docs: plan source registry store

This commit is contained in:
lda
2026-06-03 23:01:17 +07:00 Verified
parent 6a9c71b5d9
commit d530f089be
3 changed files with 706 additions and 0 deletions
+2
View File
@@ -118,6 +118,8 @@ implementation state.
source registry is designed. Config can bootstrap sources, but server-owned
dynamic source changes need persistence, validation, and auth rules before
they are safe.
- The store-backed source registry design is recorded in
[2026-06-03 store-backed source registry](./superpowers/specs/2026-06-03-store-backed-source-registry-design.md).
- Longer term: make the MCP frontend an adapter over these neutral workflow,
source-admin, and config-admin surfaces so the old `wf_mcp` server entry
point can shrink or retire.
@@ -0,0 +1,433 @@
# Source Registry Store Slice 1 Implementation Plan
> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
**Goal:** Add validated source registry models and a filesystem store for desired server-owned source configuration, without wiring it into startup or mutation commands yet.
**Architecture:** The new registry is desired configuration state and stays separate from existing auth/catalog storage. Models live in `wf_mcp` for this first slice because current connection/source registry semantics are still MCP-provider-specific; the file-store interface is small enough to move later. Runtime merge, RPC/CLI mutation, and config reconciliation are explicitly deferred.
**Tech Stack:** Python 3.14, Pydantic v2, existing `wf_mcp.connections.parse_connection_id`, existing `RESERVED_CONNECTION_IDS`, pytest, ruff, basedpyright.
---
## Scope
In scope:
- Typed `SourceRegistryFile` model.
- Typed `McpSourceRegistryEntry` model.
- Typed `StdioSourceTransport` and `HttpSourceTransport` models.
- Duplicate id validation.
- Reserved id validation.
- ID validation using existing connection id rules.
- `SourceRegistryStore` protocol.
- `FileSourceRegistryStore` using `<store_root>/source_registry.json`.
- Atomic filesystem writes.
- Tests for load missing file, save/load round trip, validation errors, and path.
Out of scope:
- Startup merge with config.
- Runtime hydration from registry.
- Mutating API/CLI/RPC commands.
- Auth record changes.
- Catalog deletion or cleanup.
- SQL/remote stores.
## File Structure
- Create `src/wf_mcp/source_registry.py`
- Pydantic models and validation.
- Protocol and file store.
- No dependency on `WfMcpService`.
- Modify `src/wf_mcp/storage/__init__.py`
- No change in this slice unless the implementor decides a store export is needed.
- Prefer exporting from `wf_mcp.source_registry`, not overloading `wf_mcp.storage`.
- Create `tests/wf_mcp/test_source_registry.py`
- Direct model/store tests.
- Modify `docs/current_roadmap.md`
- Mark Slice 1 complete after implementation.
---
### Task 1: Add failing source registry model tests
**Files:**
- Create: `tests/wf_mcp/test_source_registry.py`
- [ ] **Step 1: Write model validation tests**
Create `tests/wf_mcp/test_source_registry.py`:
```python
from __future__ import annotations
import pytest
from wf_mcp.source_registry import (
HttpSourceTransport,
McpSourceRegistryEntry,
SourceRegistryFile,
StdioSourceTransport,
)
def _entry(source_id: str = "github.work") -> McpSourceRegistryEntry:
return McpSourceRegistryEntry(
id=source_id,
provider="github",
account="work",
transport=StdioSourceTransport(
command="npx",
args=("-y", "@modelcontextprotocol/server-github"),
env={"GITHUB_TOKEN": "${GITHUB_TOKEN}"},
),
auth_ref=source_id,
metadata={"purpose": "tests"},
)
def test_source_registry_entry_keeps_identity_and_transport_structural() -> None:
entry = _entry()
assert entry.id == "github.work"
assert entry.provider == "github"
assert entry.account == "work"
assert entry.profile is None
assert entry.transport.kind == "stdio"
assert entry.transport.command == "npx"
assert entry.auth_ref == "github.work"
def test_source_registry_accepts_http_transport() -> None:
entry = McpSourceRegistryEntry(
id="github.http",
provider="github",
account="work",
transport=HttpSourceTransport(url="https://example.test/mcp"),
)
assert entry.transport.kind == "http"
assert str(entry.transport.url) == "https://example.test/mcp"
def test_source_registry_rejects_duplicate_ids() -> None:
with pytest.raises(ValueError, match="duplicate source id 'github.work'"):
SourceRegistryFile(sources=[_entry("github.work"), _entry("github.work")])
def test_source_registry_rejects_reserved_ids() -> None:
with pytest.raises(ValueError, match="reserved"):
_entry("wf.admin")
def test_source_registry_rejects_unsafe_ids() -> None:
with pytest.raises(ValueError, match="connection id"):
_entry("../bad")
```
- [ ] **Step 2: Run tests and verify failure**
Run:
```bash
uv run pytest tests/wf_mcp/test_source_registry.py -q
```
Expected: FAIL because `wf_mcp.source_registry` does not exist yet.
