Skip to content

pyhooks Python package

pyhooks

A Python library that lets Vivaria agents interact with Vivaria. pyhooks also contains other code shared between METR agents.

Actions

Functions that agents can use to implement actions, e.g. running bash and Python commands.

Source code in pyhooks/pyhooks/__init__.py
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
class Actions:
    """
    Functions that agents can use to implement actions, e.g. running bash and Python commands.
    """

    def __init__(self, envs: CommonEnvs | None = None):
        self.envs = envs or CommonEnvs.from_env()

    async def run_bash(self, script: str, timeout: float) -> str:
        await self.check_safety(script)
        return await run_bash(script, timeout)

    async def run_python(self, script: str, timeout: float) -> str:
        return await run_python(script, timeout)

    async def check_safety(self, action: str):
        safety_policy_notice = (
            await trpc_server_request(
                "mutation",
                "checkActionSafety",
                {
                    "runId": self.envs.run_id,
                    "agentBranchNumber": self.envs.branch,
                    "action": action,
                },
                envs=self.envs,
            )
        )["notice"]

        if safety_policy_notice:
            raise ActionViolatesSafetyPolicyException(safety_policy_notice)

Hooks

Bases: BaseModel

Source code in pyhooks/pyhooks/__init__.py
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
class Hooks(BaseModel):
    class Config:
        arbitrary_types_allowed = True

    def __init__(
        self,
        task_id: str | None = None,
        envs: CommonEnvs | None = None,
    ):
        super().__init__()
        self._task_id = task_id or env.TASK_ID
        self._envs = envs or CommonEnvs.from_env()

    @property
    def task_id(self) -> str:
        if not self._task_id:
            raise Exception("TASK_ID not set")
        return self._task_id

    def _send_background_request(
        self,
        reqtype: str,
        route: str,
        data: dict,
        session: aiohttp.ClientSession | None = None,
    ):
        try:
            # Try to get the currently running event loop
            loop = asyncio.get_running_loop()
            # If successful, create a task in the running loop
            return loop.create_task(
                self._send_trpc_server_request(reqtype, route, data, session)
            )
        except RuntimeError:
            # No event loop is running, so we create a new one and run the task
            loop = asyncio.new_event_loop()
            asyncio.set_event_loop(loop)

            async def coro():
                return await self._send_trpc_server_request(
                    reqtype,
                    route,
                    data,
                    session,
                )

            task = loop.run_until_complete(coro())
            loop.close()
            return task

    async def _send_trpc_server_request(
        self,
        reqtype: str,
        route: str,
        data: dict,
        session: aiohttp.ClientSession | None = None,
        record_pause_on_error: bool = True,
    ) -> Any:
        return await trpc_server_request(
            reqtype,
            route,
            data,
            session=session,
            record_pause_on_error=record_pause_on_error,
            envs=self._envs,
        )

    def main(self, main_function: Callable):
        async def error_handler_wrapper():
            try:
                import pdb_attach

                pdb_attach.listen(50000)
            except Exception as e:
                print("Failed to start pdb attach", repr(e))
            nonlocal main_function
            exit_code = 0
            try:
                await main_function(self)
            except SystemExit as e:
                if e.code is not None:
                    exit_code = e.code
            except Exception as e:
                if env.TESTING:
                    print("fatal error:", e, file=sys.stderr)
                exit_code = 1
                await self._send_trpc_server_request(
                    "mutation",
                    "logFatalError",
                    self.make_trace_entry(
                        {
                            "detail": str(e),
                            "from": "agent",
                            "trace": traceback.format_exc(),
                            "extra": None,
                        }
                    ),
                )
            finally:
                current_task = asyncio.current_task()
                all_tasks = [x for x in asyncio.all_tasks() if x is not current_task]
                all_tasks = await asyncio.gather(*all_tasks)
                return exit_code

        exit_code = asyncio.run(error_handler_wrapper())
        exit(exit_code)

    def make_trace_entry(self, x: dict[str, Any]) -> dict[str, Any]:
        result = self._new_base_event() | {"content": x}
        return result

