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

279 lines
12 KiB
Python

"""Config-driven extension loading.
Entry points are named as `module.path:install`, resolved through the same
`resolve_variable` helper the guardrails provider already uses. Load order is
the config list order — explicit and reproducible, which matters because the
middleware stack is position-sensitive.
"""
from __future__ import annotations
import logging
from collections.abc import Mapping, Sequence
from dataclasses import dataclass
from typing import Any, Literal
from deerflow_extension_api import API_VERSION
from pydantic import BaseModel, ConfigDict, Field
from deerflow.extensions.registry import ExtensionRegistry, LoadedExtensions
from deerflow.persistence.migrations._env_filters import register_extension_table_prefix
from deerflow.reflection import resolve_variable
logger = logging.getLogger(__name__)
DiagnosticLevel = Literal["debug", "info", "warning", "error"]
class ExtensionSpec(BaseModel):
"""One entry of the `plugins:` list in config.yaml."""
model_config = ConfigDict(extra="forbid")
enabled: bool = Field(
default=True,
description="When false, skip the extension without resolving or importing it",
)
name: str | None = Field(
default=None,
description="Stable operator-facing name recorded by the extension manager",
)
package: str | None = Field(
default=None,
description="Installed Python distribution recorded by the extension manager",
)
use: str = Field(description="Entry point path, e.g. 'my_extension:install'")
config: dict[str, Any] = Field(
default_factory=dict,
description="Extension-private configuration, passed to install() verbatim",
)
required: bool = Field(
default=False,
description="When true, a load failure aborts startup instead of being skipped",
)
table_prefix: str | None = Field(
default=None,
min_length=1,
description=(
"Table-name prefix this extension owns, if it persists data under its own "
"MetaData and migration chain. Registered with "
"deerflow.persistence.migrations._env_filters so alembic revision --autogenerate "
"excludes those tables instead of reflecting them from a live database and "
"proposing to drop them. Registered from two processes: here for the Gateway, "
"and from migrations/env.py -- reading this declaration, never importing the "
"extension -- for alembic, which never starts a Gateway. Omit the key to "
"declare no prefix; an empty string is rejected here rather than treated "
"as absent, so that one declaration cannot mean 'no prefix' to one of "
"those two processes and 'a prefix matching every table' to the other."
),
)
@dataclass(frozen=True)
class Diagnostic:
"""A load- or run-time problem attributed to a specific extension.
The repository has no structured diagnostics channel today; this is a
deliberately minimal one whose only job is keeping failures attributable.
"""
level: DiagnosticLevel
source: str
message: str
@classmethod
def error(cls, source: str, message: str) -> Diagnostic:
return cls("error", source, message)
@classmethod
def warning(cls, source: str, message: str) -> Diagnostic:
return cls("warning", source, message)
@classmethod
def info(cls, source: str, message: str) -> Diagnostic:
return cls("info", source, message)
@classmethod
def debug(cls, source: str, message: str) -> Diagnostic:
return cls("debug", source, message)
class ExtensionLoadError(RuntimeError):
"""Raised when an extension marked `required: true` fails to load."""
def _parse_version(version: object) -> tuple[int, ...] | None:
if not isinstance(version, str):
return None
try:
return tuple(int(part) for part in str.split(version, "."))
except ValueError:
return None
def _compatible(declared: str, current: str) -> bool:
"""One-directional, with the semver window for the contract's life stage.
Pre-1.0 minors may break, so the window is same major.minor with patches
additive: host >= declared.
From 1.0 on contracts only grow within a major, so a newer host stays
compatible with older extensions while an extension written against a
newer minor is refused — it would reach for contract additions the host
does not implement. Unparseable versions are refused, not waved through."""
declared_parts = _parse_version(declared)
current_parts = _parse_version(current)
if not declared_parts or not current_parts:
return False
width = max(len(declared_parts), len(current_parts), 2)
declared_padded = declared_parts + (0,) * (width - len(declared_parts))
current_padded = current_parts + (0,) * (width - len(current_parts))
if declared_padded[0] != current_padded[0]:
return False
if declared_padded[0] == 0 and declared_padded[1] != current_padded[1]:
return False
return current_padded >= declared_padded
def _range_for(declared: str) -> str:
"""The pip window matching ``_compatible``'s rules, for the actionable
refusal message. Falls back to an exact request when the declared version
is unparseable — the message must survive the version that caused it."""
parts = _parse_version(declared)
if not parts:
return f"=={declared}"
if parts[0] == 0:
minor = parts[1] if len(parts) > 1 else 0
return f">={declared},<0.{minor + 1}"
return f">={declared},<{parts[0] + 1}.0"
def load_extensions(specs: Sequence[ExtensionSpec]) -> tuple[LoadedExtensions, list[Diagnostic]]:
"""Resolve and install every configured extension.
Fail-open by default: a broken extension is skipped with a diagnostic so
the Gateway still starts. `required: true` flips that to fail-closed for
extensions whose absence changes behaviour rather than just observability.
