merge(search-with-proof): named exhaustion limits and partial evidence [skip ci]

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

benchmarks/search/pilot_run.py

@@ -88,6 +88,9 @@ class Pilot:
             fixed_effect_observed=None,
             outcome="no_evidence",
             end_reason="exhausted",
+            end_detail=None,
+            task_metrics=[],
+            executor_task_metrics=[],
             tasks_submitted=0,
             executor_tasks=0,
             verify_tasks=0,
@@ -139,9 +142,35 @@ class Pilot:
 
     @contextmanager
     def executor(self, revision, deadline=900):
+        before = time.monotonic()
+        metrics = dict(self.worker.metrics)
+        self.executor_end = "ok"
+        submitted = self.record["executor_tasks"]
+        try:
+            with self._executor(revision, deadline) as task:
+                yield task
+        except BudgetExhausted as exc:
+            self.executor_end = exc.end_detail
+            raise
+        except Exception:
+            self.executor_end = "execution_failure"
+            raise
+        finally:
+            if self.record["executor_tasks"] > submitted:
+                self.record["executor_task_metrics"].append(
+                    {
+                        "executor_task": submitted + 1,
+                        **{key: value - metrics[key] for key, value in self.worker.metrics.items()},
+                        "wall_seconds": time.monotonic() - before,
+                        "end": self.executor_end,
+                    }
+                )
+
+    @contextmanager
+    def _executor(self, revision, deadline=900):
         deadline = min(deadline, self.deadline - time.monotonic())
         if deadline < 1:
-            raise BudgetExhausted("task deadline", usd=0, calls=0)
+            raise BudgetExhausted("task deadline", usd=0, calls=0, end_detail="task_wall")
         acquired_at = time.monotonic()
         empty = io.BytesIO()
         with tarfile.open(fileobj=empty, mode="w"):
@@ -151,7 +180,7 @@ class Pilot:
         with SearchExecutorTask(self.lane, empty.getvalue(), deadline=deadline) as task:
             remaining = deadline - (time.monotonic() - acquired_at)
             if remaining < 1:
-                raise BudgetExhausted("acquisition deadline", usd=0, calls=0)
+                raise BudgetExhausted("acquisition deadline", usd=0, calls=0, end_detail="task_wall")
             timeout = min(180, remaining)
             result = task.client.submit(
                 "run",
@@ -163,7 +192,7 @@ class Pilot:
                 raise RepositoryFailure("checkout failed")
             remaining = deadline - (time.monotonic() - acquired_at)
             if remaining < 1:
-                raise BudgetExhausted("acquisition deadline", usd=0, calls=0)
+                raise BudgetExhausted("acquisition deadline", usd=0, calls=0, end_detail="task_wall")
             result = task.client.submit("read_file", {"path": self.row["file"]}, timeout_seconds=min(30, remaining))
             if not isinstance(result.get("input_bytes"), int) or result["input_bytes"] <= 0:
                 self.failure("acquisition", "candidate_input_unreadable")
@@ -173,7 +202,7 @@ class Pilot:
 
     def admit_task(self):
         if self.record["tasks_submitted"] >= 32:
-            raise BudgetExhausted("executor task cap", usd=0, calls=0)
+            raise BudgetExhausted("executor task cap", usd=0, calls=0, end_detail="tasks")
         self.record["tasks_submitted"] += 1
 