---
### Task 2: Implement source registry models
**Files:**
- Create: `src/wf_mcp/source_registry.py`
- Test: `tests/wf_mcp/test_source_registry.py`
- [ ] **Step 1: Add models and validators**
Create `src/wf_mcp/source_registry.py`:
```python
from __future__ import annotations
import json
from collections.abc import Mapping, Sequence
from pathlib import Path
from typing import Annotated, Literal, Protocol
from pydantic import AnyHttpUrl, BaseModel, ConfigDict, Field, field_validator, model_validator
from .connections import parse_connection_id
from .shared.names import RESERVED_CONNECTION_IDS
class SourceRegistryModel(BaseModel):
"""Base model for persisted source registry state; reject misspelled fields."""
model_config = ConfigDict(extra="forbid")
class StdioSourceTransport(SourceRegistryModel):
kind: Literal["stdio"] = "stdio"
command: str = Field(min_length=1)
args: tuple[str, ...] = ()
env: dict[str, str] = Field(default_factory=dict)
class HttpSourceTransport(SourceRegistryModel):
kind: Literal["http"] = "http"
url: AnyHttpUrl
headers: dict[str, str] = Field(default_factory=dict)
SourceTransport = Annotated[
StdioSourceTransport | HttpSourceTransport,
Field(discriminator="kind"),
]
class McpSourceRegistryEntry(SourceRegistryModel):
"""Desired MCP source configuration persisted by server-owned mutation."""
id: str
kind: Literal["mcp"] = "mcp"
enabled: bool = True
provider: str = Field(min_length=1)
account: str = Field(min_length=1)
profile: str | None = None
transport: SourceTransport
auth_ref: str | None = None
metadata: dict[str, object] = Field(default_factory=dict)
@field_validator("id")
@classmethod
def validate_id(cls, value: str) -> str:
parse_connection_id(value)
if value in RESERVED_CONNECTION_IDS:
raise ValueError(f"source id {value!r} is reserved")
return value
class SourceRegistryFile(SourceRegistryModel):
version: Literal[1] = 1
sources: list[McpSourceRegistryEntry] = Field(default_factory=list)
@model_validator(mode="after")
def validate_unique_source_ids(self) -> SourceRegistryFile:
seen: set[str] = set()
for source in self.sources:
if source.id in seen:
raise ValueError(f"duplicate source id {source.id!r}")
seen.add(source.id)
return self
def source_map(self) -> dict[str, McpSourceRegistryEntry]:
return {source.id: source for source in self.sources}
class SourceRegistryStore(Protocol):
"""Persistence boundary for desired server-owned source configuration."""
def load_registry(self) -> SourceRegistryFile:
"""Return the stored registry, or an empty registry when absent."""
...
def save_registry(self, registry: SourceRegistryFile) -> None:
"""Persist one validated registry atomically."""
...
```
- [ ] **Step 2: Run model tests**
Run:
```bash
uv run pytest tests/wf_mcp/test_source_registry.py -q
```
Expected: model tests PASS except store tests are not added yet.
