adopt-workflow 0.3.0__py3-none-any.whl

adopt_workflow/__init__.py

@@ -0,0 +1,63 @@
+"""The Workflow facade, the in-process test backend, the purity checker.
+
+Contracts §10.2 · implementation spec §4.14 · PRD F14.
+
+Invariants this package carries: **no DBOS symbol appears here** -- the only
+importer in the programme is `plane_workflow.dbos_backend`, in the closed
+repository -- workflow bodies are pure, checked at decoration time by
+`assert_pure` and across the tree by the `workflow-body-purity` contract, and
+**no Build 0 OSS-side command uses durable workflows**, so the OSS CLI never
+requires Postgres.
+"""
+
+from adopt_workflow.api import (
+    TERMINAL_STATUSES,
+    Backoff,
+    RetryPolicy,
+    StepContext,
+    WorkflowClient,
+    WorkflowContext,
+    WorkflowHandle,
+    WorkflowStatus,
+    backoff_delays_ms,
+    validate_idempotency_key,
+)
+from adopt_workflow.decorators import (
+    REGISTRY,
+    ScheduledDefinition,
+    StepDefinition,
+    WorkflowDefinition,
+    clear_registry,
+    resolve,
+    scheduled,
+    step,
+    workflow,
+)
+from adopt_workflow.inproc import InProcessWorkflowClient, Journal
+from adopt_workflow.purity import assert_pure, find_impure_workflow_bodies
+
+__all__ = [
+    "REGISTRY",
+    "TERMINAL_STATUSES",
+    "Backoff",
+    "InProcessWorkflowClient",
+    "Journal",
+    "RetryPolicy",
+    "ScheduledDefinition",
+    "StepContext",
+    "StepDefinition",
+    "WorkflowClient",
+    "WorkflowContext",
+    "WorkflowDefinition",
+    "WorkflowHandle",
+    "WorkflowStatus",
+    "assert_pure",
+    "backoff_delays_ms",
+    "clear_registry",
+    "find_impure_workflow_bodies",
+    "resolve",
+    "scheduled",
+    "step",
+    "validate_idempotency_key",
+    "workflow",
+]

adopt_workflow/api.py

@@ -0,0 +1,262 @@
+"""The `Workflow` facade: contracts §10.2, and nothing a backend can widen.
+
+**Why a facade at all.** Implementation spec §1.1 locks DBOS for durable
+execution and source spec §13 requires that the documented Temporal migration
+stay a migration. That is only true if no DBOS symbol appears outside one module,
+which means every caller in the programme talks to the shapes declared here.
+`no-dbos` enforces the second half; this file is the first.
+
+**What the seam owns and what a backend owns.** The seam owns the vocabulary --
+status values, the retry policy and its caps, the handle, the two contexts. A
+backend owns *when* things run and *where* they are persisted, and nothing else.
+A backend that could add a status value or lift a retry cap would make the
+facade advisory, and the two backends would then differ in ways only production
+discovers.
+
+**`run_id` needs no new prefix.** Contracts §1.1 registers `run_` for "one CLI
+invocation or one unit of work" and §10.2 names every keyed parameter `run_id`.
+A workflow run is one unit of work, so it is a `run_` id -- adding a `wf_` prefix
+would have been a §1.1 change made to avoid reusing the entry that already
+describes this.
+"""
+
+from collections.abc import Callable, Mapping
+from typing import Any, Final, Literal, Protocol, runtime_checkable
+
+from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
+
+from adopt_const import (
+    IDEMPOTENCY_KEY_MAX_CHARS,
+    WORKFLOW_STEP_BACKOFF_BASE_MS,
+    WORKFLOW_STEP_BACKOFF_MAX_MS,
+    WORKFLOW_STEP_MAX_ATTEMPTS,
+)
+from adopt_obs import AdoptError, ErrorCode
+
+__all__ = [
+    "TERMINAL_STATUSES",
+    "Backoff",
+    "RetryPolicy",
+    "StepContext",
+    "WorkflowClient",
+    "WorkflowContext",
+    "WorkflowHandle",
+    "WorkflowStatus",
+    "backoff_delays_ms",
+    "validate_idempotency_key",
+]
+
+#: The run lifecycle. `completed`, `failed` and `cancelled` are terminal; the
+#: kill-and-resume drill asserts `completed` after a resume, so a resumed run
+#: must reach the same value a run that never died would.