     def prepare(self, revision, spec):
@@ -247,6 +276,9 @@ class Pilot:
         self.phase = task.step
         before = time.monotonic()
         self.deadline = before + task.limits.get("task_wall_seconds", 900)
+        metrics_before = dict(self.worker.metrics)
+        result = {}
+        end = "execution_failure"
         try:
             if task.step == "verify":
                 result = self.verification()
@@ -267,6 +299,7 @@ class Pilot:
                         self.failure("verify/" + side.outcome, "side_verification_failed")
                 if result.outcome != "demonstrated":
                     self.failure("verify", "no_consistent_differential" if result.outcome == "inconclusive" else result.outcome)
+                end = result.outcome
                 return dict(status="ok", outcome=result.outcome, evidence_refs=["verify:owned"], cost_usd=0)
             if task.step == "explore":
                 if self.args.dry_run:
@@ -277,20 +310,32 @@ class Pilot:
                         self.worker.executor = remote.client
                         remaining = task.limits["task_wall_seconds"] - (time.monotonic() - before)
                         if remaining < 1:
-                            raise BudgetExhausted("explore deadline", usd=0, calls=0)
+                            raise BudgetExhausted("explore deadline", usd=0, calls=0, end_detail="task_wall")
                         result = self.worker(replace(task, limits={**task.limits, "task_wall_seconds": remaining}))
+                        self.executor_end = result.get("end_detail") or result.get("status", "execution_failure")
             else:
                 result = self.worker(task)
-            if self.worker.budget_exhausted:
-                raise BudgetExhausted("run ceiling", usd=0, calls=0)
+            end = result.get("end_detail") or result.get("status", "execution_failure")
             if result.get("status") != "ok":
                 self.failure(task.step, result.get("failure", "worker_execution_failed"))
-                if self.worker.metrics["model_calls"] == 0:
+                if self.worker.metrics["model_calls"] == 0 and not result.get("budget_exhaustion"):
                     raise RuntimeError(self.record["failures"][-1]["reason"])
                 if not self.verifications:
                     self.record["outcome"] = "inconclusive"
             return result
+        except BudgetExhausted as exc:
+            end = exc.end_detail
+            raise
         finally:
+            self.record["task_metrics"].append(
+                {
+                    "task_id": task.id,
+                    "step": task.step,
+                    **{key: value - metrics_before[key] for key, value in self.worker.metrics.items()},
+                    "wall_seconds": time.monotonic() - before,
+                    "end": end,
+                }
+            )
             wall = self.record["wall_seconds_by_phase"]
             wall[task.step] = wall.get(task.step, 0) + time.monotonic() - before
 
@@ -300,6 +345,7 @@ class Pilot:
         try:
             state = coordinator.run()
             self.record["end_reason"] = state["end_reason"]
+            self.record["end_detail"] = state.get("end_detail")
             if not self.verifications and any(f.get("step") == "explore" for f in state["facts"]):
                 self.record["outcome"] = "plausible_unproven"
         except Exception as exc:

src/openultrasast/plane/budget.py

@@ -23,8 +23,9 @@ _USAGE_FIELDS = ("prompt_tokens", "prompt_cache_hit_tokens", "completion_tokens"
 class BudgetExhausted(RuntimeError):
     """Raised before a call would start once the usd or calls ceiling is reached."""
 
-    def __init__(self, message: str, *, usd: float | None, calls: int) -> None:
-        super().__init__(message)
+    def __init__(self, message: str, *, usd: float | None, calls: int, end_detail: str = "spend") -> None:
+        super().__init__(f"{end_detail}: {message}")
+        self.end_detail = end_detail
         self.usd = usd
         self.calls = calls
 

src/openultrasast/search/board.py

@@ -27,7 +27,7 @@ class Board:
         self._lock = threading.RLock()
         self.head: str | None = None
         self._state: dict[str, Any] = dict(
-            search_id=search_id, version=0, previous=None, facts=[], intents=[], hints=[], end_reason=None, checkpoint={}
+            search_id=search_id, version=0, previous=None, facts=[], intents=[], hints=[], end_reason=None, end_detail=None, checkpoint={}
         )
 