---
### Task 3: Add failing file store tests
**Files:**
- Modify: `tests/wf_mcp/test_source_registry.py`
- [ ] **Step 1: Append file store tests**
Append:
```python
from pathlib import Path
from wf_mcp.source_registry import FileSourceRegistryStore
def test_file_source_registry_store_loads_empty_registry_when_missing(
tmp_path: Path,
) -> None:
store = FileSourceRegistryStore(tmp_path)
registry = store.load_registry()
assert registry.version == 1
assert registry.sources == []
assert store.path == tmp_path / "source_registry.json"
def test_file_source_registry_store_round_trips_registry(tmp_path: Path) -> None:
store = FileSourceRegistryStore(tmp_path)
registry = SourceRegistryFile(sources=[_entry("github.work")])
store.save_registry(registry)
loaded = store.load_registry()
assert loaded.source_map()["github.work"].provider == "github"
assert loaded.source_map()["github.work"].transport.kind == "stdio"
def test_file_source_registry_store_validates_loaded_registry(tmp_path: Path) -> None:
store = FileSourceRegistryStore(tmp_path)
store.path.write_text(
'{"version": 1, "sources": [{"id": "wf.admin", "provider": "wf", '
'"account": "admin", "transport": {"kind": "stdio", "command": "x"}}]}',
encoding="utf-8",
)
with pytest.raises(ValueError, match="reserved"):
store.load_registry()
```
- [ ] **Step 2: Run tests and verify failure**
Run:
```bash
uv run pytest tests/wf_mcp/test_source_registry.py -q
```
Expected: FAIL because `FileSourceRegistryStore` does not exist yet.
---
### Task 4: Implement file store
**Files:**
- Modify: `src/wf_mcp/source_registry.py`
- Test: `tests/wf_mcp/test_source_registry.py`
- [ ] **Step 1: Add file store implementation**
Append to `src/wf_mcp/source_registry.py`:
```python
class FileSourceRegistryStore:
"""Filesystem implementation for desired source registry state."""
def __init__(self, root: Path) -> None:
self.root = root
self.root.mkdir(parents=True, exist_ok=True)
@property
def path(self) -> Path:
return self.root / "source_registry.json"
def load_registry(self) -> SourceRegistryFile:
if not self.path.exists():
return SourceRegistryFile()
data = json.loads(self.path.read_text(encoding="utf-8"))
return SourceRegistryFile.model_validate(data)
def save_registry(self, registry: SourceRegistryFile) -> None:
# Validate again at the store boundary so callers cannot persist stale or
# partially constructed model-like objects after mutation.
validated = SourceRegistryFile.model_validate(
registry.model_dump(mode="json")
)
payload = json.dumps(validated.model_dump(mode="json"), indent=2)
tmp_path = self.path.with_name(f"{self.path.name}.tmp")
tmp_path.write_text(payload, encoding="utf-8")
tmp_path.replace(self.path)
```
- [ ] **Step 2: Add `__all__`**
At the bottom of `src/wf_mcp/source_registry.py`, add:
```python
__all__ = [
"FileSourceRegistryStore",
"HttpSourceTransport",
"McpSourceRegistryEntry",
"SourceRegistryFile",
"SourceRegistryStore",
"SourceTransport",
"StdioSourceTransport",
]
```
- [ ] **Step 3: Run tests**
Run:
```bash
uv run pytest tests/wf_mcp/test_source_registry.py -q
```
Expected: PASS.
---
### Task 5: Documentation and verification
**Files:**
- Modify: `docs/current_roadmap.md`
- [ ] **Step 1: Update roadmap**
In `docs/current_roadmap.md`, near the store-backed source registry note, add:
```markdown
- First source registry implementation slice complete: validated registry
models plus `FileSourceRegistryStore` exist, but startup merge and mutation
commands are still deferred.
