search-with-proof: diagnosable failures, pip install race, no duplicate tasks

Reasons carry type plus a sanitised message tail; command failures carry phase,
exit code and bounded stderr. Scratch walk tolerates directories pip removes
mid-walk; writable /workspace/home and runtime paths. Capacity retries delete a
possibly created task before resubmitting. Docker smoke checks the image first.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

benchmarks/ax/batch.py

@@ -25,6 +25,7 @@ import yaml
 
 from openultrasast.config import load_dotenv
 from openultrasast.plane.memory import PRESIGNED_ARCHIVE_SUFFIX, S3Store, open_store
+from openultrasast.search.task_storage import diagnostic, exception_reason
 
 
 class EgressError(ValueError):
@@ -54,6 +55,12 @@ def apply_when_available(lane, document, end):
             lane.ax("apply", "-f", "-", manifest=document, timeout=min(30, end - lane.clock()))
             return
         except NoFreeWorkers:
+            # A failed apply is not proof that the server did not create a task.
+            # Never send another create while the preceding task can still run.
+            try:
+                lane.ax("delete", "task", document["metadata"]["name"])
+            except Exception as exc:
+                raise RuntimeError("capacity cleanup failed: " + exception_reason(exc)) from None
             lane.pause(min(random.uniform(15, 30), max(0, end - lane.clock())))
     raise CapacityDeadline("no free workers available before deadline")
 
@@ -185,7 +192,11 @@ class AXLane:
             if slot:
                 AX_SLOTS.release()
         if done is None or done.returncode:
-            return {"status": "instrument_failure", "ax_attempts": attempts}
+            return {
+                "status": "instrument_failure",
+                "reason": diagnostic(done.stderr) if done is not None else "TimeoutError: task result deadline",
+                "ax_attempts": attempts,
+            }
         record = json.loads((output / "result.json").read_text())
         return {**record, "worker_ip": worker, "ax_attempts": attempts}
 
@@ -203,9 +214,11 @@ class AXLane:
         )
         if done.returncode:
             if args[0] == "apply" and "no free workers available" in (done.stdout + done.stderr).lower():
-                raise NoFreeWorkers
-            # AX errors may echo manifest URLs; report only verb and exit code.
-            raise ValueError(f"ax {args[0]} failed (exit {done.returncode})")
+                raise NoFreeWorkers("no free workers available")
+            if args[0] == "delete" and "notfound" in (done.stdout + done.stderr).lower().replace(" ", ""):
+                return done
+            # AX errors may echo manifest URLs; sanitize before recording.
+            raise ValueError(f"ax {args[0]} failed (exit {done.returncode}): " + diagnostic(done.stderr))
         return done
 
     def _execute_attempt(self, item, output, attempt, attempts):
@@ -266,13 +279,13 @@ class AXLane:
                     raise SandboxFailure
                 self.pause(min(2, max(0, end - self.clock())))
             return None, worker_ip
-        except (subprocess.TimeoutExpired, TimeoutError):
-            return None, worker_ip
+        except (subprocess.TimeoutExpired, TimeoutError) as exc:
+            return subprocess.CompletedProcess([], 1, "", exception_reason(exc)), worker_ip
         except (SandboxFailure, CapacityDeadline):
             raise
         except Exception as exc:
             # Never expose exception strings which may contain object-scoped bearer URLs.
-            detail = str(exc) if isinstance(exc, EgressError) else type(exc).__name__
+            detail = exception_reason(exc)
             return subprocess.CompletedProcess([], 1, "", "AX transport failed: " + detail), worker_ip
         finally:
             attempts.append({"name": name, "phase": phase, "seconds": self.clock() - started})
@@ -283,12 +296,12 @@ class AXLane:
                 except Exception as exc:
                     # A successful list already confirmed a disappeared task is gone.
                     if phase != "Missing":
-                        errors.append("task: " + type(exc).__name__)
+                        errors.append("task: " + exception_reason(exc))
             for key in keys:
                 try:
                     self.store._delete(key)
                 except Exception as exc:
-                    errors.append("object: " + type(exc).__name__)
+                    errors.append("object: " + exception_reason(exc))
             if errors:
                 raise RuntimeError("AX cleanup failed; stop dispatching: " + ", ".join(errors))
 
@@ -414,6 +427,9 @@ def dispatch(items, workload, out, executors, *, lanes=("ax", "ax", "docker"), o
     if not isinstance(max_attempts, int) or not 1 <= max_attempts <= 2:
         raise ValueError("batch attempts must be one or two")
     out.mkdir(parents=True, exist_ok=True)
+    items = list(items)
+    if len({item.id for item in items}) != len(items):
+        raise ValueError("duplicate dispatch item id")
     pending, fallbacks = [], []
 
     def path(item):
@@ -465,15 +481,7 @@ def dispatch(items, workload, out, executors, *, lanes=("ax", "ax", "docker"), o
 
 def docker_diagnostic(text):
     """Bound diagnostics and remove bearer URLs, environment secrets and assignments."""
-    text = str(text or "")
-    for key, value in os.environ.items():
-        if len(value) >= 4 and any(word in key.upper() for word in ("TOKEN", "SECRET", "PASSWORD", "CREDENTIAL", "KEY")):
-            text = text.replace(value, "[redacted]")
-    text = re.sub(r"https?://[^\s\"']+", "[redacted-url]", text)
-    text = re.sub(
-        r"(?i)(authorization|bearer|password|token|secret|signature|credential|[\w-]*key)\s*[:= ]\s*[^\s,;]+", r"\1=[redacted]", text
-    )
-    return "\n".join(text.splitlines()[-10:])[-4000:]
+    return diagnostic(text or "", maximum=4000)
 
 
 class DockerLane(AXLane):
@@ -510,12 +518,12 @@ class DockerLane(AXLane):
             done = self.runner(command, capture_output=True, text=True, timeout=timeout)
         except (OSError, subprocess.SubprocessError) as exc:
             self._diagnostics.append(
-                {"step": command[1], "exit_code": None, "stderr": docker_diagnostic(getattr(exc, "stderr", "") or type(exc).__name__)}
+                {"step": command[1], "exit_code": None, "stderr": docker_diagnostic(getattr(exc, "stderr", "") or exception_reason(exc))}
             )
             raise
         if done.returncode:
             self._diagnostics.append({"step": command[1], "exit_code": done.returncode, "stderr": docker_diagnostic(done.stderr)})
-            raise ValueError(f"docker {command[1]} failed (exit {done.returncode})")
+            raise ValueError(f"docker {command[1]} failed (exit {done.returncode}): " + diagnostic(done.stderr))
         return done
 