+WorkflowStatus = Literal["pending", "running", "completed", "failed", "cancelled"]
+
+#: A run in one of these states will not change again. Both backends are asked
+#: for this set rather than each deciding which of their states are final.
+TERMINAL_STATUSES: Final[frozenset[str]] = frozenset({"completed", "failed", "cancelled"})
+
+#: Contracts §10.2 shows `backoff="exponential"`. `none` exists for a step whose
+#: failure is not worth waiting on; there is deliberately no `linear`, because a
+#: third curve is a third thing to reason about at 3 a.m. for no stated need.
+Backoff = Literal["exponential", "none"]
+
+
+class RetryPolicy(BaseModel):
+    """How many times a step is retried, and how long between attempts.
+
+    `max_attempts` is **capped**, not defaulted, by `WORKFLOW_STEP_MAX_ATTEMPTS`.
+    A caller asking for more is refused rather than silently clamped: a step
+    declared with 50 attempts was written against an expectation the platform
+    does not meet, and clamping hides that until someone reads a log.
+    """
+
+    model_config = ConfigDict(extra="forbid", frozen=True)
+
+    max_attempts: int = Field(default=WORKFLOW_STEP_MAX_ATTEMPTS, ge=1)
+    backoff: Backoff = "exponential"
+    base_ms: int = Field(default=WORKFLOW_STEP_BACKOFF_BASE_MS, ge=0)
+    max_ms: int = Field(default=WORKFLOW_STEP_BACKOFF_MAX_MS, ge=0)
+
+    @field_validator("max_attempts")
+    @classmethod
+    def _within_cap(cls, value: int) -> int:
+        if value > WORKFLOW_STEP_MAX_ATTEMPTS:
+            raise ValueError(
+                f"max_attempts={value} exceeds WORKFLOW_STEP_MAX_ATTEMPTS="
+                f"{WORKFLOW_STEP_MAX_ATTEMPTS}. The cap is the platform's, not the "
+                f"step's; raise it in implementation spec §2.2 and adopt_const together."
+            )
+        return value
+
+    # A cross-field rule, so it runs after the model is built rather than as a
+    # `field_validator` taking `ValidationInfo` -- that parameter is generic in
+    # `Any`, and `mypy.ini` sets `disallow_any_decorated`.
+    @model_validator(mode="after")
+    def _max_not_below_base(self) -> "RetryPolicy":
+        if self.max_ms < self.base_ms:
+            raise ValueError(
+                f"max_ms={self.max_ms} is below base_ms={self.base_ms}, so the first "
+                f"delay would already exceed the ceiling and every later one would be "
+                f"clamped to it."
+            )
+        return self
+
+
+def backoff_delays_ms(policy: RetryPolicy) -> tuple[int, ...]:
+    """The delay before each retry, in order. Length is `max_attempts - 1`.
+
+    A pure function of the policy, so the schedule can be asserted without
+    running a workflow and without either backend reproducing it independently.
+    There is no jitter: this is a per-run schedule for a single-writer local
+    backend and a DBOS queue, not a thundering herd of clients against one
+    service. Adding jitter would also make the sequence untestable by equality,
+    which is the property that keeps the two backends honest about the cap.
+    """
+    if policy.backoff == "none":
+        return tuple(0 for _ in range(policy.max_attempts - 1))
+    delays: list[int] = []
+    delay = policy.base_ms
+    for _ in range(policy.max_attempts - 1):
+        delays.append(min(delay, policy.max_ms))
+        delay *= 2
+    return tuple(delays)
+
+
+def validate_idempotency_key(key: str) -> str:
+    """Contracts §1.5: opaque, non-empty, at most `IDEMPOTENCY_KEY_MAX_CHARS`.
+
+    Refused at the seam rather than at a backend, so both backends refuse the
+    same keys. A key silently truncated by a column width is two different runs
+    that look like a replay of each other.