```
- [ ] **Step 2: Run verification**
Run:
```bash
uv run pytest tests/wf_mcp/test_source_registry.py -q
uv run ruff check src/wf_mcp/source_registry.py tests/wf_mcp/test_source_registry.py
uv run ruff format --check src/wf_mcp/source_registry.py tests/wf_mcp/test_source_registry.py
uv run basedpyright --level error src/wf_mcp/source_registry.py tests/wf_mcp/test_source_registry.py
```
Expected:
- pytest PASS
- ruff check PASS
- ruff format PASS
- basedpyright 0 errors
- [ ] **Step 3: Commit**
```bash
git add src/wf_mcp/source_registry.py tests/wf_mcp/test_source_registry.py docs/current_roadmap.md
git commit -m "feat: add source registry file store"
```
---
## Self-Review
- Spec coverage: Slice 1 only is covered: models, validation, file store, tests, docs.
- Placeholder scan: no TBD/TODO placeholders.
- Type consistency: `SourceRegistryFile`, `McpSourceRegistryEntry`, `SourceRegistryStore`, and `FileSourceRegistryStore` names are consistent.
- Deferred work is explicit: startup merge, runtime hydration, RPC/CLI mutation, and auth/catalog cleanup are not part of this slice.
@@ -0,0 +1,271 @@
# Store-Backed Source Registry Design
## Purpose
The workflow server now has neutral read-only workflow, source-admin, and
admin/config surfaces. The next platform step is safe mutation: adding,
updating, disabling, and removing server-owned sources/connections without
editing config files by hand.
This spec defines the persistence and merge rules before mutation commands are
implemented.
## Current State
`wf_mcp.storage.FileStore` persists:
- auth records under `auth/<connection_id>.json`
- catalog snapshots under `catalog/<connection_id>.json`
It does not persist the connection/source registry itself. Current connection
definitions come from config and live in memory through `ConnectionService`.
Neutral workflow config currently has:
- `client.target`: local or JSON-RPC HTTP target
- `server.store`: filesystem store root
- `server.transports`: server-hosted transports
- `server.sources`: static built-in sources such as `wf.std` / `wf.recipes`
Legacy MCP config still has `connections`.
## Goals
- Persist server-owned dynamic source/connection changes across process restarts.
- Keep config useful as bootstrap and deployment-time infrastructure.
- Preserve structural source identity. Source ids are explicit ids, not parsed
from dotted display names.
- Make mutation validation explicit and fail-fast.
- Keep disabled/missing sources visible as diagnostics, not silent deletion from
deployments or runs.
- Avoid turning catalog snapshots into source definitions. Catalogs are observed
capability state; registry entries are desired configuration state.
## Non-Goals
- No UI design.
- No auth-secret format redesign beyond referencing existing auth records.
- No SQL store in the first implementation.
- No automatic source id inference from provider/account strings.
- No workflow lifecycle changes.
## Registry Model
The persisted registry stores desired source/connection definitions:
```json
{
"version": 1,
"sources": [
{
"id": "github.work",
"kind": "mcp",
"enabled": true,
"provider": "github",
"account": "work",
"profile": null,
"transport": {
"kind": "stdio",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": {}
},
"auth_ref": "github.work",
"metadata": {}
}
]
}
```
### Source Identity
`id` is the stable workflow-facing concrete source id. Examples:
- `github.work`
- `github.personal`
- `everything.default`
- `wf.std`
The id is not decomposed for meaning. The structured fields carry meaning:
- `provider`: logical provider/server family, such as `github`
- `account`: account/workspace name, such as `work`
- `profile`: optional variant under one provider/account
- `transport`: concrete connectivity details
Transport belongs to the concrete source entry, not only to the provider. The
same provider/account may need different connection transports in different
server deployments, and one source id must carry the exact runtime transport the
server should use.
### Source Kinds
Initial persisted kind:
- `mcp`: upstream MCP source
Bootstrap-only / built-in kinds can remain config/code-owned for now:
- `stdlib`
- `docs`
- `admin`
Do not persist built-in `wf.std`, `wf.recipes`, or `wf.admin` in the dynamic
registry unless/until there is a real need.
## Store Shape
Add a registry store beside existing auth/catalog files:
```text
<store_root>/
auth/
catalog/
source_registry.json
```
Why one file first:
- The registry is small.
- Writes can be atomic by writing a temp file then replacing it.
- Whole-file validation is simple.