     def _prepare_image(self):
@@ -625,6 +633,8 @@ class SearchExecutorTask:
         expires = timedelta(seconds=self.deadline + 600)
         store = self.lane.store
         try:
+            if isinstance(self.lane, DockerLane):
+                self.lane._run_docker(["docker", "image", "inspect", self.lane.workload.image], min(30, self.deadline))
             store._put(self.keys[0], self.archive)
             urls = {
                 "REPO_URL": store.presign_get(self.keys[0], expires),
@@ -684,12 +694,12 @@ class SearchExecutorTask:
             try:
                 self.lane.ax("delete", "task", self.name)
             except Exception as error:
-                errors.append(type(error).__name__)
+                errors.append(exception_reason(error))
         for key in self.keys:
             try:
                 self.lane.store._delete(key)
             except Exception as error:
-                errors.append(type(error).__name__)
+                errors.append(exception_reason(error))
         if self.slot and not errors:
             AX_SLOTS.release()
             self.slot = False

benchmarks/search/executor_smoke.py

@@ -1,7 +1,7 @@
 """One no-model executor task; run as python -m benchmarks.search.executor_smoke.
 
 Runtime acquisition/install needs network inside the executor. Unit tests are offline.
-Records contain measurements, never command output or presigned bearer URLs.
+Records contain measurements and sanitized failure tails, never presigned bearer URLs.
 """
 
 from __future__ import annotations
@@ -16,6 +16,7 @@ from pathlib import Path
 from urllib.parse import urlsplit
 
 from benchmarks.ax.batch import AXLane, DockerLane, SearchExecutorTask, Workload, validate_image
+from openultrasast.search.task_storage import diagnostic, exception_reason
 
 # URL and pin are positional shell arguments, never interpolated into shell code.
 # Retry the initial acquisition for delayed task egress activation (at most 60 s).
@@ -55,6 +56,8 @@ def run_smoke(lane, repo_url, commit, *, deadline=900):
         "scratch_measurement": "sampled workspace allocated bytes plus command logs",
         "model_calls": 0,
     }
+    parts = urlsplit(repo_url)
+    private = tuple(p for p in parts.path.split("/") if p) + (parts.path.rsplit("/", 1)[-1].removesuffix(".git"),)
     started = time.monotonic()
     try:
         parts = urlsplit(repo_url)
@@ -77,9 +80,18 @@ def run_smoke(lane, repo_url, commit, *, deadline=900):
                 if name == "run":
                     args = {**args, "timeout_seconds": timeout}
                 before = time.monotonic()
-                result = task.client.submit(
-                    name, args, timeout_seconds=timeout, limits={"memory_bytes": 2 * 1024**3, "disk_bytes": 2 * 1024**3}
-                )
+                try:
+                    result = task.client.submit(
+                        name, args, timeout_seconds=timeout, limits={"memory_bytes": 2 * 1024**3, "disk_bytes": 2 * 1024**3}
+                    )
+                except Exception as exc:
+                    result = {
+                        "error": type(exc).__name__,
+                        "reason": exception_reason(exc, private=private),
+                        "phase": "timeout" if isinstance(exc, TimeoutError) else "result upload",
+                        "exit_code": None,
+                        "stderr": "",
+                    }
                 output_bytes = sum(result.get(k + "_bytes", len(result.get(k, "").encode())) for k in ("stdout", "stderr"))
                 if name != "run":
                     output_bytes = len(json.dumps(result).encode())
@@ -95,12 +107,27 @@ def run_smoke(lane, repo_url, commit, *, deadline=900):
                 }
                 if "input_bytes" in result:
                     row["input_bytes"] = result["input_bytes"]
+                failed = (
+                    result.get("error")
+                    or result.get("timed_out")
+                    or result.get("status") == "could_not_build"
+                    or (name == "run" and result.get("exit_code") != 0)
+                )
+                if failed:
+                    row.update(
+                        phase=result.get("phase", "timeout" if result.get("timed_out") else "run"),
+                        reason=diagnostic(
+                            result.get("reason") or result.get("error") or f"CommandFailure: exit code {result.get('exit_code')}",
+                            private=private,
+                        ),
+                        stderr=diagnostic(result.get("stderr", ""), maximum=4000, private=private),
+                    )
                 record["commands"].append(row)
                 record["scratch_peak_bytes"] = max(record["scratch_peak_bytes"], result.get("scratch_peak_bytes", 0))
                 if result.get("error") or result.get("timed_out") or result.get("status") == "could_not_build":
-                    raise ValueError(step + " executor failure")
+                    raise ValueError(step + " executor failure: " + row["reason"])
                 if name == "run" and result.get("exit_code") != 0:
-                    raise ValueError(step + " nonzero or missing exit code")
+                    raise ValueError(step + " nonzero or missing exit code: " + row["reason"])
                 if step == "install" and output_bytes == 0 and row["wall_seconds"] < 1:
                     raise ValueError("empty fast install: exit 0 with empty output in under 1 s")
                 return result
@@ -130,28 +157,7 @@ def run_smoke(lane, repo_url, commit, *, deadline=900):
             record["status"] = "ok"
     except Exception as exc:
         record["status"] = "instrument_failure"
-        # Only our fixed diagnostic strings are safe; transport errors may contain URLs.
-        record["reason"] = (
-            str(exc)
-            if isinstance(exc, ValueError)
-            and str(exc)
-            in {
-                "credential-free HTTPS repository URL required",
-                "full lowercase commit SHA required",
-                "no recognised install manifest",
-                "clone yielded zero files",
-                "file read lacks input size",
-                "trivial command did not print 1",
-                "stop not acknowledged",
-                "empty fast install: exit 0 with empty output in under 1 s",
-                *(
-                    step + suffix
-                    for step in ("clone", "list_files", "read_file", "grep", "install", "trivial", "stop")
-                    for suffix in (" executor failure", " nonzero or missing exit code")
-                ),
-            }
-            else type(exc).__name__
-        )
+        record["reason"] = exception_reason(exc, private=private)
     record["wall_seconds"] = time.monotonic() - started
     return record
 
