tests: dual-protocol acceptance, covering the transport the suite cannot see

tests/konformitaet_2026_07_28.py speaks HTTP only. Review finding #1 lived on
stdio, so 27/27 green and the defect coexisted happily: on stdio `initialize`
answered -32601 Method not found, which is the first message any client sends.

This test crosses both protocols with both transports and echoes "hello world"
through the shell tool in each of the four cells. It also carries its own
controls, because a suite that cannot go red measures nothing:

  R1  an invented version (1999-01-01) must be REJECTED, not served.
  R2  running it against the code from BEFORE the fix must FAIL. That is the
      echo guard, and it was exercised: against 4ac9296 exactly one cell goes
      red -- 3a stdio/alt initialize, -32601 -- while all four HTTP cells stay
      green. Against 175244a all eight pass.

The two protocols are deliberately NOT tested symmetrically, because they are
not symmetric: 2026-07-28 has no `initialize` (SEP-2575 removed it and the new
path drops it to -32601 on purpose), and the envelope is enforced in the HTTP
dispatcher only, so over stdio the version is announced through _meta alone. A
test that sent the same flow to both surfaces would pass for the wrong reason.

Run: python3 tests/dual_protocol_acceptance.py <lmcp-dir> [port]
Measured 2026-08-11 in a throwaway Incus container on boltzmann (Debian 13,
lua5.4 + lua-socket), the sources checked by sha256 on both ends before the run.
A finer point than the review had: `tools/call` was never version-gated, so it
worked on stdio even before the fix. Only `initialize` was broken.

.gitignore

@@ -0,0 +1,13 @@
+# Generated by windows/sync.sh — see windows/README.md
+windows/pkg/lmcp.lua
+windows/pkg/server.lua
+windows/pkg/json.lua
+
+# Bundled Lua + LuaSocket runtime for the Windows MSI; downloaded
+# separately, not in git.
+windows/pkg/lua/
+
+# Editor / OS noise
+*.swp
+*.swo
+.DS_Store

envelope.lua

@@ -0,0 +1,54 @@
+local M = {}
+function M.check(header, body)
+    if not header or not body then
+        return false, -32600, "Invalid Request"
+    end
+
+    -- Rule 1: header['mcp-protocol-version'] missing
+    if not header['mcp-protocol-version'] then
+        return false, -32020, 'Missing required header: MCP-Protocol-Version'
+    end
+
+    -- Rule 2: header['mcp-method'] missing
+    if not header['mcp-method'] then
+        return false, -32020, 'Missing required header: Mcp-Method'
+    end
+
+    -- Rule 2: header['mcp-method'] ~= body.method
+    if header['mcp-method'] ~= body.method then
+        return false, -32020, 'Mcp-Method header does not match body'
+    end
+
+    -- Rule 3: Mcp-Name header required for tools/call, resources/read, prompts/get
+    local method = body.method
+    local params = body.params or {}
+
+    if method == 'tools/call' or method == 'prompts/get' then
+        if not header['mcp-name'] or header['mcp-name'] ~= params.name then
+            return false, -32020, 'Mcp-Name header does not match body'
+        end
+    elseif method == 'resources/read' then
+        if not header['mcp-name'] or header['mcp-name'] ~= params.uri then
+            return false, -32020, 'Mcp-Name header does not match body'
+        end
+    end
+
+    -- Rule 4: _meta required fields (located under params._meta)
+    local meta = (body.params or {})._meta or {}
+    if not meta['io.modelcontextprotocol/protocolVersion'] then
+        return false, -32602, 'Missing required _meta field: io.modelcontextprotocol/protocolVersion'
+    end
+    if not meta['io.modelcontextprotocol/clientCapabilities'] then
+        return false, -32602, 'Missing required _meta field: io.modelcontextprotocol/clientCapabilities'
+    end
+
+    -- Rule 5: header['mcp-protocol-version'] ~= meta['io.modelcontextprotocol/protocolVersion']
+    if header['mcp-protocol-version'] ~= meta['io.modelcontextprotocol/protocolVersion'] then
+        return false, -32020, 'Header MCP-Protocol-Version does not match _meta protocolVersion'
+    end
+
+    -- Success
+    return true, nil, nil
+end
+
+return M