     @property
@@ -54,7 +54,7 @@ class Board:
                 raise PermissionError("only the coordinator may write the board")
             if self._state["end_reason"] is not None:
                 raise ValueError("search already ended")
-            if changes.keys() - {"facts", "intents", "hints", "end_reason", "checkpoint"}:
+            if changes.keys() - {"facts", "intents", "hints", "end_reason", "end_detail", "checkpoint"}:
                 raise ValueError("unknown board fields")
             state = {**self.state, **copy.deepcopy(changes), "version": self._state["version"] + 1, "previous": self.head}
             self._validate(state)

src/openultrasast/search/coordinator.py

@@ -67,6 +67,7 @@ class Coordinator:
         self.board, self.executor = board, executor
         self.budget, self.clock, self.push_budget = budget or SearchBudget(), clock, push_budget
         self.state = board.state
+        self.state.setdefault("end_detail", None)
         self.progress = self.state["checkpoint"] or dict(
             started_at=clock(),
             spent_usd=0.0,
@@ -84,10 +85,25 @@ class Coordinator:
         self.progress["wall_seconds"] = max(0, self.clock() - self.progress["started_at"])
         self.state["checkpoint"] = self.progress
         self.board.commit(
-            writer=self.board.coordinator, **{k: self.state[k] for k in ("facts", "intents", "hints", "end_reason", "checkpoint")}
+            writer=self.board.coordinator,
+            **{k: self.state[k] for k in ("facts", "intents", "hints", "end_reason", "end_detail", "checkpoint")},
         )
 
-    def _end(self, reason: str) -> dict[str, Any]:
+    def _end(self, reason: str, end_detail: str | None = None) -> dict[str, Any]:
+        if reason == "budget_spent" and end_detail not in {
+            "spend",
+            "task_spend",
+            "reason_rounds",
+            "demonstrations",
+            "tasks",
+            "task_wall",
+            "search_wall",
+            "worker_steps",
+            "worker_wall",
+            "run_ceiling",
+        }:
+            raise ValueError("budget_spent requires an exhausted limit")
+        self.state["end_detail"] = end_detail
         self.state["end_reason"] = reason
         self._save()
         return self.board.state
@@ -110,16 +126,21 @@ class Coordinator:
         while True:
             remaining = self._remaining()
             if remaining <= self.budget.cleanup_seconds or self.progress["tasks"] >= self.budget.tasks:
-                raise BudgetExhausted("task or wall budget exhausted", usd=self.meter.spent, calls=self.progress["tasks"])
+                raise BudgetExhausted(
+                    "task or wall budget exhausted",
+                    usd=self.meter.spent,
+                    calls=self.progress["tasks"],
+                    end_detail="search_wall" if remaining <= self.budget.cleanup_seconds else "tasks",
+                )
             maximum = self.budget.max_call_usd if step != "verify" else 0
             reservation = self.meter.reserve(maximum)
             push_reservation = None
             try:
                 if self.push_budget is not None:
                     push_reservation = self.push_budget.reserve(maximum)
-            except BudgetExhausted:
+            except BudgetExhausted as exc:
                 reservation.release()
-                raise
+                raise BudgetExhausted("push ceiling", usd=exc.usd, calls=exc.calls, end_detail="run_ceiling") from exc
             task_id = "task-" + str(self.progress["tasks"] + 1)
             limits = asdict(self.budget)
             limits["task_wall_seconds"] = min(self.budget.task_wall_seconds, remaining - self.budget.cleanup_seconds)
@@ -164,7 +185,16 @@ class Coordinator:
                 or self._remaining() <= 0
                 or self.clock() - dispatched_at > limits["task_wall_seconds"]
             ):
-                raise BudgetExhausted("executor exceeded cost or wall contract", usd=self.meter.spent, calls=self.progress["tasks"])
+                raise BudgetExhausted(
+                    "executor exceeded cost or wall contract",
+                    usd=self.meter.spent,
+                    calls=self.progress["tasks"],
+                    end_detail="task_spend"
+                    if result.get("status") == "invalid_cost"
+                    else "search_wall"
+                    if self._remaining() <= 0
+                    else "task_wall",
+                )
             result = {**result, "task_id": task_id}
             # Persist results together with their effects in run(), not a "done"
             # checkpoint which could lose the result if the host then crashes.
@@ -209,13 +239,13 @@ class Coordinator:
             self._save()
         while True:
             if self._remaining() <= self.budget.cleanup_seconds:
-                return self._end("budget_spent")
+                return self._end("budget_spent", "search_wall")
             try:
                 intent = next((i for i in self.state["intents"] if i["status"] == "open"), None)
                 demo = next((f for f in self.state["facts"] if f.get("demo") and f["id"] not in self.progress["verified"]), None)
                 if demo is not None:
                     if self.progress["demonstrations"] >= self.budget.demonstrations:
-                        return self._end("budget_spent")
+                        return self._end("budget_spent", "demonstrations")
                     self.progress["demonstrations"] += 1
                     result = self._call("verify", {"fact_id": demo["id"], "demo": demo["demo"]})
                     self.progress["verified"].append(demo["id"])
@@ -259,7 +289,7 @@ class Coordinator:
                         intent["failure"] = "invalid_result" if result.get("status") == "ok" else result.get("status", "execution_failure")
                 else:
                     if self.progress["rounds"] >= self.budget.reason_rounds:
-                        return self._end("budget_spent")
+                        return self._end("budget_spent", "reason_rounds")
                     self.progress["rounds"] += 1
                     result = self._call("reason", {})
                     if result.get("status") == "ok" and not result.get("intents") and "complete" not in result:
@@ -275,6 +305,8 @@ class Coordinator:
                         if result.get("status") == "ok" and not result.get("intents") and "complete" not in result:
                             self.progress["failures"].append({"task_id": result["task_id"], "reason": "reason_refused"})
                             return self._end("reason_refused")
+                    if result.get("budget_exhaustion"):
+                        return self._end("budget_spent", result["budget_exhaustion"]["end_detail"])
                     proposals = result.get("intents", [])
                     if result.get("status") != "ok" or not isinstance(proposals, list) or len(proposals) > self.budget.intents_per_round:
                         return self._end("exhausted")
@@ -304,9 +336,11 @@ class Coordinator:
                         )
                     if not added:
                         return self._end("exhausted")
+                if result.get("budget_exhaustion"):
+                    return self._end("budget_spent", result["budget_exhaustion"]["end_detail"])
                 self._save()
-            except BudgetExhausted:
-                return self._end("budget_spent")
+            except BudgetExhausted as exc:
+                return self._end("budget_spent", exc.end_detail)
 
