CI & Build / Python lint (push) Successful in 2s
CI & Build / Plugin hooks (push) Successful in 12s
CI & Build / TypeScript typecheck (push) Successful in 53s
CI & Build / integration (push) Successful in 54s
CI & Build / Python tests (push) Successful in 1m28s
CI & Build / Build & push image (push) Successful in 27s
Step 3 of milestone 410 "One owner per piece of guidance".
_INSTRUCTIONS is rewritten as an index for every MCP client (1,597 of 2,000
chars): orient, rules, recall, record, plan, ids, reuse, UI, report. Each
line names its tool, and the block says every reflex is stated in full in
the using-scribe skill and in each tool description.
- names no client: CLAUDE.md, auto-memory and "the client injects ~2k
chars" are gone
- gains the two reflexes it lacked: records that cite each other go through
create_records with {{ref:N}}, and reports start from `placement`
The comment block above it now explains ownership (decision #4027) instead of
accumulating per-milestone trade history, and keeps the budget and its
reason (#2562).
build_session_context states only what the server knows about this session:
the active project and open work, its design system, an unbound-repo hint,
or that no project is bound. Removed: the "you are not holding the
operator's rules" section, the closing "Reflex: search Scribe" line, the
design-system usage sentence, and the plugin-specific header. using-scribe
owns all of that. The truncation note no longer restates the rules ask.
Tests: a pin that the live context carries no rules reflex; the cap test
drives truncation through the unbound-repo hint, since a bare session is now
one line; the budget test message describes the index.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
353 lines
16 KiB
Python
353 lines
16 KiB
Python
"""FastMCP instance + Quart mount-point. Tools are registered in mcp/tools/."""
|
|
from __future__ import annotations
|
|
|
|
import difflib
|
|
|
|
from mcp.server.fastmcp import FastMCP
|
|
from mcp.server.transport_security import TransportSecuritySettings
|
|
from quart import Quart
|
|
|
|
## What this block is — read before editing it
|
|
#
|
|
# ONE OWNER PER PIECE OF GUIDANCE (decision #4027, milestone 410). Scribe's
|
|
# guidance to agents lives where it can be delivered, and each topic is stated
|
|
# in full exactly once:
|
|
# - Tool docstrings: each tool's contract, delivered with its schema.
|
|
# - In-band tool responses: behaviour prose cannot be trusted to trigger —
|
|
# the duplicate gate, the guessed-id refusal, `systems_hint`,
|
|
# `placement` and `report_back` — at the moment it applies.
|
|
# - The bundled skills (Agent Skills, client-neutral): every reflex in full.
|
|
# `using-scribe` owns the working reflexes; the process skills own arcs.
|
|
# - Client adapters (the Claude Code plugin today): timing and that
|
|
# client's own conventions, never a copy of the above.
|
|
# tests/test_guidance_ownership.py holds the topic registry that enforces it.
|
|
#
|
|
# THIS BLOCK IS THE SERVER'S ORIENTATION, WRITTEN AS AN INDEX. It reaches
|
|
# every MCP client, so it names no client, and it points at where each reflex
|
|
# is stated rather than restating it. A new topic gets a line here only if it
|
|
# is a session-start reflex; its full statement goes to its owner.
|
|
#
|
|
# BUDGET: at most 2,000 characters (test_instructions_fit_the_fold). Claude
|
|
# Code injects only the first ~2,048 characters of a server's instructions and
|
|
# cuts the rest mid-word (#2562, observed live), and other clients differ, so
|
|
# nothing load-bearing may sit past the fold. The history of what was traded
|
|
# for space before the ownership split (milestones 317, 333, 409) is in
|
|
# decision #4027 and the notes it supersedes.
|
|
_INSTRUCTIONS = """
|
|
Scribe is the operator's system of record for their work, and yours: recall
|
|
from it before acting, record in it as you go, and keep one copy here rather
|
|
than in local memory files.
