merge(search-with-proof): readiness probe, longer tool turns, 40-step tasks [skip ci]

Co-Authored-By: Claude Fable 5.1 <[email protected]>

# Conflicts:
#	benchmarks/ax/search_verify.py
#	benchmarks/search/pilot_run.py

benchmarks/ax/search_verify.py

@@ -9,6 +9,7 @@ import uuid
 from pathlib import Path
 
 from benchmarks.ax.batch import AXLane, CapacityDeadline, Item, SearchExecutorTask, Workload, validate_image
+from openultrasast.search.budget import verification_run_seconds
 from openultrasast.search.verify import RunObservation, SideRecord
 from openultrasast.search.verify_task import MAX_RESULT_BYTES, pack_checkout
 
@@ -66,7 +67,13 @@ class AXSideDispatcher:
             raise ValueError("verification accepts checkout and declarative inputs only")
         # The trusted host selects an executor for this side. Build/package work
         # finishes there before any verifier worker slot is acquired.
-        workload = Workload(self.image, frozenset({"VERIFY_INPUT_URL", "RESULT_URL"}), timeout_seconds * 6 + 60, validate_result, kind="search-verify")
+        workload = Workload(
+            self.image,
+            frozenset({"VERIFY_INPUT_URL", "RESULT_URL"}),
+            verification_run_seconds(demo, timeout_seconds) + 60,
+            validate_result,
+            kind="search-verify",
+        )
         lane = self.lane or AXLane(self.args, workload)
         spec = {"demo": demo, "family": family, "timeout_seconds": timeout_seconds}
         try:
@@ -113,4 +120,5 @@ class AXSideDispatcher:
             phase=last.phase,
             exit_code=last.exit_code,
             stderr=last.stderr,
+            stdout=last.stdout,
         )

benchmarks/search/pilot_run.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import argparse
+import copy
 import hashlib
 import io
 import json
@@ -34,7 +35,7 @@ from openultrasast.plane.memory import FileStore, open_store
 from openultrasast.provider.openrouter import OpenRouterChatClient
 from openultrasast.search import _sandbox
 from openultrasast.search.board import Board
-from openultrasast.search.budget import SearchBudget
+from openultrasast.search.budget import SearchBudget, verification_run_seconds, verification_timeout
 from openultrasast.search.coordinator import Coordinator
 from openultrasast.search.demo import FAMILY_ORACLES, load_demo, validate_oracle
 from openultrasast.search.executor import InProcessExecutor
@@ -142,6 +143,8 @@ class Pilot:
                 command_phase=diagnostic(command.phase),
                 exit_code=command.exit_code,
                 stderr=diagnostic(command.stderr, maximum=1500, private=private),
+                stdout=diagnostic(getattr(command, "stdout", ""), maximum=1500, private=private),
+                ready_seconds=getattr(command, "ready_seconds", 0),
             )
         self.record["failures"].append(entry)
 
@@ -224,26 +227,35 @@ class Pilot:
             wall = self.record["wall_seconds_by_phase"]
             wall["build_export"] = wall.get("build_export", 0) + time.monotonic() - before
 
-    def verification(self):
+    def verification(self, limits=None):
         if not FAMILY_ORACLES[self.family]:
             return VerificationRecord("no_oracle", (), 0, "no configured owned oracle")
         demo = load_demo(self.root / "demo")
         validate_oracle(self.family, demo["oracle"])
+        timeout = verification_timeout(demo, limits or self.worker.limits)
         if self.args.dry_run:
             sides = self.sides if self.args.side == "vulnerable" else (self.sides[1], self.sides[1])
-            return verify(*sides, self.root / "demo", self.family, oracle=demo["oracle"])
+            return verify(*sides, self.root / "demo", self.family, timeout_seconds=timeout, oracle=demo["oracle"])
         revisions = [self.row[self.args.side], self.row["fixed"]]
         if self.args.verify_lane == "ax":
             # Side paths are opaque lookup keys; no checkout or opposite revision reaches the brain.
             sides = [Side(self.root / str(i)) for i in range(2)]
             pins = {s.checkout: pin for s, pin in zip(sides, revisions, strict=True)}
             parent = self
+            verify_lane = copy.copy(self.lane)
+            verify_lane.workload = Workload(
+                self.args.image,
+                frozenset({"VERIFY_INPUT_URL", "RESULT_URL"}),
+                verification_run_seconds(demo, timeout) + 60,
+                validate_result,
+                kind="search-verify",
+            )
 
             class CountingLane:
                 def __call__(self, *a, **kw):
                     parent.admit_task()
                     parent.record["verify_tasks"] += 1
-                    return parent.lane(*a, **kw)
+                    return verify_lane(*a, **kw)
 