     @staticmethod
     def _refs(refs: Any) -> bool:

src/openultrasast/search/executor.py

@@ -314,7 +314,7 @@ class InProcessExecutor:
             self.phase = "run"
             timed_out = False
             disk_exceeded = False
-            failure = None
+            failure: Exception | None = 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.

src/openultrasast/search/worker.py

@@ -173,6 +173,8 @@ class Worker:
         self.run_budget = run_budget
         self.metrics = dict(reserved=0.0, settled=0.0, model_calls=0, tokens_in=0, tokens_out=0)
         self.budget_exhausted = False
+        self.task_calls_start = 0
+        self.observations: list[tuple[str, str]] = []
         selected_prices = prices if prices is not None else getattr(client, "prices", None)
         if not isinstance(selected_prices, Prices) or any(not math.isfinite(v) or v < 0 for v in vars(selected_prices).values()):
             raise ValueError("nonnegative finite model prices required before admission")
@@ -192,26 +194,36 @@ class Worker:
     def _remaining(self) -> int:
         remaining = min(60, self.deadline - time.monotonic())
         if remaining < 1:
-            raise BudgetExhausted("worker wall budget spent", usd=0, calls=0)
+            raise BudgetExhausted("worker wall budget spent", usd=0, calls=0, end_detail="worker_wall")
         return int(remaining)
 