@@ -173,7 +179,7 @@ def main(argv=None):
         lane = (DockerLane if args.lane == "docker" else AXLane)(args, Workload(args.image, frozenset(), args.deadline, None))
         record = run_smoke(lane, args.repo_url, args.commit, deadline=args.deadline)
     except Exception as exc:
-        record = {"status": "instrument_failure", "reason": "lane setup: " + type(exc).__name__}
+        record = {"status": "instrument_failure", "reason": "lane setup: " + exception_reason(exc)}
     args.out.parent.mkdir(parents=True, exist_ok=True)
     args.out.write_text(json.dumps(record, indent=2, sort_keys=True) + "\n")
     print(json.dumps(record))

benchmarks/search/pilot_select.py

@@ -19,6 +19,7 @@ from urllib.parse import urlsplit
 
 from benchmarks.unseen import eligibility, ledger
 from openultrasast.model.taxonomy import load_families
+from openultrasast.search.task_storage import exception_reason
 
 LANGUAGES = ("python", "javascript", "typescript", "php", "java")
 FAMILIES = ("injection", "path", "output_encoding", "untrusted_destination")
@@ -352,7 +353,16 @@ def main(argv=None) -> int:
     except Exception as exc:
         if created:
             private.unlink(missing_ok=True)
-        print(json.dumps({"status": "instrument_failure", "reason": type(exc).__name__}))
+        print(
+            json.dumps(
+                {
+                    "status": "instrument_failure",
+                    "reason": exception_reason(
+                        exc, private=tuple(part for row in locals().get("rows", []) for part in str(row.get("repo", "")).split("/") if part)
+                    ),
+                }
+            )
+        )
         return 1
 
 

plane/Dockerfile.search-task

@@ -28,5 +28,8 @@ RUN python3 -m venv /venv \
 RUN groupadd --gid 10001 search \
  && useradd --create-home --uid 10001 --gid 10001 search
 # gVisor is the outer boundary; retain inner bwrap when its capability probe succeeds.
+# The Python base image sets GPG_KEY to the public CPython release-signing fingerprint (not a secret).
+# Clear it so the executor's guard against inherited credentials stays strict and still passes.
+ENV GPG_KEY=
 USER root
 # Do not replace the AX runner's ENTRYPOINT or CMD.

src/openultrasast/search/executor.py

@@ -34,7 +34,7 @@ from . import _sandbox, task_storage
 
 MAX_COMMAND_BYTES = 65536
 MAX_RESULT_BYTES = 65536
-SAFE_ENV = {"PATH": "/usr/bin:/bin", "LANG": "C.UTF-8"}
+SAFE_ENV = {"PATH": "/venv/bin:/opt/java/openjdk/bin:/usr/local/bin:/usr/bin:/bin", "LANG": "C.UTF-8", "JAVA_HOME": "/opt/java/openjdk"}
 
 
 def clean_environment(_source: Mapping[str, str] | None = None) -> dict[str, str]:
@@ -111,7 +111,7 @@ class InProcessExecutor:
     ) -> dict[str, Any]:
         result = self.execute(Command(self.seq + 1, name, args, timeout_seconds, limits or {}))
         if result.get("error") in {"ValueError", "KeyError", "TypeError", "FileNotFoundError", "OSError"}:
-            raise ValueError("repository command rejected")
+            raise ValueError(result.get("reason", "repository command rejected"))
         return result
 