             dispatcher = AXSideDispatcher(
                 self.args,
@@ -256,6 +268,7 @@ class Pilot:
                 *sides,
                 self.root / "demo",
                 self.family,
+                timeout_seconds=timeout,
                 oracle=demo["oracle"],
                 task_dispatcher=dispatcher,
                 browser=BrowserExecutor(task_boundary=True) if demo["oracle"] == "xss" else None,
@@ -263,7 +276,7 @@ class Pilot:
         if _sandbox.isolation_mode() != "userns":
             raise _sandbox.IsolationUnavailable("VM verification requires userns")
         sides = []
-        spec = dict(demo=demo, family=self.family, timeout_seconds=5)
+        spec = dict(demo=demo, family=self.family, timeout_seconds=timeout)
         for revision in revisions:
             archive = self.prepare(revision, spec)
             bundle = self.root / ("built-" + uuid.uuid4().hex)
@@ -280,6 +293,7 @@ class Pilot:
             *sides,
             self.root / "verify-demo",
             self.family,
+            timeout_seconds=timeout,
             oracle=demo["oracle"],
             browser=BrowserExecutor() if demo["oracle"] == "xss" else None,
         )
@@ -293,7 +307,7 @@ class Pilot:
         end = "execution_failure"
         try:
             if task.step == "verify":
-                result = self.verification()
+                result = self.verification(task.limits)
                 self.verifications.append(result.outcome)
                 self.record["outcome"] = result.outcome
                 if self.args.side == "fixed" and result.sides:
@@ -393,7 +407,7 @@ def main(argv=None):
     parser.add_argument("--dry-run", action="store_true")
     parser.add_argument("--verify-lane", choices=("vm", "ax"), default="vm")
     parser.add_argument("--model", default=DEFAULT_DETECTOR_MODEL)
-    parser.add_argument("--max-tokens", type=int, default=1024)
+    parser.add_argument("--max-tokens", type=int, default=4096)
     parser.add_argument("--usd-per-mtok-in", type=float)
     parser.add_argument("--usd-per-mtok-out", type=float)
     parser.add_argument("--image")
@@ -501,7 +515,9 @@ def main(argv=None):
         try:
             store = open_store(args.board_memory) if args.board_memory else FileStore(root / "boards")
             if not args.dry_run:
-                lane = AXLane(args, Workload(args.image, frozenset({"VERIFY_INPUT_URL", "RESULT_URL"}), 900, validate_result, kind="search-verify"))
+                lane = AXLane(
+                    args, Workload(args.image, frozenset({"VERIFY_INPUT_URL", "RESULT_URL"}), 900, validate_result, kind="search-verify")
+                )
         except Exception as exc:
             aborted_reason = exception_reason(exc)
         for index in indices if aborted_reason is None else []:

src/openultrasast/search/budget.py

@@ -9,7 +9,11 @@ class SearchBudget:
     # Recalibrated 2026-10-05 on the first three measured searches: settled cost was ~$0.003 per model call,
     # but worst-case reservations (~$0.027 per call) exhausted the old $0.05 task allowance after 2-4 calls.
     spend_usd: float = 2.00
+    # Both ceilings bound each task: $0.40 and 40 model steps.
     max_call_usd: float = 0.40
+    max_steps: int = 40
+    http_start_seconds: int = 90
+    cli_start_seconds: int = 30
     reason_rounds: int = 4
     intents_per_round: int = 3
     demonstrations: int = 3
@@ -28,6 +32,9 @@ class SearchBudget:
             if value < 0 or (value == 0 and name not in ("retries", "spend_usd")):
                 raise ValueError(f"invalid {name}")
             if name in (
+                "max_steps",
+                "http_start_seconds",
+                "cli_start_seconds",
                 "reason_rounds",
                 "intents_per_round",
                 "demonstrations",
@@ -37,3 +44,17 @@ class SearchBudget:
                 "disk_bytes",
             ) and not isinstance(value, int):
                 raise ValueError(f"{name} must be an integer")
+
+
+def verification_timeout(schema: dict, limits: dict | None = None) -> int:
+    """Resolve the mode's start/command allowance from the search task budget."""
+    key = "http_start_seconds" if schema["start"]["mode"] == "http" else "cli_start_seconds"
+    value = (limits or {}).get(key, getattr(SearchBudget(), key))
+    if type(value) is not int or not 1 <= value <= 300:
+        raise ValueError("invalid verification timeout")
+    return value
+
+
+def verification_run_seconds(schema: dict, timeout: int) -> int:
+    # One readiness operation, all declared steps, and driver shutdown/handshake.
+    return timeout * (1 + len(schema["steps"])) + 5

src/openultrasast/search/verify.py

@@ -21,13 +21,13 @@ import time
 from collections.abc import Callable
 from concurrent.futures import ThreadPoolExecutor
 from contextlib import suppress
-from dataclasses import asdict, dataclass, field
+from dataclasses import asdict, dataclass, field, replace
 from pathlib import Path
 from typing import Any, cast
 
 from ..sandbox import SandboxJob, SandboxResult
 from . import _sandbox, task_storage
-from .budget import SearchBudget
+from .budget import SearchBudget, verification_run_seconds, verification_timeout
 from .demo import BUILD_RECIPES, FAMILY_ORACLES, RUNTIMES, capture_output, load_demo, substitute, validate_oracle, write_start_files
 from .oracles import BrowserExecutor, Canary, Oracle, oracle_for
 from .task_storage import CommandFailure, diagnostic, exception_reason
@@ -65,6 +65,7 @@ class SideRecord:
     phase: str = ""
     exit_code: int | None = None
     stderr: str = ""
+    stdout: str = ""
 