helpers/stash_recall.py

@@ -0,0 +1,29 @@
+import urllib.request, json, threading, queue, sys
+BASE="http://192.168.88.184:8080"
+q=queue.Queue()
+def sse():
+    try:
+        r=urllib.request.urlopen(BASE+"/sse", timeout=30)
+        for raw in r:
+            s=raw.decode(errors="replace").strip()
+            if s.startswith("data:"): q.put(s[5:].strip())
+    except Exception as e: q.put("ERR:"+str(e))
+threading.Thread(target=sse,daemon=True).start()
+try:
+    ep=q.get(timeout=10)
+    if ep.startswith("ERR:"): print("stash unreachable:",ep); sys.exit(1)
+    purl=BASE+ep if ep.startswith("/") else ep
+    def post(o):
+        urllib.request.urlopen(urllib.request.Request(purl,data=json.dumps(o).encode(),headers={"Content-Type":"application/json"}),timeout=15).read()
+    post({"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"apropos","version":"1"}}})
+    q.get(timeout=10)
+    post({"jsonrpc":"2.0","method":"notifications/initialized"})
+    query=sys.argv[1] if len(sys.argv)>1 else ""
+    limit=int(sys.argv[2]) if len(sys.argv)>2 else 3
+    post({"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"recall","arguments":{"query":query,"limit":limit}}})
+    d=json.loads(q.get(timeout=25))
+    txt=d.get("result",{}).get("content",[{}])[0].get("text","[]")
+    facts=json.loads(txt)
+    if not facts: print("(no memory found for: %s)"%query); sys.exit(0)
+    for f in facts: print("- %s (score %.2f)" % (f.get("content","").strip(), f.get("score",0)))
+except Exception as e: print("recall error:",e); sys.exit(1)

hub.lua

@@ -29,6 +29,8 @@ local PROBE_TTL_UP = tonumber(os.getenv("LMCP_HUB_PROBE_TTL_UP") or "30")
 local PROBE_TTL_DOWN_MIN = tonumber(os.getenv("LMCP_HUB_PROBE_TTL_DOWN_MIN") or "60")
 local PROBE_TTL_DOWN_MAX = tonumber(os.getenv("LMCP_HUB_PROBE_TTL_DOWN_MAX") or "900")
 local PROBE_BUDGET = tonumber(os.getenv("LMCP_HUB_PROBE_BUDGET") or "3")
+-- TCP port that answers "is this host there" for ssh-only backends.
+local SSH_PROBE_PORT = os.getenv("LMCP_HUB_SSH_PORT") or "22"
 local LMCP_TIMEOUT = tonumber(os.getenv("LMCP_HUB_LMCP_TIMEOUT") or "6")
 local SSH_TIMEOUT = tonumber(os.getenv("LMCP_HUB_SSH_TIMEOUT") or "10")
 local SSH_HARD_TIMEOUT = tonumber(os.getenv("LMCP_HUB_SSH_HARD_TIMEOUT") or "30")
@@ -272,13 +274,19 @@ end
 -- bash fan-out of curl calls. Total wall clock ≈ PROBE_BUDGET.
 local function probe_all_parallel(force)
     local now = os.time()
-    local need = {}
+    local need, need_ssh = {}, {}
     for name, b in pairs(backends) do
-        if b.lmcp_url and (force or not cache_fresh(status[name], now)) then
-            need[#need+1] = b
+        if force or not cache_fresh(status[name], now) then
+            if b.lmcp_url then
+                need[#need+1] = b
+            elseif b.ssh_host then
+                -- ssh-only: no lmcp endpoint to ask, but "is the box there" is still
+                -- answerable cheaply. See the SSH probe note below.
+                need_ssh[#need_ssh+1] = b
+            end
         end
     end
-    if #need == 0 then return end
+    if #need == 0 and #need_ssh == 0 then return end
 