     def _call(self, messages: list[dict[str, Any]], tools: list[dict[str, Any]], meter: SpendBudget) -> dict[str, Any]:
         timeout = self._remaining()
-        # UTF-8 bytes upper-bound content tokens; include conservative per-message
-        # chat framing. A provider with extra hidden billed tokens needs new pricing.
-        input_bound = len(json.dumps([messages, tools], ensure_ascii=False).encode()) + 4096 * (len(messages) + 1)
+        # Pilot measured ~27k input tokens against the old ~100k reservation (3-4x).
+        # Use bytes/2 plus 1024 framing tokens per message; retain max_tokens for output.
+        # This calibrated bound is checked at settlement, not a tokenizer guarantee.
+        input_bound = math.ceil(len(json.dumps([messages, tools], ensure_ascii=False).encode()) / 2) + 1024 * (len(messages) + 1)
         maximum = (
             input_bound * max(self.prices.input_per_m, self.prices.cache_hit_per_m) + self.max_output_tokens * self.prices.output_per_m
         ) / 1_000_000
+        try:
             reservation = meter.reserve(maximum)
+        except BudgetExhausted as exc:
+            raise BudgetExhausted(
+                "next call reservation exceeds task allowance",
+                usd=exc.usd,
+                calls=int(self.metrics["model_calls"]) - self.task_calls_start,
+                end_detail="task_spend",
+            ) from exc
         run_reservation = None
         try:
             if self.run_budget is not None:
                 run_reservation = self.run_budget.reserve(maximum)
-        except BudgetExhausted:
+        except BudgetExhausted as exc:
             reservation.release()
-            self.budget_exhausted = True
-            raise
+            raise BudgetExhausted(
+                "run ceiling", usd=exc.usd, calls=int(self.metrics["model_calls"]) - self.task_calls_start, end_detail="run_ceiling"
+            ) from exc
         self.metrics["reserved"] += maximum
         self.metrics["model_calls"] += 1
         actual = maximum
@@ -326,6 +338,7 @@ class Worker:
             {"role": "user", "content": task.snapshot + "\nIntent: " + json.dumps(task.payload)},
         ]
         self.evidence = set()
+        self.observations = []
         for index in range(self.max_steps):
             messages[0]["content"] = (
                 EXPLORE_PROMPT
@@ -369,13 +382,20 @@ class Worker:
                     ref = "tool:" + str(len(self.evidence) + 1)
                     self.evidence.add(ref)
                     result["evidence_ref"] = ref
+                    if not result.get("error"):
+                        text = json.dumps(result, ensure_ascii=False)
+                        if len(text) > 2000:
+                            text = text[:2000] + " [truncated]"
+                        self.observations.append((ref, name + ": " + text))
                 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)})
-        return {"status": "execution_failure", "failure": "tool step limit exceeded"}
+        raise BudgetExhausted("tool step limit exceeded", usd=meter.spent, calls=self.max_steps, end_detail="worker_steps")
 
     def __call__(self, task: SearchTask) -> dict[str, Any]:
         self.budget_exhausted = False
+        self.task_calls_start = int(self.metrics["model_calls"])
+        self.observations = []
         self.limits = task.limits
         self.deadline = time.monotonic() + task.limits.get("task_wall_seconds", 900)
         meter = SpendBudget(task.limits["max_call_usd"])
@@ -392,6 +412,27 @@ class Worker:
                 "cost_usd": meter.spent,
                 "isolation_mode": getattr(self.executor, "isolation_mode", "brain-only"),
             }
+        except BudgetExhausted as exc:
+            self.budget_exhausted = True
+            result = {"status": "execution_failure", "failure": exception_reason(exc)}
+            if task.step == "explore" and self.observations:
+                # Preserve literal executor observations, without inventing a security conclusion
+                # or making another paid call to summarize them. Bound board growth explicitly.
+                observations = self.observations[-6:]
+                result = {
+                    "status": "ok",
+                    "fact": {
+                        "text": "Partial tool observations (search incomplete):\n" + "\n".join(text for _, text in observations),
+                        "evidence_refs": [ref for ref, _ in observations],
+                    },
+                }
+            return {
+                **result,
+                "cost_usd": meter.spent,
+                "end_detail": exc.end_detail,
+                "budget_exhaustion": {"end_detail": exc.end_detail, "usd": exc.usd, "calls": exc.calls},
+                "isolation_mode": getattr(self.executor, "isolation_mode", "brain-only"),
+            }
         except _sandbox.IsolationUnavailable as exc:
             return {
                 "status": "instrument_failure",

