Nan Gao 13f0a7f263
feat(extensions): let an out-of-tree extension observe what the agent did (#4863)
* feat(extensions): let an out-of-tree extension observe what the agent did

DeerFlow's extension system can contribute middleware, services and routes,
but an extension cannot answer basic questions about a run without reaching
into host internals. Several of the facts it would need are destroyed by the
operations that produce them:

  * The middleware chain injects and rewrites a lot of context — date
    reminders, recalled memory, compaction summaries, durable-context data,
    image payloads, activated skill bodies. Downstream, none of it is
    attributable: at the model-call boundary an injected HumanMessage is
    indistinguishable from the user's own, and anything wanting to tell them
    apart has to pattern-match prompt wording, which breaks on the next copy
    edit.

  * Two runs of "the same agent" are only comparable if the chain enforced the
    same limits, prompts and thresholds. Recovering that from outside means
    reading private attributes and guessing which of them change behaviour — a
    guess that rots silently as middlewares gain fields.

  * The lead-agent factory resolves a model after runtime overrides, renders a
    prompt, filters tools through authorization and composes a stack, all
    inside one synchronous call, and none of it survives: a middleware sees its
    neighbours but not the prompt, the run worker sees a graph but not what
    went into it.

  * Summarization is destructive by design. N messages leave the context and
    one summary enters it; afterwards only the summary exists, so "which
    messages became this?" is not reconstructible.

This adds seven neutral facilities so those facts are recorded where they are
still true, and releases the contract package as 0.2.0.

Message provenance
  Producers stamp `deerflow_content_kind` / `deerflow_producer_kind` onto the
  messages they inject or rewrite. Stamping is unconditional — a fact whose
  presence depends on whether an observer is installed is not a fact — and the
  keys are server-owned, so provenance cannot be forged from a request.

Middleware self-description
  Twelve middlewares declare their own behaviour-affecting parameters through
  a duck-typed `release_policy_parameters()`. Long text is hashed rather than
  embedded: a declaration is an identity, not a copy of the prompt.

Agent assembly descriptor
  `assemble_lead_agent()` returns the graph plus a descriptor whose fingerprint
  answers "did anything about this agent change between these two runs?".
  `make_lead_agent()` keeps its graph-only signature — it is the LangGraph
  Server ABI declared in langgraph.json. Tools and skills are sorted before
  hashing because their assembly order is incidental; middlewares are not,
  because stack order decides what wraps what. Host build identity is reported
  but excluded from the fingerprint, so a redeploy does not invalidate every
  agent's identity.

Context compaction observation
  Summarization emits the content hashes of the messages it is about to remove
  joined to the summary that replaced them. Content is the only identity
  available at that seam: the summary does not become a message, and what later
  projects it into a request renders it bounded and escaped rather than
  verbatim.

Neutral policy, transform and MCP-source facts
  Guardrail decisions are published to runtime context under a `__`-prefixed
  key; result-rewriting middlewares append a declared, ordered transform trail;
  MCP tools carry their credential-free logical origin.

Extension route identity
  Contributed routes are session-authenticated and cannot opt out, but
  "logged in" and "administrator" are different questions. Extensions get a
  neutral projection of the caller rather than the host's auth context, and
  `require_admin` fails closed when identity cannot be determined.

Extension-owned tables
  An extension that persists data owns its own MetaData and migration chain, so
  its tables are absent from Base.metadata and `alembic revision --autogenerate`
  proposes dropping them. Extensions declare a table prefix, which is rejected
  at registration if it would shadow a host table.

The contract package stays dependency-free and imports no host code; every new
Protocol method has a default so later additions remain additive. The loader's
pre-1.0 rule requires an exact major.minor match, so extensions written against
0.1 are now refused at startup with an actionable install hint rather than
loading into a host that implements a different surface.

uv.lock records the contract package's new version, so `uv sync --locked` still
resolves on a fresh checkout.

