Merge 0107bd5387 into ff5c8cc8e4

strix/config/settings.py

@@ -112,6 +112,12 @@ class RuntimeSettings(BaseSettings):
     backend: str = Field(default="docker", alias="STRIX_RUNTIME_BACKEND")
     # Max screenshot/image tool outputs kept live per agent context (0 = none).
     max_context_images: int = Field(default=3, ge=0, alias="STRIX_MAX_CONTEXT_IMAGES")
+    # Wall-clock budget for the Caido readiness probe (session_manager ->
+    # bootstrap_caido -> _login_as_guest). The sandbox entrypoint chowns a
+    # large toolchain tree before starting caido-cli, which can take minutes
+    # on a loaded host or CI runner; a fixed attempt count gives up too early
+    # there. Raise this on slow/shared hosts instead of patching the retry loop.
+    caido_boot_wait_s: float = Field(default=180.0, gt=0, alias="STRIX_CAIDO_BOOT_WAIT_S")
 
 
 class TelemetrySettings(BaseSettings):

strix/runtime/caido_bootstrap.py

@@ -13,6 +13,7 @@ import asyncio
 import contextlib
 import json
 import logging
+import time
 from typing import TYPE_CHECKING
 
 
@@ -33,16 +34,26 @@ async def _login_as_guest(
     session: BaseSandboxSession,
     *,
     container_url: str,
-    attempts: int = 10,
+    max_wait_s: float = 180.0,
 ) -> str:
     """``session.exec`` curl to fetch a guest token; retry until ready.
 
     Caido's GraphQL listener may not be up the instant the container
-    starts. The retry loop also doubles as the Caido readiness probe —
-    no separate TCP healthcheck needed.
+    starts — the sandbox entrypoint chowns a large toolchain tree first,
+    which can take minutes on a loaded host or CI runner. The retry loop
+    also doubles as the Caido readiness probe — no separate TCP
+    healthcheck needed — so it is bounded by wall-clock time rather than
+    a fixed attempt count, to accommodate slow boots without making fast
+    hosts wait needlessly.
     """
+    deadline = time.monotonic() + max_wait_s
     last_err: str | None = None
-    for i in range(1, attempts + 1):
+    attempt = 0
+    while True:
+        remaining = deadline - time.monotonic()
+        if remaining <= 0:
+            break
+        attempt += 1
         result = await session.exec(
             "curl",
             "-fsS",
@@ -53,7 +64,7 @@ async def _login_as_guest(
             "-d",
             _LOGIN_AS_GUEST_BODY,
             f"{container_url}/graphql",
-            timeout=15,
+            timeout=min(15.0, remaining),
         )
         if result.ok():
             try:
@@ -72,10 +83,16 @@ async def _login_as_guest(
         else:
             stderr = result.stderr.decode("utf-8", errors="replace")[:200]
             last_err = f"curl exit {result.exit_code}: {stderr}"
-        logger.debug("loginAsGuest attempt %d/%d failed: %s", i, attempts, last_err)
-        await asyncio.sleep(min(2.0 * i, 8.0))
 
-    raise RuntimeError(f"loginAsGuest failed after {attempts} attempts: {last_err}")
+        remaining = deadline - time.monotonic()
+        if remaining <= 0:
+            break
+        logger.debug("loginAsGuest attempt %d failed: %s", attempt, last_err)
+        await asyncio.sleep(min(2.0 * attempt, 8.0, remaining))
+
+    raise RuntimeError(
+        f"loginAsGuest failed after {attempt} attempts over {max_wait_s:.0f}s: {last_err}"
+    )
 
 
 async def bootstrap_caido(
@@ -83,6 +100,7 @@ async def bootstrap_caido(
     *,
     host_url: str,
     container_url: str,
+    boot_wait_s: float = 180.0,
 ) -> Client:
     """Connect to the in-container Caido sidecar and select a fresh project."""
     # The Caido SDK (and its generated GraphQL schema) is slow to import and is
@@ -93,7 +111,9 @@ async def bootstrap_caido(
 
     logger.info("Bootstrapping Caido client (host=%s, container=%s)", host_url, container_url)
 
-    access_token = await _login_as_guest(session, container_url=container_url)
+    access_token = await _login_as_guest(
+        session, container_url=container_url, max_wait_s=boot_wait_s
+    )
 
     client = Client(host_url, auth=TokenAuthOptions(token=access_token))
     try:

strix/runtime/session_manager.py

@@ -359,6 +359,7 @@ async def create_or_reuse(
                     session,
                     host_url=host_caido_url,
                     container_url=container_caido_url,
+                    boot_wait_s=load_settings().runtime.caido_boot_wait_s,
                 ),
                 name=f"caido-bootstrap-{scan_id}",
             )

tests/test_caido_bootstrap.py

@@ -1,20 +1,134 @@
-"""A bootstrap that dies mid-setup must not leave its transport behind.
+"""Caido bootstrap tests.
 