     def execute(self, command: Command) -> dict[str, Any]:
@@ -126,13 +126,28 @@ class InProcessExecutor:
             raise ValueError("out of order sequence")
         started = time.monotonic()
         self.deadline = started + command.timeout_seconds
+        self.phase = "run"
         try:
             result = self._execute(command)
-        except task_storage.ScratchLimit:
+        except task_storage.ScratchLimit as exc:
             self.stopped = True
-            result = {"status": "could_not_build", "reason": "scratch limit", "stopped": True}
-        except (ValueError, OSError, KeyError, TypeError, TimeoutError) as exc:
-            result = {"error": type(exc).__name__, "timed_out": isinstance(exc, TimeoutError)}
+            result = {
+                "status": "could_not_build",
+                "reason": task_storage.exception_reason(exc, private=(str(self.repo), self.repo.name)),
+                "phase": "scratch guard",
+                "exit_code": None,
+                "stderr": "",
+                "stopped": True,
+            }
+        except (ValueError, OSError, KeyError, TypeError, TimeoutError, subprocess.SubprocessError) as exc:
+            result = {
+                "error": type(exc).__name__,
+                "reason": task_storage.exception_reason(exc, private=(str(self.repo), self.repo.name)),
+                "phase": self.phase,
+                "exit_code": None,
+                "stderr": "",
+                "timed_out": isinstance(exc, TimeoutError),
+            }
         result["wall_seconds"] = time.monotonic() - started
         result = _bounded({"seq": command.seq, **result}, min(MAX_RESULT_BYTES, command.limits.get("result_bytes", MAX_RESULT_BYTES)))
         self.seq, self.last_digest, self.last_result = command.seq, digest, copy.deepcopy(result)
@@ -179,7 +194,9 @@ class InProcessExecutor:
     def _execute(self, command: Command) -> dict[str, Any]:
         name, args = command.name, command.args
         if self.task_boundary:
+            self.phase = "scratch guard"
             task_storage.check(task_storage.WORKSPACE if task_storage.WORKSPACE.is_dir() else self.repo)
+            self.phase = "run"
         if name == "prepare_verification":
             return self._prepare_verification(args)
         if name == "stop":
@@ -240,10 +257,18 @@ class InProcessExecutor:
                 wall_timeout_seconds=self._remaining(),
             )
         return {
+            **(
+                {
+                    "phase": "timeout" if result.timed_out else "run",
+                    "reason": "TimeoutError: command deadline" if result.timed_out else f"CommandFailure: exit code {result.exit_code}",
+                }
+                if result.timed_out or result.exit_code
+                else {}
+            ),
             "isolation_mode": self.isolation_mode,
             "exit_code": result.exit_code,
             "stdout": result.stdout[:16384],
-            "stderr": result.stderr[-16384:],
+            "stderr": task_storage.diagnostic(result.stderr, maximum=4000),
             "timed_out": result.timed_out,
             "truncated": len(result.stdout) > 16384 or len(result.stderr) > 16384,
         }
@@ -259,12 +284,15 @@ class InProcessExecutor:
         initial_bytes = self._tree_bytes()
         guard_root = task_storage.WORKSPACE if task_storage.WORKSPACE.is_dir() else self.repo
         peak_bytes = task_storage.check(guard_root)
+        self.phase = "spawn"
         with tempfile.TemporaryFile() as out, tempfile.TemporaryFile() as err:
             proc = subprocess.Popen(
                 argv, cwd=self.repo, env=clean_environment(), stdout=out, stderr=err, start_new_session=True, preexec_fn=constrain
             )
+            self.phase = "run"
             timed_out = False
             disk_exceeded = False
+            failure = None
             # Monitor the full disk-backed workspace, including package caches.
             # An unnamespaced child can still hardcode RAM-backed paths: this
             # guard is not a replacement for the operator's memory limit.
@@ -290,16 +318,36 @@ class InProcessExecutor:
                         break
                     except subprocess.TimeoutExpired:
                         self._remaining()
-            except TimeoutError:
+            except task_storage.ScratchLimit as exc:
+                disk_exceeded = True
+                failure = exc
+                self.stopped = True
+            except TimeoutError as exc:
                 timed_out = True
+                failure = exc
+            except OSError as exc:
+                self.phase = "scratch guard"
+                failure = exc
             finally:
                 with suppress(ProcessLookupError):
                     os.killpg(proc.pid, signal.SIGKILL)
                 proc.wait()
             out.seek(0)
             err.seek(max(0, os.fstat(err.fileno()).st_size - 16384))
+            stderr = task_storage.diagnostic(err.read().decode(errors="replace"), maximum=4000)
+            phase = "scratch guard" if disk_exceeded else "timeout" if timed_out else self.phase
+            if disk_exceeded:
+                self.stopped = True
+            failed = failure or (task_storage.ScratchLimit("scratch limit") if disk_exceeded else None)
+            reason = (
+                task_storage.exception_reason(failed)
+                if failed
+                else (f"CommandFailure: exit code {proc.returncode}" if proc.returncode else "")
+            )
             return {
-                **({"status": "could_not_build", "reason": "scratch limit"} if disk_exceeded else {}),
+                **({"status": "could_not_build", "stopped": self.stopped} if disk_exceeded else {}),
+                **({"reason": reason, "phase": phase} if reason else {}),
+                **({"error": type(failure).__name__} if failure and not (timed_out or disk_exceeded) else {}),
                 "isolation_mode": "task-boundary",
                 "exit_code": proc.returncode,
                 "timed_out": timed_out,
@@ -309,7 +357,7 @@ class InProcessExecutor:
                 "stdout_bytes": os.fstat(out.fileno()).st_size,
                 "stderr_bytes": os.fstat(err.fileno()).st_size,
                 "stdout": out.read(16384).decode(errors="replace"),
-                "stderr": err.read(16384).decode(errors="replace"),
+                "stderr": stderr,
                 "truncated": os.fstat(out.fileno()).st_size > 16384 or os.fstat(err.fileno()).st_size > 16384,
             }
 
@@ -345,10 +393,8 @@ class InProcessExecutor:
                 elif recipe["recipe"] == "gradle":
                     argv += ["assemble"]
                 built = self._task_run(argv, self._remaining(), {"disk_bytes": task_storage.DEFAULT_SCRATCH_BYTES})
-                if built.get("reason") == "scratch limit":
-                    raise task_storage.ScratchLimit("scratch limit")
-                if built["exit_code"] or built["timed_out"]:
-                    return {"status": "could_not_build", "reason": "build recipe failed: " + built["stderr"][-2000:]}
+                if built.get("disk_limit_exceeded") or built["exit_code"] or built["timed_out"]:
+                    return {**built, "status": "could_not_build"}
             # Validate internal generated links before materializing them as
             # regular archive members. No host/external files may be exported.
             _tree(self.repo, generated_links=True)
@@ -361,6 +407,7 @@ class InProcessExecutor:
             (bundle / "spec.json").write_text(json.dumps(spec))
             task_storage.check(task_storage.WORKSPACE if task_storage.WORKSPACE.is_dir() else bundle)
             blob = pack_checkout(bundle)
+            self.phase = "result upload"
             URLTransport().put(self.prepared_put_url, blob, self._remaining())
             return {"status": "ok", "archive_bytes": len(blob), "sha256": hashlib.sha256(blob).hexdigest()}
 
@@ -403,7 +450,7 @@ class URLTransport:
         except urllib.error.HTTPError as exc:
             if exc.code == 404:
                 return None
-            raise OSError("object read failed") from None
+            raise OSError(f"object read failed: HTTP {exc.code}") from None
 
     def put(self, url: str, data: bytes, timeout: float) -> None:
         with self.opener.open(
@@ -434,10 +481,14 @@ class ObjectStoreExecutor:
         payload = command.validate()
         end = time.monotonic() + timeout_seconds
         self.uncertain = True
+        last_error = ""
+        phase = "result upload"
         while time.monotonic() < end:
             try:
                 # A lost PUT acknowledgement is retried with the exact same sequence.
+                phase = "command upload"
                 self.transport.put(self.command_url, payload, max(0.001, end - time.monotonic()))
+                phase = "result read"
                 raw = self.transport.get(self.result_url, max(0.001, end - time.monotonic()))
                 if raw is not None:
                     if len(raw) > MAX_RESULT_BYTES:
@@ -451,10 +502,10 @@ class ObjectStoreExecutor:
                         self.seq, self.uncertain = command.seq, False
                         self.isolation_mode = result.get("isolation_mode", self.isolation_mode)
                         return result
-            except (OSError, TimeoutError):
-                pass
+            except (OSError, TimeoutError) as exc:
+                last_error = task_storage.exception_reason(exc)
             time.sleep(min(self.poll_seconds, max(0, end - time.monotonic())))
-        raise TimeoutError("executor result deadline")
+        raise TimeoutError(f"executor {phase} deadline; {last_error or 'no result received'}")
 
 
 def serve(
@@ -481,8 +532,11 @@ def serve(
                 transport.put(result_put_url, json.dumps(result).encode(), min(10, max(0.001, end - time.monotonic())))
                 if result.get("stopped"):
                     return
-        except (OSError, TimeoutError):
-            pass
+        except (OSError, TimeoutError) as exc:
+            if executor.last_result:
+                executor.last_result["result_upload_reason"] = task_storage.exception_reason(exc)
+                executor.last_result["result_upload_phase"] = "result upload"
+
         time.sleep(min(poll_seconds, max(0, end - time.monotonic())))
     raise TimeoutError("executor task deadline")
 