    # Don't wait for log, action, observation, frameStart, or frameEnd. Instead, run them in the background

    def log(self, *content: Any):
        return self.log_with_attributes(None, *content)

    def log_with_attributes(self, attributes: dict | None, *content: Any):
        entry = self.make_trace_entry({"content": content, "attributes": attributes})
        return self._send_background_request("mutation", "log", entry)

    def log_image(self, image_url: str, description: str | None = None):
        entry = self.make_trace_entry(
            {"content": [{"image_url": image_url, "description": description}]}
        )
        return self._send_background_request("mutation", "log", entry)

    def action(self, action: dict):
        entry = self.make_trace_entry({"action": action})
        return self._send_background_request("mutation", "action", entry)

    def observation(self, observation: dict):
        entry = self.make_trace_entry({"observation": observation})
        return self._send_background_request("mutation", "observation", entry)

    async def log_error(self, detail: Any, extra: Any = None):
        # don't cause another error just because error failed (would be messy)
        entry = self.make_trace_entry(
            {
                "detail": str(detail),
                "from": "agent",
                "trace": "".join(traceback.format_stack()[:-2]),
                "extra": extra,
            }
        )
        await self._send_trpc_server_request("mutation", "logError", entry)

    def start_frame(self, name: str):
        req = self.make_trace_entry({"name": name})
        return self._send_background_request("mutation", "frameStart", req)

    def end_frame(self):
        req = self.make_trace_entry({})
        return self._send_background_request("mutation", "frameEnd", req)

    def save_state(self, state: Any):
        req = self.make_trace_entry({"state": state})
        return self._send_background_request("mutation", "saveState", req)

    def frame(self, name: str):
        def decorator(func):
            @functools.wraps(func)
            async def wrapper(*args, **kwargs):
                self.start_frame(name)
                result = await func(*args, **kwargs)
                self.end_frame()
                return result

            return wrapper

        return decorator

    # do wait for submit, generate
    async def getTask(self) -> TaskInfo:
        res = await self._send_trpc_server_request(
            "query",
            "getTaskInstructions",
            {
                "taskId": self.task_id,
                "runId": self._envs.run_id,
                "agentBranchNumber": self._envs.branch,
            },
        )
        return TaskInfo(**res)

    async def submit(self, submission: str):
        if not isinstance(submission, str):
            raise TypeError(f"submission must be a string, got {type(submission)}")

        async with aiohttp.ClientSession(
            # No timeout because scoring the submission can take a long time
            timeout=aiohttp.ClientTimeout(),
        ) as session:
            await self._send_trpc_server_request(
                "mutation",
                "submit",
                self.make_trace_entry({"value": submission}),
                session=session,
            )

        exit(0)

    async def score(self) -> ScoreResult:
        async with aiohttp.ClientSession(
            # No timeout because scoring the task environment can take a long time
            timeout=aiohttp.ClientTimeout(),
        ) as session:
            res = await self._send_trpc_server_request(
                "mutation",
                "score",
                {"runId": self._envs.run_id, "agentBranchNumber": self._envs.branch},
                session=session,
            )
            return ScoreResult(**res)

    async def scoreLog(self) -> list[ScoreLogEntry]:
        async with aiohttp.ClientSession(
            # No timeout because scoring the task environment can take a long time
            timeout=aiohttp.ClientTimeout(),
        ) as session:
            res = await self._send_trpc_server_request(
                "query",
                "getScoreLog",
                {"runId": self._envs.run_id, "agentBranchNumber": self._envs.branch},
                session=session,
            )
            return [ScoreLogEntry(**x) for x in res]

    async def generate(
        self,
        settings: MiddlemanSettings,
        template: str | None = None,
        templateValues: dict[str, Any] | None = None,
        prompt: str | None = None,
        messages: list[OpenaiChatMessage] | None = None,
        description: Optional[str] = None,
        functions: Optional[Any] = None,
        extraParameters: dict[str, Any] | None = None,
    ) -> MiddlemanResult:
        genReq = GenerationRequest(
            settings=settings,
            template=template,
            templateValues=templateValues,
            messages=messages,
            description=description,
            functions=functions,
            prompt=prompt,
            extraParameters=extraParameters,
        )
        req = self._new_base_event() | {"genRequest": genReq.model_dump()}
        return MiddlemanResult(
            **(
                await self._send_trpc_server_request(
                    "mutation",
                    "generate",
                    req,
                )
            )
        )