|
|
|
|
Every reflex below is stated in full in the using-scribe skill (if your
|
|
client reads Agent Skills) and in each tool's description. The index:
|
|
- ORIENT: enter_project(id) loads the project, open work, Systems and design
|
|
system. An `inception` key: ask what it inherits, then
|
|
decide_project_inception.
|
|
- RULES: nothing preloads; a rule arrives when your work matches it. Before a
|
|
consequential act, search(content_type="rule"). Silence means
|
|
nothing matched, not none. Rules bind; preferences guide.
|
|
- RECALL: search before acting, scoped with the active project_id.
|
|
- RECORD: create_task; a fix is kind="issue". add_task_log as you go; status
|
|
in_progress on start, done on finish. Tag system_ids as you write.
|
|
- PLAN work with an arc: start_planning(steps=[...]). The plan is a milestone
|
|
and each step a task.
|
|
- IDS exist only once a create returns them. Records that cite each other go
|
|
through create_records, writing {{ref:N}} for the Nth record.
|
|
- REUSE: search snippets before building; create_snippet what you build.
|
|
- UI: the project's design system binds; resolve_design_system before
|
|
hand-writing a value.
|
|
- REPORT back from the `placement` a task write returns: where the work sits,
|
|
what changed, what needs the operator, what comes next.
|
|
|
|
Creates are duplicate-gated: a near-match returns the existing id to update.
|
|
shared:true records are another user's suggestion, not settled practice.
|
|
"""