@@ -534,11 +588,14 @@ def main() -> None:
     prepared_url = os.environ.get("PREPARED_PUT_URL")
     # Clearing Python's environment cannot remove initial /proc/self/environ.
     # Refuse a misconfigured task before any repository input is opened.
-    if any(
-        value and any(word in key.upper() for word in ("KEY", "TOKEN", "SECRET", "PASSWORD", "CREDENTIAL"))
+    inherited = sorted(
+        key
         for key, value in os.environ.items()
-    ):
-        raise SystemExit("executor refuses inherited credentials")
+        if value and any(word in key.upper() for word in ("KEY", "TOKEN", "SECRET", "PASSWORD", "CREDENTIAL"))
+    )
+    if inherited:
+        # Names only, never values, so a misconfigured task can be diagnosed from the worker log.
+        raise SystemExit("executor refuses inherited credentials: " + ", ".join(inherited))
     scratch_limit = task_storage.configure()
     os.environ.clear()
     os.environ.update(clean_environment())

src/openultrasast/search/probe.py

@@ -22,6 +22,7 @@ from typing import Any
 
 from . import _sandbox
 from .oracles import BrowserExecutor
+from .task_storage import exception_reason
 from .verify import Side, verify
 
 FAMILIES = ("sql", "path", "command", "ssrf", "xss")
@@ -128,13 +129,13 @@ def probe(root: Path) -> dict[str, Any]:
                     if side.reason.startswith("verification unavailable or refused:"):
                         raise _sandbox.IsolationUnavailable(f"{family}/{name}: {side.reason}")
     except Exception as exc:
-        record.update(status="instrument_failure", exception=f"{type(exc).__name__}: {exc}")
+        record.update(status="instrument_failure", exception=exception_reason(exc))
         namespaces = record["namespaces"]
         if namespaces and all(namespaces[name]["exit_code"] != 0 for name in ("user", "pid", "net", "ipc", "uts", "mount")):
             record["isolation_mode"] = "task-boundary"
             record["requires_fresh_side_tasks"] = True
         if record["isolation_check"] is None:
-            record["isolation_check"] = {"status": "instrument_failure", "exception": str(exc)}
+            record["isolation_check"] = {"status": "instrument_failure", "exception": exception_reason(exc)}
     finally:
         record["wall_seconds"] = time.monotonic() - start
         record["peak_rss_kib"] = {
@@ -165,8 +166,9 @@ def main(argv: list[str] | None = None) -> int:
             with urllib.request.urlopen(request, timeout=30) as response:
                 if not 200 <= response.status < 300:
                     return 1
-        except Exception:
-            # Presigned URL is a bearer secret; never print upload exceptions.
+        except Exception as exc:
+            record.update(status="instrument_failure", phase="result upload", reason=exception_reason(exc))
+            print(json.dumps(record), flush=True)
             return 1
     return 0 if record["status"] == "ok" else 1
 

src/openultrasast/search/task_storage.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import os
+import re
 import stat
 import subprocess
 import tempfile
@@ -23,6 +24,7 @@ def environment() -> dict[str, str]:
         return {}
     root = str(WORKSPACE)
     return {
+        "HOME": root + "/home",
         "TMPDIR": root + "/tmp",
         "npm_config_cache": root + "/.npm",
         "MAVEN_OPTS": "-Dmaven.repo.local=" + root + "/.m2",
@@ -39,6 +41,7 @@ def configure() -> int:
     env = environment()
     if env:
         Path(env["TMPDIR"]).mkdir(parents=True, exist_ok=True)
+        Path(env["HOME"]).mkdir(parents=True, exist_ok=True)
         os.environ.update(env)
         # tempfile may have cached /tmp before this entry point ran.
         tempfile.tempdir = env["TMPDIR"]
@@ -53,7 +56,8 @@ def check(root: Path | None = None, limit: int | None = None, *, additional_byte
         raise OSError("scratch root unavailable")
 
     def unreadable(error: OSError) -> None:
-        raise error
+        if not isinstance(error, FileNotFoundError):
+            raise error
 
     total = additional_bytes
     if total > limit:
@@ -91,3 +95,25 @@ def communicate(proc: subprocess.Popen[Any], data: bytes, timeout: float, root:
             first = False
             if time.monotonic() >= end:
                 raise
+
+
+def diagnostic(value: object, *, maximum: int = 300, private: tuple[str, ...] = ()) -> str:
+    """Sanitize before truncation so a clipped secret cannot escape redaction."""
+    text = str(value)
+    for key, secret in os.environ.items():
+        if len(secret) >= 4 and any(word in key.upper() for word in ("KEY", "TOKEN", "SECRET", "PASSWORD", "CREDENTIAL")):
+            text = text.replace(secret, "[redacted]")
+    text = re.sub(r"(?i)bearer\s+[^\s,;\"']+", "Bearer [redacted]", text)
+    text = re.sub(r"(?i)(?:https?|s3)://[^\s<>\"']+", "[redacted-url]", text)
+    text = re.sub(
+        r"(?i)([\w-]*(?:authorization|signature|credential|token|secret|password|key))\s*[:=]\s*[^\s,;&\"']+", r"\1=[redacted]", text
+    )
+    for name in sorted(set(private), key=len, reverse=True):
+        if name:
+            text = text.replace(name, "[repository]")
+    return "\n".join(text.splitlines()[-20:])[-maximum:]
+
+
+def exception_reason(exc: BaseException, *, private: tuple[str, ...] = ()) -> str:
+    prefix = type(exc).__name__ + ": "
+    return prefix + diagnostic(exc, maximum=max(0, 300 - len(prefix)), private=private)

src/openultrasast/search/verify.py

@@ -29,6 +29,7 @@ from ..sandbox import SandboxJob, SandboxResult
 from . import _sandbox, task_storage
 from .demo import BUILD_RECIPES, RUNTIMES, capture_output, load_demo, substitute
 from .oracles import BrowserExecutor, Canary, Oracle, oracle_for
+from .task_storage import exception_reason
 