    async def count_prompt_tokens(
        self,
        settings: MiddlemanSettings,
        messages: list[OpenaiChatMessage],
        functions: Optional[Any] = None,
        extraParameters: dict[str, Any] | None = None,
    ) -> int:
        """Returns the number of prompt tokens that a generation request will use."""
        genReq = GenerationRequest(
            settings=settings,
            messages=messages,
            functions=functions,
            extraParameters=extraParameters,
        )
        req = {"genRequest": genReq.model_dump()}
        res = await self._send_trpc_server_request("mutation", "countPromptTokens", req)
        return res["tokens"]

    async def burn_tokens(
        self,
        n_prompt_tokens: int,
        n_completion_tokens: int,
        n_serial_action_tokens: int | None = None,
    ):
        req = self._new_base_event() | {
            "n_prompt_tokens": n_prompt_tokens,
            "n_completion_tokens": n_completion_tokens,
            "n_serial_action_tokens": n_serial_action_tokens,
        }
        await self._send_trpc_server_request(
            "mutation",
            "burnTokens",
            req,
        )

    async def generate_one(
        self,
        settings: MiddlemanSettings,
        template: str | None = None,
        templateValues: dict[str, Any] | None = None,
        prompt: str | None = None,
        messages: list[OpenaiChatMessage] | None = None,
        description: Optional[str] = None,
        extraParameters: dict[str, Any] | None = None,
    ) -> str:
        if settings.n != 1:
            raise Exception(
                "in generate_one, n must be 1. use generate for n>1 and full middleman output"
            )
        result = await self.generate(
            settings=settings,
            template=template,
            templateValues=templateValues,
            messages=messages,
            description=description,
            prompt=prompt,
            extraParameters=extraParameters,
        )
        if result.error is not None or result.outputs is None:
            raise Exception("Generation error", result.error)
        return result.outputs[0].completion

    async def generate_many(
        self,
        settings: MiddlemanSettings,
        template: str | None = None,
        templateValues: dict[str, Any] | None = None,
        prompt: str | None = None,
        messages: list[OpenaiChatMessage] | None = None,
        description: Optional[str] = None,
        extraParameters: dict[str, Any] | None = None,
    ) -> list[str]:
        result = await self.generate(
            settings=settings,
            template=template,
            templateValues=templateValues,
            messages=messages,
            description=description,
            prompt=prompt,
            extraParameters=extraParameters,
        )
        if result.error is not None or result.outputs is None:
            raise Exception("Generation error", result.error)
        return [x.completion for x in result.outputs]

    async def rate_options(
        self,
        rating_model: str,
        rating_template: str,
        transcript: str,
        options: list[RatingOption],
        description: Optional[str] = None,
    ) -> RatedOption:
        trace_entry = self.make_trace_entry(
            {
                "options": [x.dict() for x in options],
                "description": description,
                "ratingModel": (rating_model),
                "ratingTemplate": rating_template,
                "transcript": transcript,
            }
        )
        chosen_option = await self._send_trpc_server_request(
            "mutation",
            "rateOptions",
            trace_entry,
        )
        entry_key = {
            "runId": trace_entry["runId"],
            "index": trace_entry["index"],
            "agentBranchNumber": trace_entry["agentBranchNumber"],
        }
        while chosen_option is None:
            print("Waiting for human interaction")
            chosen_option = await self._send_trpc_server_request(
                "query",
                "retrieveRatings",
                entry_key,
            )
        return RatedOption(**chosen_option)

    async def embed(self, req):
        return await self._send_trpc_server_request("mutation", "embeddings", req)

    def get_tokenizer(self, tokenizer_name: str = "cl100k_base"):
        try:
            return tiktoken.get_encoding(tokenizer_name)
        except Exception:
            return tiktoken.get_encoding("cl100k_base")