+    """
+    if not key:
+        raise AdoptError(
+            ErrorCode.WORKFLOW_DUPLICATE_START,
+            message="an idempotency key is required to start a workflow",
+            hint=(
+                "Contracts §1.5: every retriable operation takes a key, because "
+                "every message is assumed to be delivered twice."
+            ),
+        )
+    if len(key) > IDEMPOTENCY_KEY_MAX_CHARS:
+        raise AdoptError(
+            ErrorCode.WORKFLOW_DUPLICATE_START,
+            message=f"idempotency key is {len(key)} characters, over the "
+            f"{IDEMPOTENCY_KEY_MAX_CHARS}-character limit",
+            hint="Hash the caller's key rather than truncating it; a truncated key collides.",
+        )
+    return key
+
+
+class WorkflowHandle(BaseModel):
+    """What `start` returns, and what a replayed `start` returns unchanged."""
+
+    model_config = ConfigDict(extra="forbid", frozen=True)
+
+    run_id: str
+    name: str
+    version: int
+    idempotency_key: str
+    status: WorkflowStatus
+
+
+@runtime_checkable
+class StepContext(Protocol):
+    """Contracts §10.2, verbatim: `run_id`, `attempt`, and `dedupe`."""
+
+    run_id: str
+    attempt: int
+
+    def dedupe(self, key: str) -> bool:
+        """`True` when this is the first commit for `key`; `False` on a replay.
+
+        The exactly-once boundary. Steps run at least once, so the effect and
+        its dedupe record must commit **together** -- a backend that writes the
+        record after the effect has a window in which a crash duplicates the
+        effect, and that window is exactly what the durability drill opens.
+        """
+        ...
+
+
+@runtime_checkable
+class WorkflowContext(Protocol):
+    """What a workflow body is handed.
+
+    Deliberately tiny. Everything non-deterministic reaches the body through
+    `step`, because a body is **replayed**: on resume the engine re-executes it
+    and expects the same decisions. `workflow-body-purity` enforces the
+    negative half of that at lint and import time; this Protocol is the positive
+    half -- the only door out of a body.
+    """
+
+    run_id: str
+
+    def step(self, fn: Callable[..., Any], /, *args: Any, **kwargs: Any) -> Any:
+        """Execute `fn` as a durable step, or replay its recorded result."""
+        ...
+
+
+@runtime_checkable
+class WorkflowClient(Protocol):
+    """Contracts §10.2, plus `list` *(CR-42)*.
+
+    `list` is named by PRD F14.1 and omitted from §10.2's Protocol, and
+    implementation spec §7.4's rollback surface needs it: flipping
+    `ADOPT_FEATURE_DBOS_BACKEND` off requires in-flight runs to drain first, and
+    "drain" is unobservable without a way to enumerate them.
+    """
+
+    def start(
+        self,
+        fn: Callable[..., Any],
+        args: Mapping[str, Any],
+        *,
+        idempotency_key: str,
+    ) -> WorkflowHandle: ...
+
+    def signal(self, run_id: str, name: str, payload: Mapping[str, Any]) -> None: ...
+
+    def status(self, run_id: str) -> WorkflowStatus: ...
+
+    def result(self, run_id: str, *, timeout_s: int) -> Any: ...
+
+    def cancel(self, run_id: str) -> None: ...
+
+    def close(self) -> None:
+        """Stop executing work and release what the client holds *(CR-43)*.
+
+        **A backend that can start workers must be able to stop them**, and
+        §10.2 declared no way to. That is not a test convenience: implementation
+        spec §7.4's rollback flips `ADOPT_FEATURE_DBOS_BACKEND` off after
+        in-flight runs drain, and "drained" followed by a process that is still
+        dequeuing is not drained.
+
+        The durability drill is where the omission became visible. Its parent
+        process holds a client across three tests while each test spawns a child
+        that is meant to be the *only* executor -- and a queue is shared by every
+        worker pointed at it, so the parent silently dequeued the child's run and
+        executed it in the wrong process. That is DBOS behaving correctly and the
+        seam giving the caller no way to say "I am no longer a worker".