tests/test_search_coordinator.py

@@ -164,3 +164,56 @@ def test_invalid_identifiers_refused_before_paths(tmp_path):
             Board(FileStore(tmp_path), bad, coordinator="host")
         with pytest.raises(ValueError):
             SearchTask(bad, "search-1", "reason", "", {}, asdict(SearchBudget()))
+
+
+@pytest.mark.parametrize("detail", ["task_spend", "worker_steps", "worker_wall", "run_ceiling"])
+def test_worker_limit_preserves_fact_and_survives_resume(tmp_path, detail):
+    def execute(task):
+        if task.step == "reason":
+            return {"status": "ok", "cost_usd": 0, "intents": [{"description": "inspect"}]}
+        return {
+            "status": "ok",
+            "cost_usd": 0.01,
+            "budget_exhaustion": {"end_detail": detail, "usd": 0.01, "calls": 1},
+            "fact": {"text": "Confirmed partial observation", "evidence_refs": ["tool:1"]},
+        }
+
+    b = board(tmp_path)
+    state = Coordinator(b, execute).run()
+    assert state["end_reason"] == "budget_spent" and state["end_detail"] == detail
+    assert state["facts"][0]["text"] == "Confirmed partial observation"
+    assert state["intents"][0]["status"] == "concluded"
+    assert state["checkpoint"]["spent_usd"] == 0.01
+    assert Board.resume(b.store, "search-1", b.head, coordinator="host").state == state
+
+
+@pytest.mark.parametrize("detail", ["spend", "tasks", "reason_rounds", "demonstrations", "task_wall", "search_wall", "run_ceiling"])
+def test_coordinator_identifies_each_own_limit(tmp_path, detail):
+    from openultrasast.plane.budget import SpendBudget
+
+    now = [0]
+
+    def execute(task):
+        if detail in {"task_wall", "search_wall"}:
+            now[0] = 1000 if detail == "task_wall" else 20000
+        if task.step == "reason":
+            return {"status": "ok", "cost_usd": 0, "intents": [{"description": str(task.id)}]}
+        if task.step == "explore":
+            return {"status": "ok", "cost_usd": 0, "fact": {"text": "candidate", "evidence_refs": ["tool:1"], "demo": "demo/"}}
+        return {"status": "ok", "cost_usd": 0, "outcome": "inconclusive"}
+
+    limits = {
+        "spend": {"spend_usd": 0},
+        "tasks": {"tasks": 1},
+        "reason_rounds": {"reason_rounds": 1},
+        "demonstrations": {"demonstrations": 1},
+    }
+    state = Coordinator(
+        board(tmp_path),
+        execute,
+        budget=SearchBudget(**limits.get(detail, {})),
+        clock=lambda: now[0],
+        push_budget=SpendBudget(0) if detail == "run_ceiling" else None,
+    ).run()
+    assert state["end_reason"] == "budget_spent"
+    assert state["end_detail"] == detail

tests/test_search_pilot_run.py

@@ -63,6 +63,8 @@ def test_ceiling_refuses_before_stub_call(tmp_path):
     record = json.loads(output.read_text().splitlines()[0])
     assert record["model_calls"] == 0
     assert record["end_reason"] == "budget_spent"
+    assert record["end_detail"] == "run_ceiling"
+    assert record["task_metrics"][0]["end"] == "run_ceiling"
 