     local script_parts = {}
     for _, b in ipairs(need) do
@@ -289,6 +297,21 @@ local function probe_all_parallel(force)
             PROBE_BUDGET, b.name, auth, url, b.name
         )
     end
+    -- SSH probe. The design note above rejects checking ssh because a session costs
+    -- 3-6s per offline host — true for a SESSION. A bare TCP connect to 22 answers the
+    -- only question a host card asks ("is it there") with no handshake and no auth:
+    -- measured on this host, a dead target costs 1.05s and a live one milliseconds, and
+    -- it rides the same parallel fan-out, so wall clock stays one budget window.
+    -- Without nc we report nothing rather than guessing DOWN — a wrong claim is worse
+    -- than the "no probe result" the dashboard already renders as unknown.
+    for _, b in ipairs(need_ssh) do
+        local host = b.ssh_host:gsub("'", "'\\''")
+        script_parts[#script_parts+1] = string.format(
+            "(if command -v nc >/dev/null 2>&1; then " ..
+            "nc -z -w%d '%s' %s >/dev/null 2>&1 && echo '%s SSHUP 0' || echo '%s SSHDOWN 0'; " ..
+            "else echo '%s SSHSKIP 0'; fi) &",
+            PROBE_BUDGET, host, SSH_PROBE_PORT, b.name, b.name, b.name)
+    end
     script_parts[#script_parts+1] = "wait"
 
     local t0 = monotonic()
@@ -302,8 +325,14 @@ local function probe_all_parallel(force)
         local name, code, t = line:match("^(%S+)%s+(%S+)%s+([%d%.]+)")
         if name then
             seen[name] = true
-            local is_up = (code == "200")
-            if is_up then
+            if code == "SSHUP" then
+                apply_probe_result(name, true, nil, "ssh", nil)
+            elseif code == "SSHDOWN" then
+                apply_probe_result(name, false, "ssh port unreachable", nil, nil)
+            elseif code == "SSHSKIP" then
+                -- nc missing: leave it unprobed rather than assert a state.
+                seen[name] = nil
+            elseif code == "200" then
                 apply_probe_result(name, true, nil, "lmcp", nil)
             else
                 apply_probe_result(name, false, "lmcp code=" .. code, nil, nil)
@@ -316,7 +345,7 @@ local function probe_all_parallel(force)
             apply_probe_result(b.name, false, "probe fan-out missing", nil, nil)
         end
     end
-    logreq("probe_all_parallel n=%d elapsed=%.2fs", #need, dt)
+    logreq("probe_all_parallel lmcp=%d ssh=%d elapsed=%.2fs", #need, #need_ssh, dt)
 end
 
 -- ---- Call-tool dispatcher ----------------------------------------------
@@ -631,6 +660,62 @@ server:tool("remote_search_files", "find-by-pattern on a fleet host.",
     } }
 )
 