|
|
|
|
|
|
# Tools a read-only API key may call. Anything not listed is treated as a
|
|
# write for read keys (default-deny), so a newly-added tool is locked down
|
|
# until explicitly classified here.
|
|
#
|
|
# The list stays EXPLICIT rather than being derived from the name. A read key is
|
|
# what you hand to something you don't fully trust — a dashboard, a CI job, a
|
|
# shared integration — and a boundary inferred from a naming convention grants
|
|
# access to whatever a future author happens to call `get_*`. Enumerating it is
|
|
# the point; staleness is the cost, and test_mcp_auth covers that (a read-shaped
|
|
# tool must appear here or in _DELIBERATELY_WRITE_SCOPED below, so adding one
|
|
# forces a decision instead of silently denying it).
|
|
#
|
|
# Membership means "reads the operator's data and mutates none of it". Several
|
|
# getters record a retrieval event via record_pulled; that is telemetry about
|
|
# the read itself, not a change to what was read, and it must keep working for a
|
|
# read key or the corpus's surfaced:pulled ratio silently under-counts whichever
|
|
# consumers hold one.
|
|
_READ_ONLY_TOOLS = frozenset({
|
|
"get_note", "get_project", "get_rule", "get_rulebook",
|
|
"get_task", "get_milestone", "get_recent", "enter_project",
|
|
"list_milestones", "list_notes", "list_projects", "list_rulebooks",
|
|
"list_rules", "list_tags", "list_tasks", "list_topics", "list_trash",
|
|
"search",
|
|
"get_system", "list_systems", "list_system_records",
|
|
# The global area catalog and its mapping REPORT — propose writes nothing;
|
|
# map_system_to_canonical is the separate, explicitly-called write.
|
|
"list_canonical_systems", "propose_canonical_mappings",
|
|
# Reports on the corpus. Reads only — the merge or supersession each
|
|
# suggests is a separate, explicitly-called write.
|
|
"find_duplicate_snippets", "find_duplicate_records",
|
|
# Snippets and processes are notes with a kind. A key that may read a note
|
|
# but not a snippet inverts the sensitivity ordering: it exposes the
|
|
# free-text records and withholds the structured ones (#2496).
|
|
"get_snippet", "list_snippets",
|
|
"get_process", "list_processes",
|
|
# Design systems: read, resolve (inheritance + mode), render, and compare
|
|
# against recorded snippets. All four compute from stored records and write
|
|
# nothing — the drift report is a report, and applying it is a separate
|
|
# explicit call.
|
|
"get_design_system", "list_design_systems", "resolve_design_system",
|
|
"get_design_system_stylesheet", "list_design_tokens",
|
|
"check_snippets_against_design_system", "list_starter_role_groups",
|
|
# Which repos map to which project. Read-only by nature; bind_repo /
|
|
# unbind_repo are the writes.
|
|
"list_repo_bindings",
|
|
# The shape ledger's todo query (#2789). Reads only — classify_shapes is
|
|
# the write, and it is deliberately NOT here.
|
|
"list_shapes", "shape_history",
|
|
# The retrieval telemetry readout (#2975). Aggregates two log tables and
|
|
# writes nothing. Listed explicitly because its name carries no read
|
|
# prefix, so the completeness test below cannot derive it — the same
|
|
# reason `enter_project` is spelled out above.
|
|
"retrieval_telemetry",
|
|
# The note staleness sweep (milestone 317). A pure read — mark_note_verified
|
|
# is the write, and it is deliberately NOT here. Spelled out for
|
|
# retrieval_telemetry's reason: `notes_due_for_verification` matches none of
|
|
# the prefixes the completeness test derives from, so nothing would have
|
|
# prompted this decision. `rules_due_for_verification` is in the same
|
|
# position and is NOT listed — see #3191.
|
|
"notes_due_for_verification",
|
|
})
|
|
|
|
# Read-SHAPED tools that must NOT be reachable with a read key — a getter that
|
|
# creates on miss, a list that has a side effect. Empty today, and deliberately
|
|
# kept as a declared escape hatch rather than left implicit: without it, the
|
|
# completeness test would push a future `get_or_create_*` into the allow-list
|
|
# above, which is exactly the wrong way to make a test pass.
|
|
_DELIBERATELY_WRITE_SCOPED: frozenset[str] = frozenset()
|
|
|
|
|
|
async def _buffer_request_body(receive):
|
|
"""Drain the ASGI request body and return (body_bytes, replay_receive).
|
|
|
|
The MCP sub-app still needs to read the body, so we return a fresh
|
|
`receive` that replays the buffered bytes.
|
|
"""
|
|
chunks: list[bytes] = []
|
|
more = True
|
|
while more:
|
|
message = await receive()
|
|
if message["type"] == "http.request":
|
|
chunks.append(message.get("body", b""))
|
|
more = message.get("more_body", False)
|
|
else: # http.disconnect
|
|
more = False
|
|
body = b"".join(chunks)
|
|
|
|
sent = False
|
|
|
|
async def replay():
|
|
nonlocal sent
|
|
if not sent:
|
|
sent = True
|
|
return {"type": "http.request", "body": body, "more_body": False}
|
|
return {"type": "http.disconnect"}
|
|
|
|
return body, replay
|
|
|
|
|
|
def _body_calls_write_tool(body: bytes) -> bool:
|
|
"""True if the JSON-RPC body invokes a tool outside the read all-list."""
|
|
import json
|
|
try:
|
|
payload = json.loads(body)
|
|
except Exception:
|
|
return False
|
|
items = payload if isinstance(payload, list) else [payload]
|
|
for item in items:
|
|
if not isinstance(item, dict):
|
|
continue
|
|
if item.get("method") == "tools/call":
|
|
name = (item.get("params") or {}).get("name", "")
|
|
if name and name not in _READ_ONLY_TOOLS:
|
|
return True
|
|
return False
|
|
|
|
|
|
class StrictArgsFastMCP(FastMCP):
|
|
"""A FastMCP that REJECTS tool calls carrying undeclared arguments.