+
+        Closing is **idempotent** and does not cancel or roll back running
+        work: durable runs outlive the client by design, and the next process to
+        recover them picks them up. A backend holding nothing may do nothing.
+        """
+        ...
+
+    def recover(self) -> list[WorkflowHandle]:
+        """Re-drive every non-terminal run; return what was resumed.
+
+        Also CR-42. Implementation spec §4.14 says a crash between two step
+        records "replays the step", which presumes an entry point where the
+        replay begins -- DBOS reaches it at launch, and the in-process backend
+        has to be told. Declaring it here is what lets **one** durability suite
+        drive both: a drill that called `recover()` on one backend and relied on
+        a constructor side effect on the other would be two drills wearing one
+        name.
+        """
+        ...
+
+    # `list` is declared last: it shadows the builtin for every annotation after
+    # it in the class body, and the contract fixes the method name.
+    def list(self, *, status: WorkflowStatus | None = None) -> list[WorkflowHandle]: ...

adopt_workflow/decorators.py

@@ -0,0 +1,176 @@
+"""`@workflow`, `@step`, `@scheduled` -- contracts §10.2 and PRD F14.
+
+**Registration, not execution.** A decorator here records what a function *is*
+and validates that it may be one. Running it is a backend's job, and keeping the
+two apart is what lets one suite run against both backends: the registry is
+shared, so `inproc` and DBOS are handed identical definitions.
+
+**`@workflow` refuses an impure body at import time.** PRD F14.4 requires the
+purity check "at import time", and this is where import time happens. The
+alternative -- checking on the first run -- means the failure appears in whatever
+environment first executes the workflow, which is usually production, and the
+symptom there is a divergent replay rather than an error naming the line.
+
+**`@scheduled` is not a workflow, and that is the point.** PRD F14.5 says
+periodic single-step work uses cron rather than a workflow, and calls registering
+one as the other "a review rejection". A review line catches it when someone
+reads carefully; this decorator catches it always -- `@scheduled` refuses to
+decorate a `@workflow`, so the rejection is mechanical and does not depend on the
+reviewer having F14.5 in mind.
+"""
+
+from collections.abc import Callable
+from dataclasses import dataclass
+from typing import Any, Final, TypeVar
+
+from adopt_obs import AdoptError, ErrorCode
+from adopt_workflow.api import RetryPolicy
+from adopt_workflow.purity import assert_pure
+
+__all__ = [
+    "REGISTRY",
+    "ScheduledDefinition",
+    "StepDefinition",
+    "WorkflowDefinition",
+    "clear_registry",
+    "resolve",
+    "scheduled",
+    "step",
+    "workflow",
+]
+
+F = TypeVar("F", bound=Callable[..., Any])
+
+#: Marks set on the decorated function. Read by the backends and by `@scheduled`
+#: when it refuses a workflow; kept as dunder-free attributes so a definition is
+#: inspectable from a test without reaching into a private registry.
+WORKFLOW_ATTR: Final[str] = "__adopt_workflow__"
+STEP_ATTR: Final[str] = "__adopt_step__"
+SCHEDULED_ATTR: Final[str] = "__adopt_scheduled__"
+
+
+@dataclass(frozen=True)
+class WorkflowDefinition:
+    """What `@workflow` records. `(name, version)` is the identity."""
+
+    name: str
+    version: int
+    fn: Callable[..., Any]
+
+
+@dataclass(frozen=True)
+class StepDefinition:
+    """What `@step` records, including the policy its retries are capped by."""
+
+    name: str
+    retries: RetryPolicy
+    fn: Callable[..., Any]
+
+
+@dataclass(frozen=True)
+class ScheduledDefinition:
+    """A periodic single-step job. Cron, deliberately not a workflow."""
+
+    name: str
+    cron: str
+    fn: Callable[..., Any]
+
+
+#: One registry for the programme, keyed by `(kind, name, version)`.
+#:
+#: A backend resolves a definition from here rather than from a closure, which is
+#: what makes a resumed run able to find the body it was executing: after a
+#: process death there is no closure left, only a name in a journal.