 
 @dataclass(frozen=True)
@@ -138,16 +139,17 @@ def _tree(root: Path, *, generated_links: bool = False) -> str:
 # Executed as verifier-owned code inside the same network sandbox as the app.
 # Redirects are refused: only the configured loopback application is reachable.
 _HTTP_DRIVER = r"""
-import http.client, importlib.util, json, os, socket, subprocess, sys, tempfile, time
+import http.client, importlib.util, json, os, socket, subprocess, sys, time
 spec = importlib.util.spec_from_file_location('demo', '/verifier-demo.py')
 demo = importlib.util.module_from_spec(spec)
 spec.loader.exec_module(demo)
 config = json.load(sys.stdin)
 schema = config['schema']
-app_errors = tempfile.TemporaryFile()
+app_errors = open("/scratch/" + config["errors"], "wb")
-app = subprocess.Popen(config['command'], stdin=subprocess.DEVNULL, stdout=subprocess.DEVNULL, stderr=app_errors)
+app_output = open("/scratch/" + config["output"], "wb")
+app = subprocess.Popen(config['command'], stdin=subprocess.DEVNULL, stdout=app_output, stderr=app_errors)
 try:
-    until = time.monotonic() + config['timeout'] / 2
+    until = time.monotonic() + config['timeout']
     while True:
         if app.poll() is not None:
             raise SystemExit(app.returncode or 1)
@@ -158,11 +160,16 @@ try:
         except OSError:
             if time.monotonic() >= until:
                 raise RuntimeError('HTTP readiness timed out')
-            time.sleep(0.03)
+            time.sleep(0.5)
+    conn = http.client.HTTPConnection('127.0.0.1', schema['start']['port'], timeout=max(0.01, until - time.monotonic()))
+    conn.request('GET', '/')
+    response = conn.getresponse()  # Any HTTP status establishes listening.
+    conn.close()
     with open('/scratch/' + config['ready'], 'w') as ready:
         ready.write('ready')
+    acknowledgment_deadline = time.monotonic() + 3
     while not os.path.exists('/scratch/' + config['go']):
-        if time.monotonic() >= until + config['timeout'] / 2:
+        if time.monotonic() >= acknowledgment_deadline:
             raise RuntimeError('verifier readiness acknowledgment timed out')
         time.sleep(0.01)
     captures, outputs = {}, []
@@ -171,7 +178,7 @@ try:
         if not path.startswith('/') or path.startswith('//') or '\r' in path or '\n' in path:
             raise ValueError('invalid expanded HTTP path')
         headers = {key: demo.substitute(value, captures) for key, value in step.get('headers', {}).items()}
-        conn = http.client.HTTPConnection('127.0.0.1', schema['start']['port'], timeout=config['timeout'] / 2)
+        conn = http.client.HTTPConnection('127.0.0.1', schema['start']['port'], timeout=config['timeout'])
         conn.request(step['method'], path, body=demo.substitute(step.get('body', ''), captures).encode(), headers=headers)
         response = conn.getresponse()
         body = response.read(1048577)
@@ -189,9 +196,8 @@ finally:
     except subprocess.TimeoutExpired:
         app.kill()
         app.wait()
-    app_errors.seek(max(0, app_errors.seek(0, 2) - 16384))
-    sys.stderr.write(app_errors.read().decode(errors="replace"))
     app_errors.close()
+    app_output.close()
 """
 
 
@@ -206,16 +212,34 @@ def _http_steps(
     nonce: str,
     runner: Callable[..., SandboxResult],
     ready_check: Callable[[], None],
+    diagnostics: dict[str, Any] | None = None,
 ) -> list[str]:
     from . import demo as demo_module
 
     ready, go = ".ready-" + secrets.token_hex(12), ".go-" + secrets.token_hex(12)
-    request = {"schema": schema, "command": command, "timeout": job(command).timeout_seconds, "nonce": nonce, "ready": ready, "go": go}
+    diagnostics = diagnostics if diagnostics is not None else {}
+    request = {
+        "schema": schema,
+        "command": command,
+        "timeout": job(command).timeout_seconds,
+        "nonce": nonce,
+        "ready": ready,
+        "go": go,
+        "errors": ready + ".err",
+        "output": ready + ".out",
+    }
+    diagnostics.update(stdout="", stderr="")
     acknowledged = False
+    started = time.monotonic()
+    driver_job = replace(
+        job(command),
+        command=("/usr/bin/python3", "-I", "-c", _HTTP_DRIVER),
+        timeout_seconds=verification_run_seconds(schema, job(command).timeout_seconds),
+    )
     with ThreadPoolExecutor(max_workers=1) as pool:
         future = pool.submit(
             runner,
-            job(("/usr/bin/python3", "-I", "-c", _HTTP_DRIVER)),
+            driver_job,
             scratch=scratch,
             mounts={**mounts, "/verifier-demo.py": Path(demo_module.__file__)},
             env=env,
@@ -223,19 +247,26 @@ def _http_steps(
             network=True,
             input_text=json.dumps(request),
         )
-        deadline = time.monotonic() + job(command).timeout_seconds
+        deadline = time.monotonic() + job(command).timeout_seconds + 5
         while not (scratch / ready).exists() and not future.done() and time.monotonic() < deadline:
             time.sleep(0.01)
+        diagnostics["ready_seconds"] = diagnostics.get("ready_seconds", 0) + time.monotonic() - started
         if (scratch / ready).exists():
             ready_check()
             (scratch / go).touch()
             acknowledged = True
         result = future.result()
+    for key, suffix in (("stdout", ".out"), ("stderr", ".err")):
+        path = scratch / (ready + suffix)
+        if path.is_file():
+            with path.open("rb") as stream:
+                stream.seek(max(0, path.stat().st_size - 16384))
+                diagnostics[key] = diagnostic(stream.read().decode(errors="replace"), maximum=1500)
     if not acknowledged or result.exit_code or result.timed_out:
         raise CommandFailure(
             "start" if not acknowledged else "run",
             result.exit_code,
-            result.stderr,
+            diagnostics.get("stderr", "") + "\n" + result.stderr,
             "HTTP application/steps failed or readiness not acknowledged",
         )
     outputs = json.loads(result.stdout)
@@ -249,6 +280,7 @@ def _side(side: Side, demo: Path, family: str, timeout: int, browser: BrowserExe
     runner = _run_task if task_boundary else _sandbox.run
     observations: list[RunObservation] = []
     timings = dict(build=0.0, ready=0.0, run=0.0)
+    app_diagnostics: dict[str, Any] = {}
 