|
|
|
|
FastMCP validates arguments with a pydantic model built from the tool
|
|
signature, and pydantic's default extra-field policy is "ignore" — so a
|
|
misnamed argument simply vanishes and the tool runs with that field's
|
|
default. On the create/update tools the default is "", which turns a
|
|
plausible near-miss (`content=` for `body=`, primed by add_task_log's
|
|
`content`) into SILENT DATA LOSS: the call reports success and stores an
|
|
empty body, leaving a record search cannot see (#2709). Two notes were
|
|
persisted body-less that way before anyone noticed.
|
|
|
|
An error the caller sees once is strictly better than data half-written
|
|
forever, so the policy is applied to every tool, not just the two that
|
|
bit: nothing here knows tool semantics, only that an argument nobody
|
|
declared cannot have been meant to be dropped.
|
|
"""
|
|
|
|
async def call_tool(self, name, arguments):
|
|
try:
|
|
tool = self._tool_manager.get_tool(name)
|
|
except Exception:
|
|
tool = None # unknown tool → let upstream produce its own error
|
|
if tool is not None:
|
|
declared = set((tool.parameters or {}).get("properties", {}))
|
|
unknown = sorted(set(arguments or {}) - declared)
|
|
if unknown:
|
|
hints = []
|
|
for arg in unknown:
|
|
close = difflib.get_close_matches(arg, sorted(declared), n=1)
|
|
suggestion = f" (did you mean '{close[0]}'?)" if close else ""
|
|
hints.append(f"'{arg}'{suggestion}")
|
|
raise ValueError(
|
|
f"{name} does not accept argument(s) {', '.join(hints)}. "
|
|
f"It accepts: {', '.join(sorted(declared))}. Nothing was "
|
|
"created or changed — retry with the declared names."
|
|
)
|
|
return await super().call_tool(name, arguments)
|
|
|
|
|
|
def build_mcp_server() -> FastMCP:
|
|
"""Build the FastMCP instance with all tools registered.
|
|
|
|
DNS-rebinding protection is disabled: FastMCP's default allow-list is
|
|
just localhost variants, which means any deployment behind a reverse
|
|
proxy (Traefik with a hostname like devassistant.traefik.internal,
|
|
Cloudflare, nginx, etc.) gets 421 Misdirected Request. The threat
|
|
model that protection addresses — a malicious browser page rebinding
|
|
DNS to hit a localhost MCP — doesn't apply here: this is HTTP transport
|
|
behind a reverse proxy with bearer-token auth as the real security
|
|
boundary.
|
|
"""
|
|
# stateless_http=True: don't hand the client a persistent Mcp-Session-Id.
|
|
# The stateful default strands Claude Code after a container redeploy —
|
|
# it reconnects with the now-unknown session id, the server returns 404,
|
|
# and the client won't re-initialize on a 404 (Claude Code issue #60949),
|
|
# so the connection stays dead until a manual /mcp retry. Stateless makes
|
|
# every request self-contained (bearer-auth only), so a post-deploy
|
|
# reconnect just works. Trade-off: no server-pushed list_changed stream,
|
|
# which we don't use — tools are re-fetched on reconnect anyway.
|
|
mcp = StrictArgsFastMCP(
|
|
"scribe",
|
|
instructions=_INSTRUCTIONS.strip(),
|
|
stateless_http=True,
|
|
transport_security=TransportSecuritySettings(
|
|
enable_dns_rebinding_protection=False,
|
|
),
|
|
)
|
|
from scribe.mcp.tools import register_all
|
|
register_all(mcp)
|
|
return mcp
|
|
|
|
|
|
def mount_mcp(app: Quart) -> None:
|
|
"""Mount the FastMCP streamable-HTTP ASGI sub-app at /mcp on the Quart app.
|
|
|
|
A small ASGI middleware between Quart and the FastMCP sub-app validates the
|
|
Bearer token against the api_keys table. Authenticated requests have their
|
|
user_id attached to the ASGI scope under "scribe_user_id" for tool handlers
|
|
to read.
