FabledScribe/src/scribe/mcp/server.py

"""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

## The delivery budget — read before editing this block
#
# Claude Code injects only the FIRST ~2,048 CHARACTERS of an MCP server's
# instructions into the system prompt; the rest is silently cut mid-word
# (#2562 — the cut was observed live at exactly offset 2,048, and ~90% of the
# previous 20k-char version of this block never reached any session). So this
# block is deliberately a MAP, not a manual, and a test pins it under the
# fold (test_instruction_surfaces_agree.py::test_instructions_fit_the_fold).
#
# Where the detail lives instead — each surface has one job:
#   - Tool docstrings: the per-tool HOW. Delivered with the tool schema, at
#     reach-for time when the client defers tools. Guidance about one tool
#     belongs there, not here.
#   - Plugin static context (plugin/hooks/scribe_static_context.md): the
#     session-level reflexes (recall-first, record-as-you-go, tag-to-Systems,
#     compaction). Always delivered in full; needs no key and no network.
#   - Plugin skills: process arcs (planning, debugging, verification…).
#     Their listing line is the always-visible trigger; the body loads on
#     match. Stored Processes become skills via /scribe:sync.
#   - The server itself: behaviors prose can't be trusted to fire (the
#     duplicate gate, the untagged-record systems_hint) act in-band in tool
#     responses, at the moment they apply.
# Grow one of those, not this block.
# BUDGET: ~1980 of the client's ~2048-char cap (#2562). Everything below is
# competing for the last ~68 characters, so an addition here is a trade, never
# an append.
#
# Milestone 317 (a note's own verify_with / expires_when, and the sweep over
# them) was DECLINED a line, deliberately, by the operator — not overlooked.
# The reasoning, so it is not re-litigated blind: this is a map, and its own
# closing line says each tool's description carries the full contract. The
# sweep is a curation act, not a session-start reflex like enter_project or
# list_always_on_rules. Spending the last of the budget on it would leave the
# map unable to grow for something more central later.
#
# The accepted cost: an agent that never opens create_note's docstring never
# learns the field exists. Guidance lives in the create_note / update_note
# docstrings and the using-scribe skill instead.
#
# Milestone 333 step 3 (2026-09-04) bought the HOW bullet's second clause —
# search(content_type="rule") before a consequential act — by TRADING OUT
# "Processes are saved procedures (follow verbatim)" and "Deletes are
# trash-recoverable". Recorded so the trade is not silently reversed:
#   - Both were already in test_instruction_surfaces_agree's DISPLACED_TOPICS
#     and already stated on a delivered surface, so nothing fell off: the
#     process reflex is in every scribe-proc-* skill listing (each says the
#     process governs and is followed verbatim), and trash recovery is in the
#     delete_*/list_trash/restore docstrings, which is where per-tool guidance
#     belongs by this block's own doctrine.
#   - What it bought is not per-tool guidance and has nowhere else to live at
#     session-start altitude. Rules were retrievable only by RESIDENCY: the
#     always-on preload put them in front of the agent, and nothing told a
#     session to go looking for one it had not been handed. The tier split is
#     therefore load-bearing on ANY install (rule 115): a delivered rule costs
#     tokens in every session forever, so a rulebook that only delivers cannot
#     grow past what one session can hold, and every rule worth keeping has to
#     become resident to bind at all. Retrieval is what lets it keep growing —
#     and retrieval fires only if something asks, which nothing told a session
#     to do. A tool-choice reflex asks least of all (#3476, #161).
#   - This states the PULL for conditional rules, exactly as the surrounding
#     line states it for always-on ones. Rule 119 makes these surfaces the
#     specification, so the same sentence lands on all three session-start
#     surfaces, and test_instruction_surfaces_agree pins it.
_INSTRUCTIONS = """
Scribe is the operator's self-hosted second brain and system of record — and
yours: recall from it before acting, record as you go. Keep no parallel copy
in local files (CLAUDE.md, auto-memory); Scribe holds the single copy.

Hierarchy: Project -> Milestone -> Task/Note. The map, by purpose:
- ORIENT: enter_project(id) at session start — rules, open tasks, recent
  notes, Systems, design system. `inception` key: ask what the project
  inherits, decide_project_inception (create_project takes the same).
- DO: create_task. Fixed a problem? kind="issue" (symptom -> root cause ->
  fix), never a work-log line on an unrelated task. Log with add_task_log;
  keep status honest — in_progress on start, done on finish.
- PLAN work with an arc: start_planning. The plan IS a milestone; each step is
  a child task, not a checkbox. No local plan .md files.
- CAPTURE: create_note. RECALL: search first — prior art exists; pass the
  active project_id to stay in scope.
- WHERE work happens: Systems. Tag records with system_ids as you write;
  create_system when the area is unmodelled.
- HOW: rules bind; preferences guide. list_always_on_rules() at start;
  before a consequential act, search(content_type="rule") — the resident
  set is not all of them.
- UI: the project's design system is binding — resolve_design_system /
  get_design_system_stylesheet before hand-writing a value.
- REUSE: search snippets before writing a helper; record what you build with
  create_snippet; classify shapes against canon (classify_shapes) — a
  consumer map is rows, never prose.

A task is a note with status (*_note vs *_task tools).
Creates are duplicate-gated: a near-match BLOCKS and returns the existing
id — update it, don't force. shared:true records are another user's — a
suggestion, not the operator's settled practice.

This is only a map — the client injects ~2k chars and cuts the rest. Each
tool's description carries its full contract: read it when you load the
tool, and trust it over habit.
"""


# 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",
    "list_always_on_rules", "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