 
 @dataclass(frozen=True)
@@ -414,10 +415,10 @@ def _side(side: Side, demo: Path, family: str, timeout: int, browser: BrowserExe
                 finally:
                     oracle.close()
         return record("observed", "fresh task run completed" if task_boundary else "three fresh runs completed")
-    except task_storage.ScratchLimit:
-        return record("could_not_build", "scratch limit")
+    except task_storage.ScratchLimit as exc:
+        return record("could_not_build", exception_reason(exc))
     except (OSError, ValueError, AssertionError, KeyError, IndexError, TypeError) as exc:
-        return record("could_not_run", f"verification unavailable or refused: {exc}")
+        return record("could_not_run", "verification unavailable or refused: " + exception_reason(exc))
 
 
 def _run_task(
@@ -508,7 +509,7 @@ def verify(
     try:
         load_demo(artefact)
     except (OSError, ValueError) as exc:
-        return VerificationRecord("inconclusive", (), time.monotonic() - started, "invalid declarative demo: " + str(exc))
+        return VerificationRecord("inconclusive", (), time.monotonic() - started, "invalid declarative demo: " + exception_reason(exc))
     mode = "task-boundary" if task_dispatcher is not None else _sandbox.isolation_mode()
     if task_dispatcher is None:
         try:

src/openultrasast/search/verify_task.py

@@ -212,8 +212,8 @@ def task_main() -> None:
             demo.mkdir()
             (demo / "demo.json").write_text(json.dumps(spec["demo"]))
             result = run_side(Side(bundle / "checkout", products=bundle / "products"), demo, spec["family"], spec["timeout_seconds"])
-        except task_storage.ScratchLimit:
-            result = SideRecord("could_not_build", (), 0, "scratch limit", isolation_mode="task-boundary")
+        except task_storage.ScratchLimit as exc:
+            result = SideRecord("could_not_build", (), 0, task_storage.exception_reason(exc), isolation_mode="task-boundary")
         upload(urls["RESULT_URL"], json.dumps(asdict(result)).encode())
 
 

src/openultrasast/search/worker.py

@@ -20,6 +20,7 @@ from . import _sandbox
 from .coordinator import SearchTask
 from .demo import DEMO_SCHEMA, load_demo, validate_demo
 from .executor import ObjectStoreExecutor
+from .task_storage import exception_reason
 
 
 def _schema(name: str, description: str, properties: dict[str, Any], required: list[str]) -> dict[str, Any]:
@@ -286,7 +287,7 @@ class Worker:
                     self.evidence.add(ref)
                     result["evidence_ref"] = ref
                 except (ValueError, OSError, KeyError, TypeError) as exc:
-                    result = {"error": type(exc).__name__}
+                    result = {"error": type(exc).__name__, "reason": exception_reason(exc)}
                 messages.append({"role": "tool", "tool_call_id": call["id"], "content": json.dumps(result)})
         return {"status": "execution_failure", "failure": "tool step limit exceeded"}
 
@@ -307,18 +308,18 @@ class Worker:
                 "cost_usd": meter.spent,
                 "isolation_mode": getattr(self.executor, "isolation_mode", "brain-only"),
             }
-        except _sandbox.IsolationUnavailable:
+        except _sandbox.IsolationUnavailable as exc:
             return {
                 "status": "instrument_failure",
-                "failure": "sandbox isolation unavailable",
+                "failure": exception_reason(exc),
                 "cost_usd": meter.spent,
                 "isolation_mode": getattr(self.executor, "isolation_mode", "brain-only"),
             }
         except Exception as exc:
-            # Provider exceptions can contain keys/URLs. Never persist their text.
+            # Preserve the diagnostic after removing credentials and URLs.
             return {
                 "status": "execution_failure",
-                "failure": type(exc).__name__,
+                "failure": exception_reason(exc),
                 "cost_usd": meter.spent,
                 "isolation_mode": getattr(self.executor, "isolation_mode", "brain-only"),
             }

tests/test_ax_batch.py

@@ -274,13 +274,13 @@ def test_capacity_backpressure_retries_same_manifest_until_deadline(tmp_path, mo
     )
     assert lane(item(), tmp_path, 0)["status"] == "ok"
     assert len(applies) >= 3 and len(set(applies)) == 1
-    assert len(deletes) == 1
+    assert len(deletes) == len(applies)
     busy[0] = True
     lane.pause = lambda seconds: now.__setitem__(0, now[0] + seconds)
     deletes.clear()
     result = lane(item(), tmp_path, 1)
     assert result["status"] == "capacity_timeout"
-    assert not deletes
+    assert deletes
     calls = []
 
     def no_vm(*args):

tests/test_search_executor.py

@@ -39,7 +39,7 @@ def test_paths_and_oversize(tmp_path):
 def test_environment_allowlist():
     assert clean_environment(
         {"OPENROUTER_API_KEY": "secret", "GEMINI_API_KEY": "secret", "PATH": "/evil", "AWS_SECRET_ACCESS_KEY": "secret"}
-    ) == {"PATH": "/usr/bin:/bin", "LANG": "C.UTF-8"}
+    ) == {"PATH": "/venv/bin:/opt/java/openjdk/bin:/usr/local/bin:/usr/bin:/bin", "LANG": "C.UTF-8", "JAVA_HOME": "/opt/java/openjdk"}
 