    async def get_input(self, description: str, default_input: str) -> str:
        "get input from user or use default if not in intervention mode"
        trace_entry = self.make_trace_entry(
            {
                "description": description,
                "defaultInput": default_input,
            }
        )
        entry_key = {
            "runId": trace_entry["runId"],
            "index": trace_entry["index"],
            "agentBranchNumber": trace_entry["agentBranchNumber"],
        }
        await self._send_trpc_server_request("mutation", "requestInput", trace_entry)
        input = await self._send_trpc_server_request(
            "query", "retrieveInput", entry_key
        )
        while input is None:
            print("Waiting for human interaction")
            input = await self._send_trpc_server_request(
                "query", "retrieveInput", entry_key
            )
            if input is None:
                await asyncio.sleep(10)
        return input

    def token_lengths(
        self, texts: list[str], tokenizer_or_model_name: str = "cl100k_base"
    ) -> list[int]:
        if "gpt-4" in tokenizer_or_model_name or "turbo" in tokenizer_or_model_name:
            tokenizer_or_model_name = "cl100k_base"
        try:
            tokenizer = self.get_tokenizer(tokenizer_or_model_name)
        except Exception as e:
            print("can't find tokenizer", tokenizer_or_model_name, repr(e))
            tokenizer = self.get_tokenizer("cl100k_base")
        return [len(x) for x in tokenizer.encode_batch(texts, disallowed_special=())]

    def token_length(self, text, tokenizer_or_model_name: str = "cl100k_base") -> int:
        return self.token_lengths([text], tokenizer_or_model_name)[0]

    def oai_message_token_lengths(self, messages: list[OpenaiChatMessage]) -> list[int]:
        return [
            x + 3
            for x in self.token_lengths(
                [
                    # TODO Handle the case where x.content is a list[dict], as it can be for
                    # gpt-4-vision-preview: https://platform.openai.com/docs/guides/vision/quick-start
                    (x.content if isinstance(x.content, str) else "")
                    + (json.dumps(x.function_call) if x.function_call else "")
                    + (x.name if x.name else "")
                    for x in messages
                ],
                "cl100k_base",
            )
        ]

    async def get_permitted_models_info(self) -> dict[str, ModelInfo]:
        global permitted_models_cache
        if permitted_models_cache:
            return permitted_models_cache
        res = await self._send_trpc_server_request(
            "query",
            "getPermittedModelsInfo",
            {},
        )
        permitted_models_info = {mi["name"]: ModelInfo(**mi) for mi in res}
        permitted_models_cache = permitted_models_info
        return permitted_models_info

    # Deprecated; use Actions#run_bash instead
    async def run_bash(self, script, timeout) -> str:
        await Actions().check_safety(script)
        return await run_bash(script, timeout)

    # Deprecated; use Actions#run_python instead
    async def run_python(self, script, timeout) -> str:
        return await run_python(script, timeout)

    def deduplicate_options(self, options: list[RatingOption]) -> list[RatingOption]:
        return deduplicate_options(options)

    async def update_agent_command_result(
        self,
        stdout_to_append: str,
        stderr_to_append: str,
        exit_status: int | None,
        agent_pid: int | None,
    ):
        req = {
            "runId": self._envs.run_id,
            "agentBranchNumber": self._envs.branch,
            "stdoutToAppend": stdout_to_append,
            "stderrToAppend": stderr_to_append,
            "exitStatus": exit_status,
            "agentPid": agent_pid,
        }
        await self._send_trpc_server_request(
            "mutation",
            "updateAgentCommandResult",
            req,
        )

    async def get_usage(self) -> RunUsageAndLimits:
        res = await self._send_trpc_server_request(
            "query",
            "getRunUsageHooks",
            {
                "runId": self._envs.run_id,
                "agentBranchNumber": self._envs.branch,
            },
        )
        return RunUsageAndLimits(**res)

    async def pause(self):
        await self._send_trpc_server_request(
            "mutation",
            "pause",
            {
                "runId": self._envs.run_id,
                "agentBranchNumber": self._envs.branch,
                "start": timestamp_now(),
                "reason": "pauseHook",
            },
            record_pause_on_error=False,
        )

    async def unpause(self):
        await self._send_trpc_server_request(
            "mutation",
            "unpause",
            {
                "runId": self._envs.run_id,
                "agentBranchNumber": self._envs.branch,
                "reason": "unpauseHook",
            },
            record_pause_on_error=False,
        )

    def _new_base_event(self) -> dict[str, Any]:
        return {
            "runId": self._envs.run_id,
            "index": random_index(),
            "agentBranchNumber": self._envs.branch,
            "calledAt": timestamp_strictly_increasing(),
        }

count_prompt_tokens(settings, messages, functions=None, extraParameters=None) async