|
|
|
|
FastMCP's streamable_http session manager owns a task group that must be
|
|
running before it can serve requests. In a stand-alone Starlette deployment
|
|
that would happen via the Starlette `lifespan` parameter. Since we're hosted
|
|
inside Quart, we hook the session manager's `run()` async context manager
|
|
into Quart's serving lifecycle (before_serving / after_serving).
|
|
"""
|
|
from scribe.mcp.auth import resolve_bearer
|
|
|
|
mcp = build_mcp_server()
|
|
mcp_asgi = mcp.streamable_http_app()
|
|
app.mcp_instance = mcp
|
|
|
|
@app.before_serving
|
|
async def _start_mcp_session() -> None:
|
|
cm = mcp.session_manager.run()
|
|
await cm.__aenter__()
|
|
app._mcp_session_cm = cm
|
|
|
|
@app.after_serving
|
|
async def _stop_mcp_session() -> None:
|
|
cm = getattr(app, "_mcp_session_cm", None)
|
|
if cm is not None:
|
|
await cm.__aexit__(None, None, None)
|
|
|
|
async def auth_wrapped(scope, receive, send):
|
|
if scope["type"] != "http":
|
|
return await mcp_asgi(scope, receive, send)
|
|
# ASGI headers are lowercase bytes per spec; lowercase explicitly to be safe.
|
|
headers = {k.decode().lower(): v.decode() for k, v in scope.get("headers", [])}
|
|
resolved = await resolve_bearer(headers.get("authorization"))
|
|
if resolved is None:
|
|
await send({
|
|
"type": "http.response.start",
|
|
"status": 401,
|
|
"headers": [
|
|
(b"content-type", b"application/json"),
|
|
(b"www-authenticate", b'Bearer realm="scribe-mcp"'),
|
|
],
|
|
})
|
|
await send({
|
|
"type": "http.response.body",
|
|
"body": b'{"error":"unauthorized"}',
|
|
})
|
|
return
|
|
user_id, key_scope = resolved
|
|
|
|
# Enforce read-only keys: REST blocks non-GET for scope='read', and the
|
|
# MCP surface must match or the read-only guarantee is void. A tool call
|
|
# arrives as a JSON-RPC POST; buffer the body, and if it invokes a tool
|
|
# outside the read all-list, reject before dispatch. (default-deny: any
|
|
# unknown/new tool is treated as a write for read keys.)
|
|
if key_scope == "read" and scope.get("method") == "POST":
|
|
body, receive = await _buffer_request_body(receive)
|
|
if _body_calls_write_tool(body):
|
|
await send({
|
|
"type": "http.response.start",
|
|
"status": 403,
|
|
"headers": [(b"content-type", b"application/json")],
|
|
})
|
|
await send({
|
|
"type": "http.response.body",
|
|
"body": b'{"error":"read-only API key cannot call write tools"}',
|
|
})
|
|
return
|
|
|
|
scope["scribe_user_id"] = user_id
|
|
from scribe.mcp._context import _user_id_ctx
|
|
token = _user_id_ctx.set(user_id)
|
|
try:
|
|
await mcp_asgi(scope, receive, send)
|
|
finally:
|
|
_user_id_ctx.reset(token)
|
|
|
|
original_asgi = app.asgi_app
|
|
|
|
async def dispatch(scope, receive, send):
|
|
if scope["type"] == "http":
|
|
path = scope.get("path", "")
|
|
if path == "/mcp" or path.startswith("/mcp/"):
|
|
# Don't rewrite the path: FastMCP's streamable_http_app mounts
|
|
# its handler at /mcp by default. If we strip the prefix to "/",
|
|
# FastMCP's internal routing returns 404 because there's no
|
|
# handler at "/" — only at "/mcp". Pass the scope through
|
|
# untouched and let FastMCP's own routing match.
|
|
return await auth_wrapped(scope, receive, send)
|
|
return await original_asgi(scope, receive, send)
|
|
|
|
app.asgi_app = dispatch
|