 
 class Mailbox:
@@ -240,3 +240,79 @@ def test_sandbox_receives_exact_remaining_wall_limit(tmp_path, monkeypatch):
     InProcessExecutor(tmp_path).submit("run", {"command": ["/bin/true"]}, timeout_seconds=0.2)
     assert 0 < seen["probe"] <= 0.2
     assert 0 < seen["wall"] <= 0.2
+
+
+def test_spawn_failure_has_diagnostic(tmp_path):
+    result = InProcessExecutor(tmp_path, task_boundary=True).execute(Command(1, "run", {"command": ["/missing-executor-program"]}))
+    assert result["phase"] == "spawn"
+    assert result["exit_code"] is None
+    assert "FileNotFoundError:" in result["reason"]
+    assert "No such file" in result["reason"]
+
+
+def test_timeout_preserves_child_diagnostics(tmp_path):
+    import sys
+
+    result = InProcessExecutor(tmp_path, task_boundary=True).execute(
+        Command(
+            1,
+            "run",
+            {"command": [sys.executable, "-c", "import sys,time; print('before timeout',file=sys.stderr,flush=True); time.sleep(10)"]},
+            0.3,
+        )
+    )
+    assert result["phase"] == "timeout"
+    assert result["exit_code"] is not None
+    assert result["stderr"] == "before timeout"
+    assert result["reason"].startswith("TimeoutError:")
+
+
+def test_upload_failure_reports_phase(tmp_path, monkeypatch):
+    app = InProcessExecutor(tmp_path, task_boundary=True)
+
+    def upload(command):
+        app.phase = "result upload"
+        raise OSError("upload HTTP 403 https://host/path?X-Amz-Signature=secret")
+
+    monkeypatch.setattr(app, "_execute", upload)
+    result = app.execute(Command(1, "prepare_verification", {}))
+    assert result["phase"] == "result upload"
+    assert "HTTP 403" in result["reason"]
+    assert "secret" not in result["reason"]
+
+
+def test_diagnostics_sanitize_before_tail_and_keep_twenty_lines(monkeypatch):
+    from openultrasast.search.task_storage import diagnostic, exception_reason
+
+    monkeypatch.setenv("SERVICE_TOKEN", "environment-secret")
+    detail = (
+        "https://host/repo?X-Amz-Signature=url-secret Authorization: Bearer bearer-secret X-Amz-Signature=bare-secret environment-secret"
+    )
+    result = exception_reason(ValueError(detail))
+    assert result.startswith("ValueError:") and len(result) <= 300
+    for secret in ("url-secret", "bearer-secret", "bare-secret", "environment-secret", "host/repo"):
+        assert secret not in result
+    assert len(diagnostic("\n".join(str(i) for i in range(40)), maximum=4000).splitlines()) == 20
+    assert len(exception_reason(ValueError("x" * 1000))) <= 300
+
+
+def test_guard_io_failure_after_child_exit_is_not_success(tmp_path, monkeypatch):
+    import sys
+
+    from openultrasast.search import task_storage
+
+    checks = []
+    original = task_storage.check
+
+    def check(*args, **kwargs):
+        checks.append(1)
+        if len(checks) >= 3:
+            raise PermissionError("cache cannot be read")
+        return original(*args, **kwargs)
+
+    monkeypatch.setattr(task_storage, "check", check)
+    result = InProcessExecutor(tmp_path, task_boundary=True).execute(Command(1, "run", {"command": [sys.executable, "-c", "pass"]}))
+    assert result["error"] == "PermissionError"
+    assert result["phase"] == "scratch guard"
+    assert result["exit_code"] is not None
+    assert "cache cannot be read" in result["reason"]

tests/test_search_executor_smoke.py

@@ -142,9 +142,9 @@ def test_docker_managed_task_uses_docker_and_no_heap_flag(tmp_path, monkeypatch)
     )
     record = smoke.run_smoke(lane, "https://github.com/example/project", "a" * 40)
     assert record["status"] == "ok"
-    assert [a[1] for a in calls] == ["run", "rm"]
-    assert "--heap-profile" not in calls[0]
-    assert calls[1][2] == "-f"
+    assert [a[1] for a in calls] == ["image", "run", "rm"]
+    assert "--heap-profile" not in calls[1]
+    assert calls[2][2] == "-f"
 
 
 def test_apply_exception_still_deletes(tmp_path, monkeypatch):
@@ -160,3 +160,104 @@ def test_apply_exception_still_deletes(tmp_path, monkeypatch):
     assert result["status"] == "instrument_failure"
     assert events == ["apply", "delete"]
     assert "secret.example" not in json.dumps(result)
+
+
+def test_smoke_retains_sanitized_command_failure(tmp_path, monkeypatch):
+    executor, lane, events = setup(tmp_path, monkeypatch)
+    original = executor._task_run
+
+    def fail(argv, timeout, limits):
+        if "pip" in argv:
+            return {
+                "exit_code": 17,
+                "stderr": "project failed: Bearer hidden https://host/project?q=secret",
+                "stdout": "",
+                "phase": "run",
+                "reason": "CommandFailure: exit code 17",
+            }
+        return original(argv, timeout, limits)
+
+    monkeypatch.setattr(executor, "_task_run", fail)
+    record = smoke.run_smoke(lane, "https://github.com/example/project", "a" * 40)
+    row = record["commands"][-1]
+    assert row["exit_code"] == 17 and row["phase"] == "run"
+    assert "CommandFailure" in row["reason"]
+    assert all(word not in json.dumps(record) for word in ("hidden", "secret", "project", "example"))
+    assert events == ["apply", "delete"]
+
+
+def test_docker_missing_image_is_diagnosable_without_pull(tmp_path, monkeypatch):
+    import subprocess
+
+    monkeypatch.setenv("S3_ENDPOINT", "https://files.example")
+    calls = []
+
+    def runner(argv, **kwargs):
+        calls.append(argv)
+        return subprocess.CompletedProcess(argv, 1, "", "No such image: pinned-image")
+
+    lane = batch.DockerLane(
+        SimpleNamespace(atespace="default"),
+        batch.Workload("image@sha256:" + "a" * 64, frozenset(), 900, None),
+        store=Store(Executor(tmp_path, task_boundary=True)),
+        runner=runner,
+    )
+    record = smoke.run_smoke(lane, "https://github.com/example/project", "a" * 40)
+    assert record["status"] == "instrument_failure"
+    assert "ValueError:" in record["reason"] and "No such image" in record["reason"]
+    assert [c[1] for c in calls] == ["image"]
+
+
+@pytest.mark.parametrize("delete_fails", [False, True])
+def test_capacity_partial_create_deleted_before_retry(tmp_path, monkeypatch, delete_fails):
+    executor, lane, events = setup(tmp_path, monkeypatch)
+    live, attempts = set(), []
+
+    def ax(*args, **kwargs):
+        events.append(args[0])
+        if args[0] == "apply":
+            assert not live
+            live.add(kwargs["manifest"]["metadata"]["name"])
+            attempts.append(1)
+            if len(attempts) == 1:
+                raise batch.NoFreeWorkers("partially provisioned")
+        elif args[0] == "delete":
+            if delete_fails:
+                raise OSError("cannot delete")
+            live.discard(args[2])
+
+    lane.ax, lane.pause = ax, lambda seconds: None
+    record = smoke.run_smoke(lane, "https://github.com/example/project", "a" * 40)
+    assert len(attempts) == (1 if delete_fails else 2)
+    assert record["status"] == ("instrument_failure" if delete_fails else "ok")
+    if not delete_fails:
+        assert events == ["apply", "delete", "apply", "delete"] and not live
+    else:
+        # Simulated leaked task deliberately retains its reservation; restore test global.
+        batch.AX_SLOTS.release()
+
+
+def test_apply_timeout_does_not_resubmit(tmp_path, monkeypatch):
+    import subprocess
+
+    executor, lane, events = setup(tmp_path, monkeypatch)
+
+    def ax(*args, **kwargs):
+        events.append(args[0])
+        if args[0] == "apply":
+            raise subprocess.TimeoutExpired("apply", 30)
+
+    lane.ax = ax
+    record = smoke.run_smoke(lane, "https://github.com/example/project", "a" * 40)
+    assert events == ["apply", "delete"]
+    assert "TimeoutExpired:" in record["reason"]
+
+
+def test_dispatch_rejects_duplicate_items_before_launch(tmp_path):
+    item = batch.Item("same", {}, "result.json", (), {}, "")
+
+    def unexpected(*args):
+        pytest.fail("duplicate inputs must be rejected before any dispatch")
+
+    with pytest.raises(ValueError, match="duplicate dispatch item"):
+        batch.dispatch([item, item], batch.Workload("image", frozenset(), 1, None), tmp_path, {"ax": unexpected}, lanes=("ax",))