- It matches config-like semantics.
Future stores can expose the same interface over SQL or another durable backend.
## Merge Rules
On server startup:
1. Load built-in sources from code/config.
2. Load config-defined connections/sources.
3. Load `source_registry.json`.
4. Merge into one desired source map.
5. Hydrate runtime `ConnectionService` / `SourceCatalogService` from that map.
Precedence:
1. Built-in reserved ids always win for reserved ids.
2. Config-defined entries win over dynamic registry entries with the same id.
3. Dynamic registry entries fill in ids not present in config.
Rationale: config is deployment bootstrap and operator-controlled. Dynamic store
state should not secretly override source definitions checked into deployment
config.
Duplicate behavior:
- Duplicate ids inside one config or one registry file are validation errors.
- A registry entry with the same id as config is allowed but ignored with a
diagnostic/event.
- A registry entry using a reserved id is invalid.
## Mutation Rules
Initial mutation commands should target the registry only, not config:
- add source
- update source
- enable/disable source
- remove source
Rules:
- Mutations validate the full registry before saving.
- Mutations write one atomic registry replacement.
- Enabling a source requires validation of source shape and transport config.
- Optional live validation can be requested, but ordinary connection failure
should not corrupt registry state.
- Removing a source deletes desired registry state only. It does not delete old
catalog/auth files in the first pass.
- Disable is preferred over remove when deployments may still reference the
source.
## Runtime Semantics
If a deployment references a missing/disabled/unreachable source:
- validation reports diagnostics (`source_missing`, `source_disabled`,
`source_unreachable`)
- run start fails or is blocked by validation
- existing deployments are not rewritten
- ordinary dead tools/sources do not become interrupts or pauses
## Events
Registry mutations should emit broker/server events:
- `source_registered`
- `source_updated`
- `source_enabled`
- `source_disabled`
- `source_removed`
- `source_registry_ignored_config_shadow`
Events are read-only through the existing `WorkflowAdminApi`.
## API Surfaces
Do not add mutation to `WorkflowApiSurface`.
Likely surfaces:
- `WorkflowSourceAdminSurface`: read-only source list/inspect already exists
- future `WorkflowSourceRegistrySurface`: mutating registry operations
- `WorkflowAdminSurface`: read-only connections/status/events already exists
The mutation surface may live in `wf_api` if it remains protocol-neutral. If it
becomes transport/provider-heavy, split it into a platform admin package instead
of bloating workflow lifecycle APIs.
## First Implementation Slices
### Slice 1: Registry Models and File Store
- Add Pydantic registry models.
- Add `SourceRegistryStore` protocol.
- Add `FileSourceRegistryStore`.
- Validate duplicate ids and reserved ids.
- Atomic write for filesystem store.
- No runtime wiring yet.
### Slice 2: Startup Merge
- Load registry during server/broker construction.
- Merge config + registry deterministically.
- Emit diagnostics/events for ignored shadowed entries.
- Preserve existing config-only behavior when registry file is absent.
### Slice 3: Read Registry Through Admin
- Add admin read method for desired registry entries if needed.
- Keep current source inventory list as runtime/observed source inventory.
- Document difference between desired registry and observed source catalog.
### Slice 4: Mutating RPC/CLI
- Add add/update/enable/disable/remove operations.
- Add JSON-RPC methods.
- Add CLI commands.
- Validate before commit.
- Do not delete auth/catalog on remove in v1.
## Open Questions
- Should dynamic registry entries support non-MCP transports in v1, or only MCP
stdio/HTTP?
- Should auth references be required for sources that need auth, or optional
until live validation?
- Should config shadowing registry entries be a warning only, or should server
startup fail in strict mode?
- Should disabled registry entries still hydrate as disabled sources so inspect
can explain them, or stay only in registry/admin output?
## Recommendation
Implement Slice 1 first. Keep it independent from runtime startup so the model
and file persistence rules become solid before they affect source hydration.
The existing `wf_mcp` store can host the first file-backed implementation, but
the interface should be neutral enough to move later. Catalog/auth remain
observed/secret state; the new registry is desired source configuration state.