Returns the number of prompt tokens that a generation request will use.

Source code in pyhooks/pyhooks/__init__.py
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
async def count_prompt_tokens(
    self,
    settings: MiddlemanSettings,
    messages: list[OpenaiChatMessage],
    functions: Optional[Any] = None,
    extraParameters: dict[str, Any] | None = None,
) -> int:
    """Returns the number of prompt tokens that a generation request will use."""
    genReq = GenerationRequest(
        settings=settings,
        messages=messages,
        functions=functions,
        extraParameters=extraParameters,
    )
    req = {"genRequest": genReq.model_dump()}
    res = await self._send_trpc_server_request("mutation", "countPromptTokens", req)
    return res["tokens"]

get_input(description, default_input) async

get input from user or use default if not in intervention mode

Source code in pyhooks/pyhooks/__init__.py
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
async def get_input(self, description: str, default_input: str) -> str:
    "get input from user or use default if not in intervention mode"
    trace_entry = self.make_trace_entry(
        {
            "description": description,
            "defaultInput": default_input,
        }
    )
    entry_key = {
        "runId": trace_entry["runId"],
        "index": trace_entry["index"],
        "agentBranchNumber": trace_entry["agentBranchNumber"],
    }
    await self._send_trpc_server_request("mutation", "requestInput", trace_entry)
    input = await self._send_trpc_server_request(
        "query", "retrieveInput", entry_key
    )
    while input is None:
        print("Waiting for human interaction")
        input = await self._send_trpc_server_request(
            "query", "retrieveInput", entry_key
        )
        if input is None:
            await asyncio.sleep(10)
    return input

Pauser

Manages delays in retrying RPCs, and sending pause/unpause requests to the server

Source code in pyhooks/pyhooks/__init__.py
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
class Pauser:
    """Manages delays in retrying RPCs, and sending pause/unpause requests to the server"""

    _envs: CommonEnvs
    _start: int
    _end: Optional[int]
    _state: State
    _sleeper: Sleeper
    _request_fn: RequestFn
    _record_pause: bool

    class State(Enum):
        NO_PAUSE = auto()
        PAUSE_REQUESTED = auto()
        PAUSE_FAILED = auto()
        PAUSE_SUCCEEDED = auto()

    def __init__(
        self,
        envs: CommonEnvs,
        sleeper: Sleeper,
        request_fn: RequestFn,
        record_pause: bool,
    ):
        self._envs = envs
        self._start = timestamp_now()
        self._end = None
        self._state = self.State.NO_PAUSE
        self._sleeper = sleeper
        self._request_fn = request_fn
        self._record_pause = record_pause

    @property
    def run_id(self) -> int:
        return cast(int, self._envs.run_id or env.RUN_ID)

    @property
    def branch(self) -> int:
        return cast(int, self._envs.branch or env.AGENT_BRANCH_NUMBER)

    async def pause(self):
        await self._try_pause_once()
        await self._sleeper.sleep()
        self._end = timestamp_now()

    async def _try_pause_once(self):
        """Tries to ensure that a single pause request was sent to the server.