     disk_peak = 0
     root: Path | None = None
@@ -270,14 +302,15 @@ def _side(side: Side, demo: Path, family: str, timeout: int, browser: BrowserExe
             time.monotonic() - started,
             diagnostic(reason, maximum=1500),
             timings["build"],
-            timings["ready"],
+            timings["ready"] + app_diagnostics.get("ready_seconds", 0),
-            timings["run"],
+            max(0.0, timings["run"] - app_diagnostics.get("ready_seconds", 0)),
             disk_peak,
             isolation_mode="task-boundary" if task_boundary else _sandbox.isolation_mode(),
             chromium_no_sandbox=bool(getattr(browser, "no_sandbox", False)),
             phase=phase,
             exit_code=result.exit_code if result is not None else None,
-            stderr=diagnostic(result.stderr, maximum=1500) if result is not None else "",
+            stderr=diagnostic(result.stderr if result is not None else app_diagnostics.get("stderr", ""), maximum=1500),
+            stdout=diagnostic(getattr(result, "stdout", app_diagnostics.get("stdout", "")), maximum=1500),
         )
 
     try:
@@ -380,6 +413,7 @@ def _side(side: Side, demo: Path, family: str, timeout: int, browser: BrowserExe
                             env=app_env,
                             namespace_pid=oracle.namespace_pid,
                         )
+                        app_diagnostics.update(stdout=ready.stdout, stderr=ready.stderr)
                         if ready.exit_code or ready.timed_out:
                             return record("could_not_run", "app readiness failed", "start", ready)
                     if family == "ssrf" and oracle.observe()[0]:
@@ -409,6 +443,7 @@ def _side(side: Side, demo: Path, family: str, timeout: int, browser: BrowserExe
                             canary.nonce,
                             runner,
                             http_ready_check,
+                            app_diagnostics,
                         )
                         for output in outputs:
                             timed("run", oracle.capture, output)
@@ -425,6 +460,7 @@ def _side(side: Side, demo: Path, family: str, timeout: int, browser: BrowserExe
                                 namespace_pid=oracle.namespace_pid,
                                 input_text=substitute(step.get("stdin", ""), captures),
                             )
+                            app_diagnostics.update(stdout=result.stdout, stderr=result.stderr)
                             if result.timed_out:
                                 return record("inconclusive", "app attack timed out")
                             if result.exit_code:
@@ -507,7 +543,7 @@ def _run_task(
         )
 
 
-def verify_side_task(side: Side, demo: Path, family: str, timeout_seconds: int = 5) -> SideRecord:
+def verify_side_task(side: Side, demo: Path, family: str, timeout_seconds: int | None = None) -> SideRecord:
     """Fresh AX task entry: one repetition, grouped by the trusted dispatcher.
 
     This explicit API is never selected as a fallback in the host verifier.
@@ -517,10 +553,10 @@ def verify_side_task(side: Side, demo: Path, family: str, timeout_seconds: int =
         return SideRecord("no_oracle", (), 0, "no owned oracle", isolation_mode="task-boundary")
     validate_oracle(family, schema["oracle"])
     kind = schema["oracle"]
-    if timeout_seconds < 1:
+    if timeout_seconds is not None and timeout_seconds < 1:
         raise ValueError("timeout must be positive")
     browser = BrowserExecutor(task_boundary=True) if kind == "xss" else None
-    return _side(side, demo, kind, timeout_seconds, browser, task_boundary=True)
+    return _side(side, demo, kind, timeout_seconds or verification_timeout(schema), browser, task_boundary=True)
 