+-- ---- Fleet-central local tools (migrated from tools.d/hertz.lua, 2026-07-18) ----
+-- These run LOCALLY on hertz (where the hub process also lives), so no ssh
+-- backend hop. Consolidated here so the hub is the single fleet-management
+-- endpoint. Still also served by hertz-tools (:8080) for now — remove there
+-- once every client (pi-agents etc.) has a @hub session.
+local function run_local(cmd, timeout)
+    local full = timeout and ("timeout " .. tostring(timeout) .. " " .. cmd) or cmd
+    local p = io.popen(full .. " 2>&1")
+    if not p then return "Error: popen failed" end
+    local out = p:read("*a")
+    p:close()
+    return out or ""
+end
+
+server:tool("apropos",
+    "Search shared fleet memory (stash) for facts about the fleet, projects, decisions, and preferences. query = 2-6 words on the topic; limit = max results (default 3). Read-only.",
+    { type = "object", properties = {
+        query = { type = "string",  description = "2-6 words describing what to recall" },
+        limit = { type = "integer", description = "max results, default 3" },
+    }, required = { "query" } },
+    function(a)
+        local q = tostring(a.query or ""):gsub("[^%w%s%-%.]", " "):gsub("%s+", " ")
+        if q:gsub("%s", "") == "" then return "Error: query required" end
+        local lim = tonumber(a.limit) or 3
+        return run_local("python3 /opt/lmcp/helpers/stash_recall.py '" .. q .. "' " .. lim, 30)
+    end,
+    { annotations = {
+        title           = "Apropos (fleet memory)",
+        readOnlyHint    = true,
+        destructiveHint = false,
+        idempotentHint  = true,
+        openWorldHint   = true,
+    } }
+)
+
+server:tool("wake_fleet",
+    "Wake a fleet NUC (pve1..pve4) via Fritz!Box Wake-on-LAN. Powers a node ON only; it cannot power anything off. Node boots in ~30-60s.",
+    { type = "object", properties = {
+        node = { type = "string", description = "Node to wake: '1'..'4' or 'pve1'..'pve4'" },
+    }, required = { "node" } },
+    function(a)
+        local node = tostring(a.node or ""):gsub("[^%w]", "")
+        if not node:match("^p?v?e?[1-4]$") then
+            return "Error: node must be 1-4 or pve1-pve4 (got: " .. tostring(a.node) .. ")"
+        end
+        return run_local("sudo /root/.local/bin/wake-pve " .. node, 15)
+    end,
+    { annotations = {
+        title           = "Wake fleet NUC",
+        readOnlyHint    = false,
+        destructiveHint = false,
+        idempotentHint  = true,
+        openWorldHint   = true,
+    } }
+)
+
 io.stderr:write(string.format("lmcp-hub starting on port %d with %d backends from %s\n",
     server.port, (function() local n = 0; for _ in pairs(backends) do n = n + 1 end; return n end)(), CONF_PATH))
 server:run()

lmcp.lua

@@ -3,10 +3,36 @@
 -- SPDX-License-Identifier: MIT
 
 local json = require('json')
+-- The version check. Phase B built it but never wired it up: measured on
+-- 2026-08-09, `versions` occurred zero times in this file, and the running
+-- server answered `initialize` with "1999-01-01" with HTTP 200 and
+-- "2025-06-18" -- verbatim the defect it was written against.
+local versions = require('versions')
+-- The envelope (envelope check 2026-07-28): required headers (Mcp-Method,
+-- MCP-Protocol-Version, Mcp-Name) + required _meta fields are checked before
+-- method dispatch. Based on job667.lua (coder run 5), two review fixes
+-- (mcp-name header, _meta under params._meta).
+local envelope = require('envelope')
 
 local lmcp = {}
 lmcp.__index = lmcp
 
+-- Module-level coroutine→ctx registry (issue #11). Weak keys so
+-- coroutines that die without explicit cleanup get GC'd out.
+-- Each ctx table carries a `server` back-reference, so any code with
+-- a coroutine handle can find both ctx and its owning lmcp instance.
+local _ctx_by_co = setmetatable({}, { __mode = "k" })
+
+-- server.lua and any other library code can call lmcp.current_ctx() to
+-- access the ctx of the currently-running dispatch coroutine. Returns
+-- nil outside coroutine context. Used by server.lua:run() to do
+-- transparent auto-cancellation of long-running shell-out polls.
+function lmcp.current_ctx()
+    local co = coroutine.running()
+    if co == nil then return nil end
+    return _ctx_by_co[co]
+end
+
 -- Read auth token from config file if present
 local function read_conf(path)
     local conf = {}
@@ -21,7 +47,12 @@ local function read_conf(path)
 end
 
 -- Protocol constants
-local MCP_VERSION = "2025-06-18"
+-- ONE truth. The version used to live twice in the tree -- here as a
+-- constant and in versions.M.SUPPORTED -- and the response line took
+-- this one. Changing the list therefore did NOT change what the server
+-- answers: measured with {"2025-06-18","2025-03-26"}, a call to
+-- initialize(2025-03-26) returned a 200 with "2025-06-18".
+local MCP_VERSION = versions.SUPPORTED[1]
 local JSONRPC = "2.0"
 