tests/test_search_task_storage.py

@@ -61,7 +61,7 @@ def test_executor_checks_caches_and_stops_at_scratch_limit(workspace, monkeypatc
         "p=Path(os.environ['PIP_CACHE_DIR']); p.mkdir(); (p/'large').write_bytes(b'x'*30000); " + ("time.sleep(10)" if wait else "")
     )
     result = app.submit("run", {"command": [sys.executable, "-c", code]}, timeout_seconds=3)
-    assert result["status"] == "could_not_build" and result["reason"] == "scratch limit"
+    assert result["status"] == "could_not_build" and result["reason"] == "ScratchLimit: scratch limit"
     assert result["stopped"] and app.stopped
 
 
@@ -126,7 +126,7 @@ def test_verifier_reports_scratch_limit(workspace, monkeypatch):
     (demo / "demo.json").write_text(json.dumps(spec()["demo"]))
     monkeypatch.setenv("OUSAST_SCRATCH_BYTES", "25000")
     result = verify_side_task(Side(repo), demo, "path")
-    assert result.outcome == "could_not_build" and result.reason == "scratch limit", result
+    assert result.outcome == "could_not_build" and result.reason == "ScratchLimit: scratch limit", result
 
 
 def test_preparation_builds_before_export_and_exports_products(workspace, monkeypatch):
@@ -165,9 +165,60 @@ def test_open_logs_count_toward_scratch_guard(workspace, monkeypatch):
     monkeypatch.setenv("OUSAST_SCRATCH_BYTES", "20000")
     app = executor.InProcessExecutor(repo, task_boundary=True)
     result = app.submit("run", {"command": [sys.executable, "-c", "print('x'*30000)"]})
-    assert result["status"] == "could_not_build" and result["reason"] == "scratch limit"
+    assert result["status"] == "could_not_build" and result["reason"] == "ScratchLimit: scratch limit"
 
 
 def test_unreadable_scratch_is_not_reported_as_empty(tmp_path):
     with pytest.raises(OSError, match="scratch root unavailable"):
         task_storage.check(tmp_path / "missing")
+
+
+def test_pip_removing_enumerated_directory_is_not_guard_failure(workspace, monkeypatch):
+    original = task_storage.os.walk
+
+    def disappearing(root, **kwargs):
+        kwargs["onerror"](FileNotFoundError(2, "removed pip build directory", str(root / "pip-build")))
+        yield from original(root, **kwargs)
+
+    monkeypatch.setattr(task_storage.os, "walk", disappearing)
+    (workspace / "input").write_bytes(b"input")
+    assert task_storage.check(workspace) >= 5
+
+
+def test_guard_still_refuses_permission_errors(workspace, monkeypatch):
+    def unreadable(root, **kwargs):
+        kwargs["onerror"](PermissionError("permission denied"))
+        return iter(())
+
+    monkeypatch.setattr(task_storage.os, "walk", unreadable)
+    with pytest.raises(PermissionError):
+        task_storage.check(workspace)
+
+
+def test_child_environment_has_disk_home_and_owned_runtime_path(workspace):
+    repo = workspace / "checkout"
+    repo.mkdir()
+    (repo / "input").write_bytes(b"proof")
+    code = (
+        "import os; from pathlib import Path; print(Path('input').stat().st_size); "
+        "print(os.environ['PATH']); p=Path.home()/'probe'; p.write_text('ok'); print(p)"
+    )
+    result = executor.InProcessExecutor(repo, task_boundary=True).submit("run", {"command": [sys.executable, "-c", code]})
+    assert result["exit_code"] == 0
+    assert result["stdout"].startswith("5\n/venv/bin:")
+    assert (workspace / "home/probe").read_text() == "ok"
+
+
+def test_scratch_failure_preserves_exit_and_stderr(workspace, monkeypatch):
+    repo = workspace / "checkout"
+    repo.mkdir()
+    monkeypatch.setenv("OUSAST_SCRATCH_BYTES", "20000")
+    code = (
+        "import sys,time; from pathlib import Path; print('pip build detail',file=sys.stderr,flush=True); "
+        "Path('big').write_bytes(b'x'*30000); time.sleep(10)"
+    )
+    result = executor.InProcessExecutor(repo, task_boundary=True).submit("run", {"command": [sys.executable, "-c", code]})
+    assert result["phase"] == "scratch guard"
+    assert result["exit_code"] is not None
+    assert result["stderr"] == "pip build detail"
+    assert result["reason"].startswith("ScratchLimit:")

tests/test_search_worker.py

@@ -113,7 +113,7 @@ def test_reservation_precedes_each_call_and_refuses_over_budget(tmp_path):
     job = task()
     job.limits["max_call_usd"] = 0.000001
     result = worker(job)
-    assert result["failure"] == "BudgetExhausted"
+    assert result["failure"].startswith("BudgetExhausted: ")
     assert len(client.calls) == 1