt.py

@@ -46,7 +46,8 @@ def _patch_engine_scaffold(
             reasoning_effort="high",
             force_required_tool_choice=False,
             timeout=300,
-        )
+        ),
+        runtime=types.SimpleNamespace(max_context_images=3),
     )
     monkeypatch.setattr(runner, "load_settings", lambda: settings)
     monkeypatch.setattr(runner, "configure_sdk_model_defaults", lambda _settings: None)

tests/test_usage_subscription.py

@@ -0,0 +1,46 @@
+"""Subscription runs track tokens but report zero cost."""
+
+from __future__ import annotations
+
+from agents.usage import Usage
+
+from strix.report.usage import LLMUsageLedger
+
+
+def _usage() -> Usage:
+    usage = Usage()
+    usage.requests = 1
+    usage.input_tokens = 1000
+    usage.output_tokens = 200
+    usage.total_tokens = 1200
+    return usage
+
+
+def test_zero_cost_ledger_keeps_tokens_but_reports_no_cost() -> None:
+    ledger = LLMUsageLedger()
+    ledger.zero_cost = True
+    ledger.record(agent_id="a", usage=_usage(), agent_name="strix", model="gpt-5.5")
+
+    record = ledger.to_record()
+    assert record["cost"] == 0.0
+    assert record["total_tokens"] == 1200
+    assert record["input_tokens"] == 1000
+    assert record["output_tokens"] == 200
+    assert ledger.total_cost == 0.0
+
+
+def test_zero_cost_ledger_ignores_observed_cost() -> None:
+    ledger = LLMUsageLedger()
+    ledger.zero_cost = True
+    ledger.record_observed_cost(4.20)
+    assert ledger.total_cost == 0.0
+
+
+def test_normal_ledger_still_estimates_cost() -> None:
+    # Sanity check the flag is opt-in: without it, an OpenAI-native model still
+    # accrues an estimated cost (proves zeroing is what suppresses it).
+    ledger = LLMUsageLedger()
+    ledger.record(agent_id="a", usage=_usage(), agent_name="strix", model="gpt-5.5")
+    assert ledger.to_record()["total_tokens"] == 1200
+    # Cost estimation depends on litellm's cost map; it should be >= 0 and not error.
+    assert ledger.total_cost >= 0.0