        Can be called successively and will only retry pausing until success."""
        match self._state:
            case self.State.NO_PAUSE:
                self._state = self.State.PAUSE_REQUESTED
                await self._send_pause()
            case self.State.PAUSE_FAILED:
                await self._send_pause()
            case self.State.PAUSE_REQUESTED, self.State.PAUSE_SUCCEEDED:
                return

    async def _send_pause(self) -> bool:
        if not self._record_pause:
            self._state = self.State.PAUSE_SUCCEEDED
            return True
        try:
            await self._request_fn(
                "mutation",
                "pause",
                {
                    "runId": self.run_id,
                    "agentBranchNumber": self.branch,
                    "reason": "pyhooksRetry",
                    "start": self._start,
                },
                record_pause_on_error=False,
                envs=self._envs,
            )
            self._state = self.State.PAUSE_SUCCEEDED
            return True
        except Exception as e:
            self._state = self.State.PAUSE_FAILED
            print("Failed to pause trpc server request", repr(e))
            return False

    async def unpause(self):
        """Sends an unpause request to the server if necessary.

        Also sends a pause request if previous pause attempts failed."""
        match self._state:
            case self.State.NO_PAUSE:
                return
            case self.State.PAUSE_REQUESTED:
                raise RuntimeError(
                    "Unpause called before pause completed (should never happen)"
                )
            case self.State.PAUSE_FAILED:
                if await self._send_pause():
                    await self._send_unpause()
                # If the pause request failed, an unpause will just make things confusing.
            case self.State.PAUSE_SUCCEEDED:
                await self._send_unpause()

    async def _send_unpause(self):
        assert self._end is not None
        if not self._record_pause:
            return
        try:
            await self._request_fn(
                "mutation",
                "unpause",
                {
                    "runId": self.run_id,
                    "agentBranchNumber": self.branch,
                    "reason": "pyhooksRetry",
                    "end": self._end,
                },
                record_pause_on_error=False,
                envs=self._envs,
            )
        except Exception as e:
            print("Failed to unpause trpc server request", repr(e))
            raise

unpause() async

Sends an unpause request to the server if necessary.

Also sends a pause request if previous pause attempts failed.

Source code in pyhooks/pyhooks/__init__.py
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
async def unpause(self):
    """Sends an unpause request to the server if necessary.

    Also sends a pause request if previous pause attempts failed."""
    match self._state:
        case self.State.NO_PAUSE:
            return
        case self.State.PAUSE_REQUESTED:
            raise RuntimeError(
                "Unpause called before pause completed (should never happen)"
            )
        case self.State.PAUSE_FAILED:
            if await self._send_pause():
                await self._send_unpause()
            # If the pause request failed, an unpause will just make things confusing.
        case self.State.PAUSE_SUCCEEDED:
            await self._send_unpause()

pyhooks.agent_output

Functionality for watching a single /agent-output/agent-branch-N directory for changes to the agent's stdout, stderr, and exit status files. When a file changes, watch_agent_output calls Hooks#update_agent_command_result with the updated stdout, stderr, and exit status. Each agent branch in an agent container starts its own copy of this script.

pyhooks.env

pyhooks.execs

exports python_exec utility for handling timeouts and output redirection (one day, RAM too)

run_python(code, timeout_sec, wait_after_kill=15.0, minimum_free_ram_bytes=800000000, log=False) async

exec the code in a separate thread and collect the stdout and stderr. If the code takes longer than timeout seconds to run, then we attempt to kill it. If it takes longer than wait_after_kill seconds to die, then we give up and return the result anyway. (The thread will die when its running system call finishes executing.)

Variables are shared between threads, so e.g. shared_box[0] += 1 works. Note that x += 1 won't work.

Source code in pyhooks/pyhooks/execs.py
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
async def run_python(
    code: str,
    timeout_sec: float,
    wait_after_kill=15.0,
    minimum_free_ram_bytes=800_000_000,
    log=False,
) -> str:
    """
    exec the code in a separate thread and collect the stdout and stderr.
    If the code takes longer than timeout seconds to run, then we attempt to kill it.
    If it takes longer than wait_after_kill seconds to die, then we give up and return the result anyway.
    (The thread will die when its running system call finishes executing.)