 
 def verify(
@@ -529,7 +565,7 @@ def verify(
     artefact: Path,
     family: str,
     *,
-    timeout_seconds: int = 5,
+    timeout_seconds: int | None = None,
     oracle: str | None = None,
     browser: BrowserExecutor | None = None,
     task_dispatcher: Callable[..., SideRecord] | None = None,
@@ -538,7 +574,7 @@ def verify(
     started = time.monotonic()
     if family in FAMILY_ORACLES and not FAMILY_ORACLES[family]:
         return VerificationRecord("no_oracle", (), time.monotonic() - started, "no configured owned oracle for " + family)
-    if timeout_seconds < 1:
+    if timeout_seconds is not None and timeout_seconds < 1:
         raise ValueError("timeout must be positive")
     try:
         schema = load_demo(artefact)
@@ -550,6 +586,7 @@ def verify(
             raise ValueError("oracle runtime unavailable")
     except (OSError, ValueError) as exc:
         return VerificationRecord("inconclusive", (), time.monotonic() - started, "invalid declarative demo: " + exception_reason(exc))
+    timeout_seconds = timeout_seconds or verification_timeout(schema)
     mode = "task-boundary" if task_dispatcher is not None else _sandbox.isolation_mode()
     if task_dispatcher is None:
         try:

src/openultrasast/search/worker.py

@@ -174,8 +174,8 @@ class Worker:
         demo: Path,
         model: str = "scripted",
         prices: Prices | None = None,
-        max_output_tokens: int = 1024,
+        max_output_tokens: int = 4096,
-        max_steps: int = 32,
+        max_steps: int = SearchBudget().max_steps,
         executor: RepositoryExecutor | None = None,
         run_budget: SpendBudget | None = None,
     ) -> None:
@@ -262,6 +262,7 @@ class Worker:
             reservation.settle(actual)
             if run_reservation is not None:
                 run_reservation.settle(actual)
+            self.finish_reason = raw["choices"][0].get("finish_reason")
             return cast(dict[str, Any], raw["choices"][0]["message"])
         except Exception:
             if reservation._active:
@@ -356,14 +357,20 @@ class Worker:
         self.observations = []
         malformed = 0
         plain_text = 0
-        for index in range(self.max_steps):
+        shortening_retry = False
+        max_steps = min(self.max_steps, task.limits.get("max_steps", self.max_steps))
+        for index in range(max_steps):
             messages[0]["content"] = (
                 EXPLORE_PROMPT
                 + "\nSingle intent: "
                 + json.dumps(task.payload.get("description", ""))
-                + f"\nCalls left including this one: {self.max_steps - index}. "
+                + f"\nCalls left including this one: {max_steps - index}. "
                 "Finish now if only one remains; report confirmed observations or failure."
             )
+            if index and index % 10 == 0:
+                messages.append(
+                    {"role": "user", "content": f"Remaining steps: {max_steps - index}. Save a demo or finish within this budget."}
+                )
             message = self._call(messages, TOOLS, meter)
             messages.append(message)
             calls = message.get("tool_calls", [])
@@ -373,6 +380,27 @@ class Worker:
                     return {"status": "execution_failure", "failure": "second plain-text reply: " + str(message.get("content", ""))[:300]}
                 messages.append({"role": "user", "content": "Use a tool or call finish; plain text is not accepted."})
                 continue
+            truncated = False
+            for call in calls:
+                try:
+                    json.loads(call["function"]["arguments"])
+                except (ValueError, TypeError):
+                    truncated = getattr(self, "finish_reason", None) == "length"
+                    if truncated:
+                        break
+            if truncated and not shortening_retry:
+                # Reject the whole turn before any tools execute; retry once, fully metered.
+                shortening_retry = True
+                for call in calls:
+                    messages.append(
+                        {
+                            "role": "tool",
+                            "tool_call_id": call["id"],
+                            "content": json.dumps({"error": "Output truncated; no tools executed. Retry with a shorter call."}),
+                        }
+                    )
+                continue
+            shortening_retry = False
             finishes = [call for call in calls if isinstance(call, dict) and call.get("function", {}).get("name") == "finish"]
             if finishes:
                 if len(calls) > 1:
@@ -432,7 +460,7 @@ class Worker:
                 except (ValueError, OSError, KeyError, TypeError) as exc:
                     result = {"error": type(exc).__name__, "reason": exception_reason(exc)}
                 messages.append({"role": "tool", "tool_call_id": call["id"], "content": json.dumps(result)})
-        raise BudgetExhausted("tool step limit exceeded", usd=meter.spent, calls=self.max_steps, end_detail="worker_steps")
+        raise BudgetExhausted("tool step limit exceeded", usd=meter.spent, calls=max_steps, end_detail="worker_steps")
 
     def __call__(self, task: SearchTask) -> dict[str, Any]:
         self.budget_exhausted = False

tests/test_search_pilot_run.py

@@ -112,7 +112,7 @@ def test_fixed_verification_never_acquires_vulnerable_revision(tmp_path, monkeyp
 
     def stop(archive, bundle):
         bundle.mkdir()
-        (bundle / "spec.json").write_text(json.dumps({"demo": schema, "family": "path", "timeout_seconds": 5}))
+        (bundle / "spec.json").write_text(json.dumps({"demo": schema, "family": "path", "timeout_seconds": 30}))
 
