agnes-the-ai-analyst/app/resource_types.py
ZdenekSrotyr 70672204fe
feat(memory): admin Edit + MEMORY_DOMAIN RBAC + ai-section UI (#141)
Cuts release 0.23.0.

## Highlights
- Single-item Edit button on every memory item card (modal hits PATCH /api/memory/admin/{id}).
- MEMORY_DOMAIN RBAC resource type — admins grant user_groups access to specific domains via /admin/access. Composes with existing audience filter (OR semantics, no-op when no grants).
- ai: section editable in /admin/server-config — admins can set ANTHROPIC_API_KEY / model / provider / base_url for the corporate-memory extractor without editing instance.yaml directly. api_key auto-masked.

## Devin findings addressed
- Modal NULL→empty fix (audience visibility wouldn't break).
- Stats endpoint granted_domains parity with list endpoint.
- Documented intentional MEMORY_DOMAIN→audience bypass.
- Documented conscious ai.base_url SSRF exclusion (legit internal LiteLLM/vLLM proxies).

See CHANGELOG [0.23.0] for full notes.
2026-04-30 11:04:41 +02:00

289 lines
11 KiB
Python

"""Resource types that can be granted to user groups.
A *resource type* identifies a class of entity admins can hand out access to
(e.g. marketplace plugins, datasets). Concrete instances live in their own
domain tables (`marketplace_plugins`, `table_registry`, …); access to a
specific instance is recorded as a row in `resource_grants` with this enum
value as ``resource_type`` and a module-defined path string as ``resource_id``.
Adding a new type — single place, no separate wiring step:
1. Add a member to :class:`ResourceType`.
2. Write a ``list_blocks(conn) -> list[Block]`` delegate that projects the
domain tables into the (block → items) tree the admin /access page
consumes. Each item must include ``resource_id`` matching the path
string used in ``resource_grants.resource_id``.
3. Register a :class:`ResourceTypeSpec` in :data:`RESOURCE_TYPES`. The
dataclass requires ``list_blocks`` — the type checker forces step 2.
4. Wire endpoints with
``Depends(require_resource_access(ResourceType.X, "<path>"))``.
No DB migration needed — this is application-level configuration. Membership
in the enum + registry is the source of truth; the DB just stores the string
value verbatim.
"""
from __future__ import annotations
import os
from dataclasses import dataclass
from enum import StrEnum
from typing import TYPE_CHECKING, Any, Callable, List
if TYPE_CHECKING:
import duckdb
class ResourceType(StrEnum):
"""Resource categories that the access-control layer understands.
Values are persisted verbatim in ``resource_grants.resource_type``.
Renaming a member is a breaking change — existing grants reference the
string. Add a new member and migrate via SQL UPDATE if needed.
"""
MARKETPLACE_PLUGIN = "marketplace_plugin"
TABLE = "table"
MEMORY_DOMAIN = "memory_domain"
# Shape returned by ``list_blocks`` delegates. Kept as plain ``dict`` to keep
# the registry decoupled from any specific ORM/repo type — UI consumes JSON.
Block = dict[str, Any]
ListBlocksFn = Callable[["duckdb.DuckDBPyConnection"], List[Block]]
@dataclass(frozen=True)
class ResourceTypeSpec:
"""Self-contained definition of a resource type.
Bundles UI copy with the projection delegate so that adding a new type
in :data:`RESOURCE_TYPES` is the single place that needs editing — no
forgotten branch in ``access-overview`` or the admin UI.
Attributes:
key: The enum member; ``key.value`` is what gets persisted.
display_name: Plural label rendered as a section header on the
admin /access page.
description: One-liner shown in the create-grant form's helper text.
id_format: Human-readable hint for ``resource_id`` shape — e.g.
``"<marketplace_slug>/<plugin_name>"``. Surfaced as input
placeholder.
list_blocks: Delegate that takes a system DB connection and returns
``[{id, name, items: [{resource_id, name, ...}]}]`` — one block
per parent entity (e.g. marketplace), one item per grantable
resource (e.g. plugin). Items must carry ``resource_id`` that
matches the path string written into ``resource_grants``.
"""
key: ResourceType
display_name: str
description: str
id_format: str
list_blocks: ListBlocksFn
# ---------------------------------------------------------------------------
# Marketplace plugin projection
# ---------------------------------------------------------------------------
def _marketplace_plugin_blocks(conn: "duckdb.DuckDBPyConnection") -> List[Block]:
"""Project marketplace_registry + marketplace_plugins into the
hierarchical (block → items) shape the admin UI renders.
One block per marketplace_registry row, ordered by registered_at.
Items inside are plugins; ``resource_id`` encodes the canonical path
``<marketplace_slug>/<plugin_name>`` that ``resource_grants.resource_id``
matches against.
"""
rows = conn.execute(
"""SELECT mr.id, mr.name, mr.registered_at,
mp.name AS plugin_name, mp.version, mp.category,
mp.description, mp.source_type
FROM marketplace_registry mr
LEFT JOIN marketplace_plugins mp ON mp.marketplace_id = mr.id
ORDER BY mr.registered_at, mr.id, mp.name"""
).fetchall()
blocks: dict[str, Block] = {}
for mr_id, mr_name, _, p_name, p_ver, p_cat, p_desc, p_src in rows:
block = blocks.setdefault(mr_id, {
"id": mr_id,
"name": mr_name,
"items": [],
})
if p_name:
block["items"].append({
"resource_id": f"{mr_id}/{p_name}",
"name": p_name,
"version": p_ver,
"category": p_cat,
"description": p_desc,
"source_type": p_src,
})
return list(blocks.values())