 
 def make_pilot(tmp_path, *, side="fixed", dry_run=False):
@@ -221,6 +223,13 @@ def test_live_wiring_with_fake_mailbox_no_network(tmp_path, monkeypatch, verify_
     assert record["end_reason"] == "goal_met"
     assert record["model_calls"] == 3
     assert record["executor_tasks"] == 3
+    assert len(record["executor_task_metrics"]) == 3
+    assert sum(t["model_calls"] for t in record["executor_task_metrics"]) == 3
+    for row in record["executor_task_metrics"]:
+        assert row["reserved"] >= row["settled"]
+        assert row["wall_seconds"] > 0
+        assert row["end"] == "ok"
+        assert {"tokens_in", "tokens_out"} <= row.keys()
     assert record["verify_tasks"] == (6 if verify_lane == "ax" else 0)
     assert record["tasks_submitted"] == (9 if verify_lane == "ax" else 3)
     assert record["input_bytes"] > 0

tests/test_search_worker.py

@@ -307,3 +307,82 @@ def test_prompt_contracts_and_demo_example(tmp_path, step):
         assert len(prompt) / 4 < 1000  # rough offline token estimate, not a tokenizer
     else:
         assert "file:line" in prompt and "final call for finish" in prompt
+
+
+@pytest.mark.parametrize("with_evidence", [False, True])
+def test_task_spend_finishes_with_confirmed_observations(tmp_path, with_evidence):
+    from openultrasast.model.endpoint import Prices
+
+    (tmp_path / "input.txt").write_text("confirmed input\n" * 400)
+
+    class UsageStub(StubModel):
+        def complete_chat_raw(self, **kwargs):
+            raw = super().complete_chat_raw(**kwargs)
+            raw["usage"] = {"prompt_tokens": 100, "completion_tokens": 100}
+            return raw
+
+    model = UsageStub([tool("read_file", path="input.txt")])
+    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
+    result = worker(job)
+    assert result["end_detail"] == result["budget_exhaustion"]["end_detail"] == "task_spend"
+    assert len(model.calls) == int(with_evidence)
+    assert worker.metrics["reserved"] >= worker.metrics["settled"]
+    if with_evidence:
+        assert result["status"] == "ok"
+        assert "confirmed input" in result["fact"]["text"]
+        assert result["fact"]["evidence_refs"] == ["tool:1"]
+        assert result["cost_usd"] == pytest.approx(0.0002)
+    else:
+        assert result["status"] == "execution_failure"
+        assert "task_spend" in result["failure"]
+        assert "fact" not in result
+
+
+def test_calibrated_reservation_covers_reported_usage(tmp_path):
+    import math
+
+    from openultrasast.model.endpoint import Prices
+    from openultrasast.plane.budget import SpendBudget
+
+    class UsageStub(StubModel):
+        def complete_chat_raw(self, **kwargs):
+            raw = super().complete_chat_raw(**kwargs)
+            raw["usage"] = {"prompt_tokens": 27000, "completion_tokens": 1024}
+            return raw
+
+    model = UsageStub([reply({})])
+    worker = Worker(model, repo=tmp_path, demo=tmp_path / "demo", prices=Prices(1, 1, 2))
+    worker.reason("facts: []\nnote: " + "x" * 85000, SpendBudget(1))
+    call = model.calls[0]
+    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"] >= worker.metrics["settled"] == pytest.approx(0.029048)
+
+
+@pytest.mark.parametrize("detail", ["worker_steps", "worker_wall", "run_ceiling"])
+def test_worker_exhaustion_is_structured(tmp_path, monkeypatch, detail):
+    from openultrasast.model.endpoint import Prices
+    from openultrasast.plane.budget import SpendBudget
+
+    worker = Worker(
+        StubModel([tool("list_files")]),
+        repo=tmp_path,
+        demo=tmp_path / "demo",
+        executor=InProcessExecutor(tmp_path),
+        max_steps=1,
+        prices=Prices(1, 1, 1),
+        run_budget=SpendBudget(0 if detail == "run_ceiling" else 1),
+    )
+    if detail == "worker_wall":
+        from openultrasast.search import worker as module
+
+        times = iter([0, 901])
+        monkeypatch.setattr(module.time, "monotonic", lambda: next(times))
+    result = worker(task("explore"))
+    assert result["budget_exhaustion"]["end_detail"] == detail
+    assert result["end_detail"] == detail