    Variables are shared between threads, so e.g. `shared_box[0] += 1` works. Note that `x += 1` won't work.
    """
    from pyhooks import Actions  # type: ignore

    await Actions().check_safety(code)

    async with aiohttp.ClientSession(
        timeout=aiohttp.ClientTimeout(
            total=60 * 25, sock_connect=60 * 25, sock_read=60 * 25
        )
    ) as session:
        try:
            async with session.post(
                "http://localhost:9712/run_python",
                json={
                    "code": code,
                    "timeout": timeout_sec,
                    "wait_after_kill": wait_after_kill,
                    "minimum_free_ram_bytes": minimum_free_ram_bytes,
                    "log": log,
                },
            ) as resp:
                # encode and decode to prevent errors from unicode surrogate characters
                return (
                    (await resp.json())["result"]
                    .encode("utf-8", errors="replace")
                    .decode("utf-8")
                )
        except Exception as e:
            print(f"Error connecting to python server: {e}", file=sys.stderr)
            return f"Unknown error. May be caused by python code timeout after 25 minutes. Details: {e}"

pyhooks.options

pyhooks.python_server

InterruptibleThread

Bases: Thread

A thread that can be interrupted with t.raiseException() Based on https://stackoverflow.com/a/325528

Source code in pyhooks/pyhooks/python_server.py
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
class InterruptibleThread(threading.Thread):
    """
    A thread that can be interrupted with t.raiseException()
    Based on https://stackoverflow.com/a/325528
    """

    def run(self):
        """
        Catch uncaught exceptions and save them to t.exc.
        Necessary to remove unwanted "Exception ignored in thread started by..." and "Exception ignored in sys.unraisablehook..."
        https://stackoverflow.com/a/31614591
        """
        self.exc = None
        try:
            self.ret = self._target(*self._args, **self._kwargs)  # type: ignore
        except Exception as e:
            self.exc = e

    def raiseException(self, ExceptionClass):
        """
        Interrupt thread with an exception.
        Exception happens after the current system call finishes executing.
        (So eg time.sleep() is not interrupted.)
        If exception isn't firing then you can try calling this in a loop.
        """
        if not self.is_alive():
            return  # do nothing
        thread_id = self.ident
        if thread_id is None:
            raise Exception("couldn't get thread identifier")
        res = ctypes.pythonapi.PyThreadState_SetAsyncExc(
            ctypes.c_long(thread_id), ctypes.py_object(ExceptionClass)
        )
        if res == 0:
            raise ValueError("invalid thread id")
        elif res != 1:
            # "if it returns a number greater than one, you're in trouble,
            # and you should call it again with exc=NULL to revert the effect"
            ctypes.pythonapi.PyThreadState_SetAsyncExc(ctypes.c_long(thread_id), None)
            raise SystemError("PyThreadState_SetAsyncExc failed")

raiseException(ExceptionClass)

Interrupt thread with an exception. Exception happens after the current system call finishes executing. (So eg time.sleep() is not interrupted.) If exception isn't firing then you can try calling this in a loop.

Source code in pyhooks/pyhooks/python_server.py
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
def raiseException(self, ExceptionClass):
    """
    Interrupt thread with an exception.
    Exception happens after the current system call finishes executing.
    (So eg time.sleep() is not interrupted.)
    If exception isn't firing then you can try calling this in a loop.
    """
    if not self.is_alive():
        return  # do nothing
    thread_id = self.ident
    if thread_id is None:
        raise Exception("couldn't get thread identifier")
    res = ctypes.pythonapi.PyThreadState_SetAsyncExc(
        ctypes.c_long(thread_id), ctypes.py_object(ExceptionClass)
    )
    if res == 0:
        raise ValueError("invalid thread id")
    elif res != 1:
        # "if it returns a number greater than one, you're in trouble,
        # and you should call it again with exc=NULL to revert the effect"
        ctypes.pythonapi.PyThreadState_SetAsyncExc(ctypes.c_long(thread_id), None)
        raise SystemError("PyThreadState_SetAsyncExc failed")

run()

Catch uncaught exceptions and save them to t.exc. Necessary to remove unwanted "Exception ignored in thread started by..." and "Exception ignored in sys.unraisablehook..." https://stackoverflow.com/a/31614591