     monkeypatch.setattr(pilot_run, "extract_checkout", stop)
     monkeypatch.setattr(pilot_run, "verify", lambda *args, **kwargs: (_ for _ in ()).throw(StopVerification()))
@@ -271,7 +271,7 @@ def test_fixed_effect_reporting_is_independent_of_differential(tmp_path, monkeyp
     pilot = make_pilot(tmp_path)
     side = SideRecord("observed", tuple(RunObservation(value, "owned", 0) for value in observed), 0, "completed")
     result = VerificationRecord("inconclusive", (side, side), 0, "no differential")
-    monkeypatch.setattr(pilot, "verification", lambda: result)
+    monkeypatch.setattr(pilot, "verification", lambda limits: result)
     task = SearchTask("task-1", "search-1", "verify", "", {}, asdict(SearchBudget()))
     assert pilot.dispatch(task)["outcome"] == "inconclusive"
     assert pilot.record["fixed_effect_observed"] is expected
@@ -538,3 +538,58 @@ def test_pilot_executor_deletes_task_when_submit_times_out(tmp_path, monkeypatch
     assert calls[-1][0][:2] == ("delete", "task")
     assert len(deleted) == 4
     assert pilot.record["executor_task_metrics"][-1]["end"] == "execution_failure"
+
+
+def test_start_failure_retains_sanitized_logs_and_readiness(tmp_path):
+    from openultrasast.search.verify import SideRecord
+
+    pilot = make_pilot(tmp_path)
+    side = SideRecord(
+        "could_not_run",
+        (),
+        1.5,
+        "timeout",
+        ready_seconds=1.1,
+        phase="start",
+        exit_code=-9,
+        stdout="x" * 2000 + " boot token=hidden",
+        stderr="y" * 2000 + " stderr password=hidden",
+    )
+    pilot.failure("verify", side.reason, side)
+    entry = pilot.record["failures"][-1]
+    assert entry["command_phase"] == "start" and entry["exit_code"] == -9
+    assert entry["ready_seconds"] == 1.1
+    assert len(entry["stdout"]) <= 1500 and len(entry["stderr"]) <= 1500
+    assert "hidden" not in entry["stdout"] + entry["stderr"]
+
+
+def test_ax_pilot_uses_verification_task_limits_for_deadline(tmp_path, monkeypatch):
+    from benchmarks.search import pilot_run
+    from openultrasast.search.probe import materialise
+
+    pilot = make_pilot(tmp_path)
+    pilot.args.verify_lane = "ax"
+    pilot.args.image = "image@sha256:" + "a" * 64
+    deadlines = []
+
+    class Lane:
+        workload = None
+
+        def __call__(self, *args):
+            deadlines.append(self.workload.deadline)
+
+    pilot.lane = Lane()
+    _, demos = materialise(tmp_path / "probe", "path")
+    schema = json.loads((demos["real"] / "demo.json").read_text())
+    (tmp_path / "demo/demo.json").write_text(json.dumps(schema))
+    captured = []
+
+    def verify(*args, **kwargs):
+        captured.append(kwargs["timeout_seconds"])
+        kwargs["task_dispatcher"].lane(None, None, 1)
+
+    monkeypatch.setattr(pilot_run, "verify", verify)
+    pilot.verification({"cli_start_seconds": 17})
+    assert captured == [17]
+    assert deadlines == [17 * (len(schema["steps"]) + 1) + 65]
+    assert pilot.lane.workload is None

tests/test_search_verify.py

@@ -381,7 +381,16 @@ def test_http_startup_effect_vetoed_before_request_steps(tmp_path):
 
     with pytest.raises(ValueError, match="startup effect"):
         _http_steps(
-            {}, (), lambda command: SandboxJob("", command, tmp_path, {}, 1, 256, 128), tmp_path, {}, {}, None, "nonce", runner, check
+            {"steps": []},
+            (),
+            lambda command: SandboxJob("", command, tmp_path, {}, 1, 256, 128),
+            tmp_path,
+            {},
+            {},
+            None,
+            "nonce",
+            runner,
+            check,
         )
     assert checks == [True]
 