+REGISTRY: Final[dict[tuple[str, str, int], object]] = {}
+
+
+def clear_registry() -> None:
+    """Empty the registry. For tests that declare throwaway workflows."""
+    REGISTRY.clear()
+
+
+def resolve(kind: str, name: str, version: int = 1) -> Any:
+    """The definition registered under `(kind, name, version)`.
+
+    Raises rather than returning `None`: a resumed run that cannot find its body
+    must fail loudly, because the alternative is a run that silently never
+    completes and a queue that never drains.
+    """
+    key = (kind, name, version)
+    if key not in REGISTRY:
+        known = ", ".join(sorted(f"{k[0]}:{k[1]}@{k[2]}" for k in REGISTRY))
+        raise AdoptError(
+            ErrorCode.WORKFLOW_STEP_EXHAUSTED,
+            message=f"no {kind} registered as {name!r} version {version}",
+            hint=(
+                f"A resumed run resolves its body by name; the module declaring it "
+                f"must be imported before the backend starts. Registered: {known or 'none'}"
+            ),
+        )
+    return REGISTRY[key]
+
+
+def workflow(*, name: str, version: int = 1) -> Callable[[F], F]:
+    """Register a deterministic, replayable workflow body.
+
+    The body is checked for purity **now**, at decoration time.
+    """
+
+    def decorate(fn: F) -> F:
+        assert_pure(fn)
+        definition = WorkflowDefinition(name=name, version=version, fn=fn)
+        key = ("workflow", name, version)
+        if key in REGISTRY:
+            raise AdoptError(
+                ErrorCode.WORKFLOW_DUPLICATE_START,
+                message=f"a workflow named {name!r} version {version} is already registered",
+                hint=(
+                    "`(name, version)` is the identity a resumed run resolves through. "
+                    "Two bodies under one key means a resume can execute the wrong one."
+                ),
+            )
+        REGISTRY[key] = definition
+        setattr(fn, WORKFLOW_ATTR, definition)
+        return fn
+
+    return decorate
+
+
+def step(*, name: str, retries: RetryPolicy | None = None) -> Callable[[F], F]:
+    """Register a step: the only place non-deterministic work may live."""
+
+    def decorate(fn: F) -> F:
+        definition = StepDefinition(name=name, retries=retries or RetryPolicy(), fn=fn)
+        REGISTRY[("step", name, 1)] = definition
+        setattr(fn, STEP_ATTR, definition)
+        return fn
+
+    return decorate
+
+
+def scheduled(*, name: str, cron: str) -> Callable[[F], F]:
+    """Register periodic single-step work. **Not** a workflow -- PRD F14.5.
+
+    Refuses to decorate a function that is already a workflow. The durable
+    machinery exists for work that must survive a crash mid-sequence; a single
+    step that runs every hour survives by running again next hour, and putting it
+    on the engine buys retention, replay and a queue nobody needed.
+    """
+
+    def decorate(fn: F) -> F:
+        if hasattr(fn, WORKFLOW_ATTR):
+            raise AdoptError(
+                ErrorCode.WORKFLOW_BODY_IMPURE,
+                message=f"{name!r} is registered as both a workflow and a scheduled job",
+                hint=(
+                    "PRD F14.5: periodic single-step work uses cron, not a workflow. "
+                    "If the job really is a multi-step sequence that must survive a "
+                    "crash halfway, drop @scheduled and start it from cron instead."
+                ),
+            )
+        definition = ScheduledDefinition(name=name, cron=cron, fn=fn)
+        REGISTRY[("scheduled", name, 1)] = definition
+        setattr(fn, SCHEDULED_ATTR, definition)
+        return fn
+
+    return decorate

adopt_workflow/inproc/__init__.py

@@ -0,0 +1,6 @@
+"""The in-process backend: no external dependency, for CI and local dev."""
+
+from adopt_workflow.inproc.client import InProcessStepContext, InProcessWorkflowClient
+from adopt_workflow.inproc.journal import Journal
+
+__all__ = ["InProcessStepContext", "InProcessWorkflowClient", "Journal"]