Source code in pyhooks/pyhooks/python_server.py
132
133
134
135
136
137
138
139
140
141
142
def run(self):
    """
    Catch uncaught exceptions and save them to t.exc.
    Necessary to remove unwanted "Exception ignored in thread started by..." and "Exception ignored in sys.unraisablehook..."
    https://stackoverflow.com/a/31614591
    """
    self.exc = None
    try:
        self.ret = self._target(*self._args, **self._kwargs)  # type: ignore
    except Exception as e:
        self.exc = e

OutputTee

Allows each thread to output to different File objects (with optional prefix). Based on https://stackoverflow.com/a/57996986

Source code in pyhooks/pyhooks/python_server.py
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
class OutputTee:
    """
    Allows each thread to output to different File objects (with optional prefix).
    Based on https://stackoverflow.com/a/57996986
    """

    def __init__(self, default_file):
        self.which_files = {}
        self.default = default_file

    def set_outputs(self, files):
        self.which_files[get_thread_id()] = files

    def write(self, message: str):
        files = self.which_files.get(get_thread_id(), [self.default])
        for file in files:
            try:
                file.write(message)
            except:  # noqa: E722
                pass

    def flush(self):
        "required for compatibility"
        files = self.which_files.get(get_thread_id(), [self.default])
        for file in files:
            try:
                file.flush()
            except:  # noqa: E722
                pass

flush()

required for compatibility

Source code in pyhooks/pyhooks/python_server.py
104
105
106
107
108
109
110
111
def flush(self):
    "required for compatibility"
    files = self.which_files.get(get_thread_id(), [self.default])
    for file in files:
        try:
            file.flush()
        except:  # noqa: E722
            pass

PrefixedFile

add a prefix to each line written to a file

Source code in pyhooks/pyhooks/python_server.py
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
class PrefixedFile:
    "add a prefix to each line written to a file"

    def __init__(self, file, prefix):
        self.file = file
        self.prefix = prefix
        self.on_newline = True

    def write(self, s: str):
        if not s:
            return
        if self.on_newline:
            s = self.prefix + s
        ends_with_newline = s[-1] == "\n"
        if ends_with_newline:
            s = s[:-1]
        s = s.replace("\n", f"\n{self.prefix}")
        self.file.write(s)
        if ends_with_newline:
            self.file.write("\n")
        self.on_newline = ends_with_newline

    def flush(self):
        self.file.flush()

worker(code, output_file, timeout_fyi, log)

Redirects outputs and performs exec

Source code in pyhooks/pyhooks/python_server.py
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
def worker(code: str, output_file, timeout_fyi: float, log: bool):
    "Redirects outputs and performs exec"

    # set up output redirection
    c = worker_counter_box[0]
    worker_counter_box[0] += 1
    stdouts: list = [output_file]
    stderrs: list = [output_file]
    if log:
        stdouts.append(PrefixedFile(real_stdout, f"[python-exec-{c}]-  "))
        stderrs.append(PrefixedFile(real_stderr, f"[python-exec-{c}]+  "))
    stdout.set_outputs(stdouts)
    stderr.set_outputs(stderrs)

    # do the exec
    try:
        ipython_shell.run_cell(code)
    except PythonExecTimeoutException:
        print(
            f"PythonExecTimeoutException: python exec timed out after {timeout_fyi} seconds",
            file=stderr,
        )
    except PythonExecOutOfMemoryException:
        print(
            "PythonExecOutOfMemoryException: python exec exceeded available memory. Python environment has been reset.",
            file=stderr,
        )
    except Exception as e:
        traceback.print_exception(type(e), e, e.__traceback__, file=stderr)

pyhooks.types

pyhooks.util

get_available_ram_bytes(base_path=_MEMORY_CGROUP_DIR)

docker-specific! normal stuff like psutil won't work

Source code in pyhooks/pyhooks/util.py
29
30
31
32
33
34
35
36
37
def get_available_ram_bytes(base_path: pathlib.Path = _MEMORY_CGROUP_DIR) -> float:
    "docker-specific! normal stuff like psutil won't work"
    current_path = (base_path / "memory.current")
    if not current_path.exists():
        # system is using cgroup v1
        current_path = (base_path / "memory/memory.usage_in_bytes")

    current = current_path.read_text()
    return _get_ram_limit_bytes(base_path) - int(current)