@@ -463,7 +472,7 @@ def test_http_failure_keeps_exception_tail(tmp_path):
     traceback = "Traceback (most recent call last):\n" + "  frame\n" * 200 + "FileNotFoundError: /dev/null\n"
     with pytest.raises(ValueError, match="FileNotFoundError: /dev/null"):
         _http_steps(
-            {},
+            {"steps": []},
             ("/bin/true",),
             lambda command: SandboxJob("", command, tmp_path, {}, 1, 256, 128),
             tmp_path,
@@ -541,3 +550,150 @@ def test_app_requires_environment_and_data_files_on_both_sides(pair):
     assert result.outcome == "demonstrated", result
     assert all(len(side.runs) == 3 for side in result.sides)
     assert all(not (side.checkout / ".demo").exists() for side in sides)
+
+
+@pytest.mark.parametrize("delay,timeout,expected", [(8, None, "observed"), (8, 1, "could_not_run")])
+def test_delayed_http_readiness_and_timeout_diagnostics(pair, monkeypatch, delay, timeout, expected):
+    try:
+        with socket.socket() as sock:
+            sock.bind(("127.0.0.1", 0))
+            port = sock.getsockname()[1]
+    except PermissionError:
+        pytest.skip("host policy prohibits localhost sockets")
+    sides, demo = pair
+    app = f"""import sys, time
+from http.server import BaseHTTPRequestHandler, HTTPServer
+print('x' * 2000 + ' boot stdout token=hidden', flush=True)
+print('y' * 2000 + ' boot stderr password=hidden', file=sys.stderr, flush=True)
+time.sleep({delay})
+class Handler(BaseHTTPRequestHandler):
+ def do_GET(self):
+  self.send_response(503); self.end_headers()
+  self.wfile.write(b'listening')
+ def log_message(self, *args): pass
+HTTPServer(('127.0.0.1', {port}), Handler).serve_forever()
+"""
+    (sides[0].checkout / "app.py").write_text(app)
+
+    def change(value):
+        value["start"].update(mode="http", port=port)
+        value["steps"] = [{"type": "http", "method": "GET", "path": "/attack"}]
+
+    edit(demo, change)
+    result = verify_side_task(sides[0], demo, "path", timeout)
+    assert result.outcome == expected, result
+    assert "boot stdout" in result.stdout and "boot stderr" in result.stderr
+    assert "hidden" not in result.stdout + result.stderr
+    assert len(result.stdout) <= 1500 and len(result.stderr) <= 1500
+    if expected == "observed":
+        assert result.ready_seconds >= 8
+        assert len(result.runs) == 1
+    else:
+        assert result.phase == "start" and result.exit_code != 0
+        assert 0.9 <= result.ready_seconds < 4
+        assert not result.runs
+
+
+def test_killed_http_driver_retains_app_logs(pair, monkeypatch):
+    from openultrasast.sandbox import SandboxResult
+    from openultrasast.search import verify as module
+
+    sides, demo = pair
+    edit(
+        demo,
+        lambda value: (
+            value["start"].update(mode="http", port=18080),
+            value.update(steps=[{"type": "http", "method": "GET", "path": "/"}]),
+        ),
+    )
+
+    def killed(job, **kwargs):
+        config = json.loads(kwargs["input_text"])
+        for key in ("output", "errors"):
+            (kwargs["scratch"] / config[key]).write_text("app log token=hidden")
+        return SandboxResult(-9, "", "", True)
+
+    monkeypatch.setattr(module, "_run_task", killed)
+    result = verify_side_task(sides[0], demo, "path", 1)
+    assert result.phase == "start" and result.exit_code == -9
+    assert result.ready_seconds >= 0
+    assert result.stdout == result.stderr == "app log token=[redacted]"
+
+
+@pytest.mark.parametrize("timeout,ready", [(90, True), (1, False)])
+def test_http_driver_polls_eight_second_listener_before_get_and_steps(tmp_path, monkeypatch, capsys, timeout, ready):
+    import http.client
+    import io
+    import subprocess
+    import sys
+    import time
+    from types import SimpleNamespace
+
+    from openultrasast.search import demo as demo_module
+    from openultrasast.search.verify import _HTTP_DRIVER
+
+    tick = [0.0]
+    polls, requests = [], []
+    monkeypatch.setattr(time, "monotonic", lambda: tick[0])
+
+    def sleep(seconds):
+        assert seconds == 0.5
+        tick[0] += seconds
+
+    monkeypatch.setattr(time, "sleep", sleep)
+
+    def connect(address, timeout):
+        assert address == ("127.0.0.1", 18080)
+        polls.append(tick[0])
+        if tick[0] < 8:
+            raise ConnectionRefusedError()
+        return SimpleNamespace(close=lambda: None)
+
+    monkeypatch.setattr(socket, "create_connection", connect)
+
+    class HTTP:
+        def __init__(self, host, port, timeout):
+            assert tick[0] >= 8
+
+        def request(self, method, path, **kwargs):
+            requests.append((method, path))
+
+        def getresponse(self):
+            return SimpleNamespace(status=503, read=lambda size: b"listening")
+
+        def close(self):
+            pass
+
+    monkeypatch.setattr(http.client, "HTTPConnection", HTTP)
+
+    def app(*args, **kwargs):
+        kwargs["stdout"].write(b"boot stdout")
+        kwargs["stderr"].write(b"boot stderr")
+        return SimpleNamespace(poll=lambda: None, terminate=lambda: None, wait=lambda **kw: None)
+
+    monkeypatch.setattr(subprocess, "Popen", app)
+    config = dict(
+        schema={"start": {"port": 18080}, "steps": [{"method": "GET", "path": "/attack"}]},
+        command=["fake-app"],
+        timeout=timeout,
+        nonce="nonce",
+        ready="ready",
+        go="go",
+        errors="errors",
+        output="output",
+    )
+    (tmp_path / "go").touch()
+    monkeypatch.setattr(sys, "stdin", io.StringIO(json.dumps(config)))
+    driver = _HTTP_DRIVER.replace("/verifier-demo.py", demo_module.__file__).replace("/scratch/", str(tmp_path) + "/")
+    if ready:
+        exec(driver, {})
+        assert requests == [("GET", "/"), ("GET", "/attack")]
+        assert json.loads(capsys.readouterr().out) == ["listening"]
+        assert tick[0] == 8
+    else:
+        with pytest.raises(RuntimeError, match="readiness timed out"):
+            exec(driver, {})
+        assert not requests and not (tmp_path / "ready").exists()
+    assert polls == [i * 0.5 for i in range(len(polls))]
+    assert (tmp_path / "output").read_text() == "boot stdout"
+    assert (tmp_path / "errors").read_text() == "boot stderr"