* fix(backend): sort gateway service imports
2026-08-23 09:57:12 +08:00

240 lines
7.5 KiB
Python

"""The extension contracts and their data types.
Compatibility rules enforced throughout this module:
* every Protocol method carries a default implementation, so adding a method
later stays additive for already-released extensions;
* every optional dataclass field carries a default, so adding a field stays
additive.
"""
from __future__ import annotations
from collections.abc import Callable, Mapping, Sequence
from dataclasses import dataclass, field
from enum import StrEnum
from typing import TYPE_CHECKING, Any, Literal, Protocol, TypeVar, runtime_checkable
from deerflow_extension_api.state import ExtensionData
if TYPE_CHECKING: # pragma: no cover - typing only
from deerflow_extension_api.assembly import AgentAssemblyObserver
from deerflow_extension_api.compaction import ContextCompactionObserver
from deerflow_extension_api.placement import AgentBuildContext, MiddlewarePlacement
F = TypeVar("F", bound=Callable[..., Any])
# --- Host projections -------------------------------------------------------
@dataclass(frozen=True)
class HostPolicySnapshot:
"""The limits the host actually enforces, projected for extensions.
A narrow projection instead of the host's AppConfig: exposing AppConfig
would pin every extension to the harness release cadence. Every field has
a default so widening this stays additive.
"""
token_budget_enabled: bool = False
max_input_tokens: int | None = None
max_output_tokens: int | None = None
max_total_tokens: int | None = None
budget_warn_fraction: float | None = None
budget_hard_fraction: float | None = None
max_subagents_per_run: int | None = None
# --- Task lifecycle ---------------------------------------------------------
class TaskOutcome(StrEnum):
COMPLETED = "completed"
ABORTED = "aborted"
FAILED = "failed"
@dataclass(frozen=True)
class TaskInfo:
"""Identity of one lead-agent or subagent execution."""
task_id: str
run_id: str
thread_id: str
kind: Literal["lead", "subagent"]
parent_task_id: str | None = None
agent_name: str | None = None
resumed: bool = False
class TaskLifecycleContributor(Protocol):
async def on_task_start(
self,
app_store: ExtensionData,
task_store: ExtensionData,
info: TaskInfo,
) -> None:
return None
async def on_task_stop(
self,
app_store: ExtensionData,
task_store: ExtensionData,
info: TaskInfo,
outcome: TaskOutcome,
) -> None:
return None
# --- System model calls not wrapped by middleware model-call hooks ----------
class SystemOperationKind(StrEnum):
GOAL = "goal"
MEMORY = "memory"
TITLE = "title"
SUMMARIZATION = "summarization"
@dataclass(frozen=True)
class SystemModelRequest:
"""Read-only snapshot taken before a system-owned model call."""
messages: Sequence[Any] = ()
model_name: str | None = None
invoke_config: Mapping[str, Any] | None = None
def __post_init__(self) -> None:
"""Normalize ``messages`` to a tuple so the snapshot is what it claims to be.
Call sites differ: goal evaluation and memory extraction pass a message list,
while title generation and summarization pass one prompt string. A bare ``str``
already satisfies ``Sequence``, so without this an observer iterating
``request.messages`` would silently walk characters. Copying a list also makes
the frozen snapshot immutable in fact, not only by dataclass declaration — the
caller keeps its own list and observations may run after the call returns.
"""
messages = self.messages
if isinstance(messages, tuple):
return
normalized = tuple(messages) if isinstance(messages, Sequence) and not isinstance(messages, str | bytes) else (messages,)
object.__setattr__(self, "messages", normalized)
@dataclass(frozen=True)
class SystemModelResult:
"""Success or failure snapshot taken after a system-owned model call."""
response: Any | None = None
error: BaseException | None = None
duration_ms: float | None = None
class SystemModelCallObserver(Protocol):
async def on_system_model_call(
self,
app_store: ExtensionData,
task_store: ExtensionData,
kind: SystemOperationKind,
request: SystemModelRequest,
result: SystemModelResult,
) -> None:
return None
# --- Middleware -------------------------------------------------------------
class MiddlewareContributor(Protocol):
def contribute_middlewares(
self,
app_store: ExtensionData,
ctx: AgentBuildContext,
) -> Sequence[MiddlewarePlacement]:
return ()
# --- Extension services ----------------------------------------------------
@dataclass(frozen=True)
class ExtensionRuntimeDeps:
"""Host capabilities bound after Gateway infrastructure is ready."""
app_store: ExtensionData | None = None
policy: HostPolicySnapshot = field(default_factory=HostPolicySnapshot)
session_factory: Any | None = None
class ExtensionService(Protocol):
async def start(self, deps: ExtensionRuntimeDeps) -> None:
return None
async def stop(self) -> None:
return None
# --- Registration surface ---------------------------------------------------
@runtime_checkable
class ExtensionRegistry(Protocol):
"""The write-only registration surface handed to ``install()``.
Structural and minimal on purpose. Every method has a default so additive
contract releases remain compatible with older registry implementations.
The host's concrete registry additionally carries host-only machinery
(attribution, positional rollback, build) that is deliberately absent here.
"""
def middlewares(self, contributor: MiddlewareContributor) -> None:
return None
def task_lifecycle(self, contributor: TaskLifecycleContributor) -> None:
return None
def system_model_observer(self, observer: SystemModelCallObserver) -> None:
return None
def agent_assembly_observer(self, observer: AgentAssemblyObserver) -> None:
return None
def context_compaction_observer(self, observer: ContextCompactionObserver) -> None:
return None
def service(self, service: ExtensionService) -> None:
return None
def routers(self, routers: Sequence[Any]) -> None:
"""Register HTTP routers constructed eagerly during extension install.
Router types stay ``Any`` so this contract package has no FastAPI
dependency. The host validates supported route shapes before mounting;
runtime resources belong in a separately registered service.
"""
return None
#: The install() entry point signature every extension exposes.
ExtensionInstall = Callable[[ExtensionRegistry, Mapping[str, Any]], None]
# --- Declaration decorator --------------------------------------------------
def extension(*, api: str, name: str | None = None) -> Callable[[F], F]:
"""Stamp an install function with the API version it was written against.
Optional. pip's dependency resolution is the primary compatibility
mechanism; this covers `--no-deps` installs and editable monorepo checkouts
where versions can skew, and turns a deep AttributeError into an
actionable startup diagnostic.
"""
def _decorate(func: F) -> F:
func.__deerflow_api__ = api # type: ignore[attr-defined]
func.__deerflow_name__ = name # type: ignore[attr-defined]
return func
return _decorate