# ---------------------------------------------------------------------------
# Table projection
# ---------------------------------------------------------------------------
def _table_blocks(conn: "duckdb.DuckDBPyConnection") -> List[Block]:
"""Project table_registry into the (block → items) shape the admin UI
renders.
One block per ``bucket`` value, ordered by bucket then table name.
Items inside are tables; ``resource_id`` is the ``table_registry.id``
primary key — that is the path string that ``resource_grants.resource_id``
matches against when enforcement lands. Bucket is purely a UI grouping
and does not enter the resource_id (mirrors the marketplace/plugin
pattern, where the marketplace itself is not a grantable resource).
Tables with NULL/empty bucket fall into a synthetic ``"(no bucket)"``
block so they are still grantable.
"""
rows = conn.execute(
"""SELECT id, name, bucket, source_type, query_mode, description
FROM table_registry
ORDER BY COALESCE(bucket, ''), name"""
).fetchall()
blocks: dict[str, Block] = {}
for tbl_id, name, bucket, source_type, query_mode, description in rows:
block_key = bucket if bucket else "(no bucket)"
block = blocks.setdefault(block_key, {
"id": block_key,
"name": block_key,
"items": [],
})
block["items"].append({
"resource_id": tbl_id,
"name": name,
"category": query_mode,
"source_type": source_type,
"description": description,
})
return list(blocks.values())
# ---------------------------------------------------------------------------
# Memory domain projection
# ---------------------------------------------------------------------------
# Mirrors VALID_DOMAINS in app/api/memory.py. Kept inline here to avoid
# importing the FastAPI module at registry-load time (circular import risk).
# If this list drifts from VALID_DOMAINS, add a runtime cross-check or merge
# the two sources — for now they're tiny and reviewed together.
_MEMORY_DOMAINS = (
"finance",
"engineering",
"product",
"data",
"operations",
"infrastructure",
)
def _memory_domain_blocks(conn: "duckdb.DuckDBPyConnection") -> List[Block]:
"""Project the (fixed) set of corporate-memory domains into the
(block → items) shape the admin UI renders.
Unlike marketplace plugins / tables, the grantable items are a fixed
enum, not a DB lookup — every deployment has the same 6 domains. One
synthetic block ``"Memory domains"`` holds them; ``resource_id`` is
the domain string (matches ``knowledge_items.domain``).
"""
return [{
"id": "memory_domains",
"name": "Memory domains",
"items": [
{
"resource_id": domain,
"name": domain,
"category": "domain",
"description": (
f"Members of granted groups see all knowledge_items "
f"with domain={domain!r}, in addition to the existing "
f"audience filter."
),
}
for domain in _MEMORY_DOMAINS
],
}]
# ---------------------------------------------------------------------------
# Registry — the one place that gets edited when adding a new resource type
# ---------------------------------------------------------------------------
RESOURCE_TYPES: dict[ResourceType, ResourceTypeSpec] = {
ResourceType.MARKETPLACE_PLUGIN: ResourceTypeSpec(
key=ResourceType.MARKETPLACE_PLUGIN,
display_name="Marketplace plugins",
description="A plugin from a registered marketplace.",
id_format="<marketplace_slug>/<plugin_name>",
list_blocks=_marketplace_plugin_blocks,
),
ResourceType.TABLE: ResourceTypeSpec(
key=ResourceType.TABLE,
display_name="Tables",
description="A registered data table.",
id_format="<table_id>",
list_blocks=_table_blocks,
),
ResourceType.MEMORY_DOMAIN: ResourceTypeSpec(
key=ResourceType.MEMORY_DOMAIN,
display_name="Memory domains",
description="A corporate-memory domain (knowledge_items.domain).",
id_format="<domain>",
list_blocks=_memory_domain_blocks,
),
}
def is_resource_type_enabled(rt: ResourceType) -> bool:
"""Whether a resource type is exposed to the admin UI + grant API.
``ResourceType.TABLE`` is gated behind ``AGNES_ENABLE_TABLE_GRANTS=1``
(default off) until the data-plane runtime check delegates to
``app.auth.access.can_access`` (currently still flows through legacy
``dataset_permissions``). Without runtime enforcement, exposing the chip
is misleading: admins grant access through /admin/access, but the user
still gets filtered out at /api/catalog. See
``docs/TODO-rbac-data-enforcement.md`` step 1.
Existing TABLE grants in ``resource_grants`` remain in the DB regardless
— this flag controls UI exposure + new-grant acceptance, not data.
"""
if rt is ResourceType.TABLE:
return os.environ.get("AGNES_ENABLE_TABLE_GRANTS", "").lower() in {
"1", "true", "yes", "on",
}
return True
def enabled_resource_types() -> list[ResourceTypeSpec]:
"""The subset of RESOURCE_TYPES currently surfaced to admins."""
return [spec for rt, spec in RESOURCE_TYPES.items() if is_resource_type_enabled(rt)]
def list_resource_types() -> list[dict[str, str]]:
"""Flat projection for /api/admin/resource-types.
Shape: ``[{key, display_name, description, id_format}]``. The
``list_blocks`` delegate is intentionally omitted — the UI consumes
blocks via ``/api/admin/access-overview`` instead. Disabled types
(e.g. TABLE without ``AGNES_ENABLE_TABLE_GRANTS``) are filtered out so
the admin UI does not advertise grants the runtime cannot enforce.
"""
return [
{
"key": spec.key.value,
"display_name": spec.display_name,
"description": spec.description,
"id_format": spec.id_format,
}
for spec in enabled_resource_types()
]