-The bootstrap now runs concurrently with the scan start, so teardown can
-cancel it at any await — including inside ``Client.connect()``, where the
-client exists but no caller will ever see it to close it.
+Covers two concerns:
+
+* the guest-login readiness probe is bounded by a configurable wall-clock
+  budget (not a fixed attempt count), and its per-attempt curl timeout never
+  runs past that deadline; and
+* a bootstrap that dies mid-setup must not leave its transport behind. The
+  bootstrap now runs concurrently with the scan start, so teardown can cancel
+  it at any await -- including inside ``Client.connect()``, where the client
+  exists but no caller will ever see it to close it.
 """
 
 from __future__ import annotations
 
 import asyncio
+import json
 import sys
 import types
+from dataclasses import dataclass, field
 from typing import Any
 
 import pytest
 
-from strix.runtime.caido_bootstrap import bootstrap_caido
+from strix.runtime.caido_bootstrap import _login_as_guest, bootstrap_caido
+
+
+@dataclass
+class _FakeResult:
+    exit_code: int
+    stdout: str = ""
+    stderr: bytes = b""
+
+    def ok(self) -> bool:
+        return self.exit_code == 0
+
+
+@dataclass
+class _FakeSession:
+    """Stands in for BaseSandboxSession; returns queued exec() results."""
+
+    results: list[_FakeResult]
+    sleeps: list[float] = field(default_factory=list)
+    timeouts: list[float] = field(default_factory=list)
+    calls: int = 0
+
+    async def exec(self, *_args: Any, **kwargs: Any) -> _FakeResult:
+        result = self.results[min(self.calls, len(self.results) - 1)]
+        self.calls += 1
+        self.timeouts.append(kwargs["timeout"])
+        return result
+
+
+_FAKE_TOKEN = "guest-token"  # noqa: S105 - test fixture, not a real credential
+
+
+def _token_result() -> _FakeResult:
+    body = {"data": {"loginAsGuest": {"token": {"accessToken": _FAKE_TOKEN}}}}
+    return _FakeResult(exit_code=0, stdout=json.dumps(body))
+
+
+def _refused_result() -> _FakeResult:
+    return _FakeResult(exit_code=7, stderr=b"curl: (7) Failed to connect")
+
+
+async def test_login_as_guest_succeeds_once_caido_is_up(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr("strix.runtime.caido_bootstrap.asyncio.sleep", _no_sleep)
+    session = _FakeSession(results=[_refused_result(), _refused_result(), _token_result()])
+
+    token = await _login_as_guest(session, container_url="http://127.0.0.1:48080", max_wait_s=30)
+
+    assert token == _FAKE_TOKEN
+    assert session.calls == 3
+
+
+async def test_login_as_guest_respects_configurable_deadline(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """A slow-booting sandbox (connection refused throughout) should be given
+    the full configured budget rather than giving up after a fixed attempt
+    count, and the raised error should report elapsed time, not attempts.
+    """
+    monkeypatch.setattr("strix.runtime.caido_bootstrap.asyncio.sleep", _no_sleep)
+
+    # Fake clock: each read advances by 25s, so a 180s budget survives
+    # ~7 attempts (the old fixed-10-attempt loop only had ~68s total).
+    fake_now = [0.0]
+
+    def _fake_monotonic() -> float:
+        fake_now[0] += 25.0
+        return fake_now[0]
+
+    monkeypatch.setattr("strix.runtime.caido_bootstrap.time.monotonic", _fake_monotonic)
+    session = _FakeSession(results=[_refused_result()])
+
+    with pytest.raises(RuntimeError) as exc_info:
+        await _login_as_guest(session, container_url="http://127.0.0.1:48080", max_wait_s=180)
+
+    assert "curl exit 7" in str(exc_info.value)
+    assert "180s" in str(exc_info.value)
+    assert session.calls >= 2
+
+
+async def test_login_as_guest_caps_final_attempt_timeout_to_remaining_budget(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """The per-attempt curl timeout must never let a single attempt run past
+    the overall deadline -- otherwise a request started just before the
+    deadline can block session creation for up to another 15s beyond the
+    configured STRIX_CAIDO_BOOT_WAIT_S budget.
+    """
+    monkeypatch.setattr("strix.runtime.caido_bootstrap.asyncio.sleep", _no_sleep)
+
+    # First read establishes the deadline; second leaves 3s remaining for the
+    # attempt; every read after that is past the deadline so the loop exits.
+    fake_now = iter([0.0, 7.0])
+
+    def _fake_monotonic() -> float:
+        return next(fake_now, 11.0)
+
+    monkeypatch.setattr("strix.runtime.caido_bootstrap.time.monotonic", _fake_monotonic)
+    session = _FakeSession(results=[_refused_result()])
+
+    with pytest.raises(RuntimeError):
+        await _login_as_guest(session, container_url="http://127.0.0.1:48080", max_wait_s=10)
+
+    assert session.timeouts == [3.0]
+
+
+async def _no_sleep(_seconds: float) -> None:
+    """asyncio.sleep stub so deadline-bound tests run instantly."""
 
 
 class _FakeExecResult:
@@ -28,7 +142,7 @@ class _FakeExecResult:
         return True
 
 
-class _FakeSession:
+class _FakeLoginSession:
     async def exec(self, *_args: Any, **_kwargs: Any) -> _FakeExecResult:
         return _FakeExecResult('{"data":{"loginAsGuest":{"token":{"accessToken":"t"}}}}')
 
@@ -62,7 +176,7 @@ async def _bootstrap_expecting(
 
     with pytest.raises(type(error)):
         await bootstrap_caido(
-            _FakeSession(),  # type: ignore[arg-type]
+            _FakeLoginSession(),  # type: ignore[arg-type]
             host_url="http://host",
             container_url="http://container",
         )