"""
registry = ExtensionRegistry()
diagnostics: list[Diagnostic] = []
loaded_sources: list[str] = []
for spec in specs:
if spec.table_prefix:
# Registered unconditionally -- even for a disabled or later-failing
# spec -- because the tables it names may already exist in the
# database from a previous run. Excluding them from alembic's view
# is the safe direction; the risk this guards against is
# autogenerate proposing to drop them, not registering one prefix
# too many.
#
# A prefix that collides with a host table name is not a
# per-extension failure `required: false` can shrug off: it
# corrupts the shared alembic filter for that host table for the
# life of the process, regardless of whether this extension ever
# loads. It always aborts startup.
try:
register_extension_table_prefix(spec.table_prefix)
except ValueError as exc:
message = str(exc)
diagnostics.append(Diagnostic.error(spec.use, message))
logger.error("Extension %s: %s", spec.use, message)
raise ExtensionLoadError(message) from exc
if not spec.enabled:
continue
try:
install = resolve_variable(spec.use)
except Exception as exc:
message = f"could not resolve extension entry point: {exc}"
diagnostics.append(Diagnostic.error(spec.use, message))
logger.error("Extension %s: %s", spec.use, message)
if spec.required:
raise ExtensionLoadError(f"required extension {spec.use} failed to load") from exc
continue
if not callable(install):
message = f"extension entry point is not callable: {type(install).__name__}"
diagnostics.append(Diagnostic.error(spec.use, message))
logger.error("Extension %s: %s", spec.use, message)
if spec.required:
raise ExtensionLoadError(f"required extension {spec.use} is not callable")
continue
try:
declared = getattr(install, "__deerflow_api__", None)
except Exception as exc:
message = f"could not inspect extension-api version marker: {type(exc).__name__}"
diagnostics.append(Diagnostic.error(spec.use, message))
logger.error("Extension %s: %s", spec.use, message)
if spec.required:
raise ExtensionLoadError(f"required extension {spec.use} could not inspect api marker") from exc
continue
if declared is not None and _parse_version(declared) is None:
message = f"extension declares invalid extension-api version marker of type {type(declared).__name__}; expected a dotted numeric string such as '0.1'"
diagnostics.append(Diagnostic.error(spec.use, message))
logger.error("Extension %s: %s", spec.use, message)
if spec.required:
raise ExtensionLoadError(f"required extension {spec.use} declares invalid api marker")
continue
if declared is not None:
# ``isinstance(..., str)`` also accepts subclasses whose
# ``__str__``/``__format__`` methods can execute plugin code while
# we build an incompatibility diagnostic. Normalize with the base
# implementation before compatibility checks and rendering.
declared = str.__str__(declared)
if declared is not None and not _compatible(declared, API_VERSION):
message = f"extension requires extension-api {declared}, host provides {API_VERSION}. Install a matching version: pip install 'deerflow-extension-api{_range_for(declared)}'"
diagnostics.append(Diagnostic.error(spec.use, message))
logger.error("Extension %s: %s", spec.use, message)
if spec.required:
raise ExtensionLoadError(f"required extension {spec.use} declares incompatible api {declared}")
continue
# Positional rollback, not registry.discard(spec.use): two specs may
# legitimately share the same `use` with different config, and
# discard-by-source would also erase an earlier, successfully
# installed instance that happens to share this spec's `use`.
mark = registry.mark()
try:
with registry.attributed_to(spec.use):
install(registry, _frozen_config(spec.config))
except Exception as exc:
registry.rollback_to(mark)
message = f"install() failed: {exc}"
diagnostics.append(Diagnostic.error(spec.use, message))
logger.exception("Extension %s: install() failed", spec.use)
if spec.required:
raise ExtensionLoadError(f"required extension {spec.use} failed to install") from exc
continue
loaded_sources.append(spec.use)
# Loading third-party code is exactly the event an operator needs positive
# confirmation of, and every other branch here is failure-only — so without
# this line a fully successful load is indistinguishable from a `plugins:`
# block the host never read. The x/y count names the difference between
# "all loaded" and "some were skipped" without repeating the per-failure
# errors already logged above.
if specs:
logger.info("Extensions loaded: %d/%d (%s)", len(loaded_sources), len(specs), ", ".join(loaded_sources) or "none")
else:
# Debug, not info: no configured plugins is the default state for almost
# every deployment, and an unconditional line would be pure boot noise.
logger.debug("No extensions configured")
return registry.build(), diagnostics
def _frozen_config(config: dict[str, Any]) -> Mapping[str, Any]:
"""Hand extensions a shallow copy of their config block.
This is a shallow copy: it stops an extension from reassigning
top-level keys on another extension's (or the caller's) config dict, but
nested structures (lists, dicts) are still shared by reference and can be
mutated in place. Use plain, top-level config values if this guarantee
matters to you.
"""
return dict(config)