tests/test_search_verify_transport.py

@@ -242,3 +242,26 @@ def test_transport_checks_family_and_demo_oracle(family, oracle):
     schema["oracle"] = "path" if oracle != "path" else "sql"
     with pytest.raises(ValueError, match="not allowed"):
         entry.validate_spec(spec)
+
+
+@pytest.mark.parametrize("mode,timeout", [("http", 90), ("cli", 30)])
+def test_verifier_deadline_uses_mode_budget_and_all_steps(tmp_path, monkeypatch, mode, timeout):
+    from openultrasast.search.budget import verification_run_seconds, verification_timeout
+
+    schema = demo()
+    schema["start"]["mode"] = mode
+    schema["steps"] *= 8
+    assert verification_timeout(schema) == timeout
+    assert verification_timeout(schema, {mode + "_start_seconds": 17}) == 17
+    workloads = []
+
+    def lane(args, workload):
+        workloads.append(workload)
+        return lambda *args: asdict(record())
+
+    monkeypatch.setattr(transport, "AXLane", lane)
+    dispatch = transport.AXSideDispatcher(
+        SimpleNamespace(), "image@sha256:" + "a" * 64, tmp_path / "out", prepare=lambda side, spec: b"archive"
+    )
+    dispatch(side=Side(tmp_path), demo=schema, family="injection", timeout_seconds=timeout, fresh_task=True)
+    assert workloads[0].deadline == verification_run_seconds(schema, timeout) + 60 == timeout * 9 + 65

tests/test_search_worker.py

@@ -327,7 +327,7 @@ def test_task_spend_finishes_with_confirmed_observations(tmp_path, with_evidence
     worker = Worker(model, repo=tmp_path, demo=tmp_path / "demo", prices=Prices(1, 1, 1), executor=InProcessExecutor(tmp_path))
     job = task("explore")
     # First reservation fits; the second includes the real file contents and cannot fit.
-    job.limits["max_call_usd"] = 0.01 if with_evidence else 0.000001
+    job.limits["max_call_usd"] = 0.014 if with_evidence else 0.000001
     result = worker(job)
     assert result["end_detail"] == result["budget_exhaustion"]["end_detail"] == "task_spend"
     assert len(model.calls) == int(with_evidence)
@@ -362,7 +362,7 @@ def test_calibrated_reservation_covers_reported_usage(tmp_path):
     expected_input = math.ceil(len(json.dumps([call["messages"], call["tools"]], ensure_ascii=False).encode()) / 2) + 1024 * (
         len(call["messages"]) + 1
     )
-    assert worker.metrics["reserved"] == pytest.approx((expected_input + 1024 * 2) / 1_000_000)
+    assert worker.metrics["reserved"] == pytest.approx((expected_input + 4096 * 2) / 1_000_000)
     assert worker.metrics["reserved"] >= worker.metrics["settled"] == pytest.approx(0.029048)
 
 
@@ -481,3 +481,36 @@ def test_plain_text_consumes_step(tmp_path):
     result = Worker(model, repo=tmp_path, demo=tmp_path / "demo", max_steps=1)(task("explore"))
     assert result["end_detail"] == "worker_steps"
     assert len(model.calls) == 1
+
+
+@pytest.mark.parametrize("finish_reason,expected_calls", [("length", 3), ("tool_calls", 2)])
+def test_truncated_arguments_retry_once_before_malformed(tmp_path, finish_reason, expected_calls):
+    broken = tool("run", command=["unused"])
+    broken["tool_calls"][0]["function"]["arguments"] = '{"command":["unterminated'
+
+    class LengthModel(StubModel):
+        def complete_chat_raw(self, **kwargs):
+            raw = super().complete_chat_raw(**kwargs)
+            raw["choices"][0]["finish_reason"] = finish_reason
+            return raw
+
+    model = LengthModel([broken, broken, tool("finish", failure="shorter call worked")])
+    worker = Worker(model, repo=tmp_path, demo=tmp_path / "demo")
+    result = worker(task("explore"))
+    assert len(model.calls) == expected_calls
+    assert all(call["extra_body"]["max_tokens"] == 4096 for call in model.calls)
+    if finish_reason == "length":
+        assert result["failure"] == "shorter call worked"
+        assert "shorter call" in model.calls[1]["messages"][-1]["content"]
+    else:
+        assert "two consecutive malformed" in result["failure"]
+
+
+def test_default_forty_steps_and_ten_step_reminders(tmp_path):
+    model = StubModel([tool("list_files")] * 40)
+    worker = Worker(model, repo=tmp_path, demo=tmp_path / "demo", executor=InProcessExecutor(tmp_path))
+    result = worker(task("explore"))
+    assert result["end_detail"] == "worker_steps"
+    assert len(model.calls) == SearchBudget().max_steps == 40
+    for index in (10, 20, 30):
+        assert f"Remaining steps: {40 - index}" in model.calls[index]["messages"][-1]["content"]