 function lmcp.new(name, opts)
@@ -32,6 +63,19 @@ function lmcp.new(name, opts)
     self.host = opts.host or "0.0.0.0"
     self.port = opts.port or 8080
     self.tools = {}
+    -- Per-instance allowlist (LMCP_TOOL_ALLOW, comma-separated). When set,
+    -- `tool()` registers ONLY those names -- built-ins like plugins.
+    -- When unset: everything as before. This is the only place where a
+    -- tool is created, hence the only place where it can be prevented;
+    -- removing one later means every future entry must know and stale out.
+    self.tool_allow = nil
+    do
+        local roh = os.getenv("LMCP_TOOL_ALLOW")
+        if roh and roh:match("%S") then
+            self.tool_allow = {}
+            for n in roh:gmatch("[^,%s]+") do self.tool_allow[n] = true end
+        end
+    end
     -- Resources primitive (MCP 2025-06-18 §Server/Resources). Storage is
     -- always present; capability is advertised iff `opts.resources` is
     -- truthy OR at least one resource/template has been registered by
@@ -69,6 +113,16 @@ function lmcp.new(name, opts)
     -- server calls `:roots(session_id, ...)`; invalidated when the client
     -- sends notifications/roots/list_changed.
     self._roots_cache = {}
+    -- Pending handler coroutines (issue #20 — concurrent dispatch).
+    -- Each entry: { co, conn, wake_at, finalise }. The scheduler tick
+    -- resumes any whose wake_at has passed and runs `finalise` on the
+    -- coroutine's return value to build the deferred response.
+    self._pending_handlers = {}
+    -- Cancellation flags (issue #11). Keyed by stringified JSON-RPC
+    -- request id. Only ever holds in-flight ids — see the
+    -- notifications/cancelled handler in handle_request which checks
+    -- for in-flight before inserting. Cleared by _finalise_dispatch.
+    self._cancelled_ids = {}
     -- Notification queue: drained by Streamable HTTP transport (issue #16).
     -- Today delivery is a no-op; we still enqueue so the emission code
     -- path is exercised. Capped + deduped to keep the queue useful.
@@ -137,10 +191,36 @@ end
 --     structuredContent (issue #13; spec-strict clients get first-class
 --     structured access)
 function lmcp:tool(name, description, params_schema, handler, opts)
+    -- Normalise empty inputSchema.properties → nil. JSON Schema allows
+    -- omitting `properties` on a `type: "object"` schema (means "any
+    -- object, no constraints"). Without this, an empty Lua properties
+    -- table goes through json.lua's is_array → emitted as `[]` →
+    -- spec-strict clients (Zod et al.) reject with
+    -- `expected: record, received: array`. The same gotcha already
+    -- bit `ping` in v1.0.0-rc1 (project_json_empty_table_gotcha
+    -- memory). v1.1.1 fix.
+    local schema = params_schema or { type = "object" }
+    if type(schema.properties) == "table" and next(schema.properties) == nil then
+        -- Clone the schema and drop the empty `properties` key. Avoids
+        -- mutating the caller's table (in case they re-use it across
+        -- registrations).
+        local clean = {}
+        for k, v in pairs(schema) do
+            if k ~= "properties" then clean[k] = v end
+        end
+        schema = clean
+    end
+    -- Allow-list: silently refuse, so a plugin offering a tool that is not
+    -- allowed does not crash -- it simply does not exist. `tools/list` and
+    -- `tools/call` both read the same registry, so an unregistered tool is
+    -- neither visible nor callable.
+    if self.tool_allow and not self.tool_allow[name] then
+        return self
+    end
     self.tools[name] = {
         name = name,
         description = description,
-        inputSchema = params_schema or { type = "object", properties = {} },
+        inputSchema = schema,
         handler = handler,
         annotations  = opts and opts.annotations  or nil,
         outputSchema = opts and opts.outputSchema or nil,
@@ -388,18 +468,75 @@ function lmcp:log(level, logger, data)
 end
 
 -- JSON-RPC response helpers
+-- MCP 2026-07-28: "The result MUST include a resultType field to indicate the
+-- type of the result." Hence here and not at the fourteen call sites:
+-- there should not be a single place that could forget it.
+--
+-- NEVER mutate the passed object. `json.empty_object` is a singleton
+-- (json.lua:235), two callers pass it straight through, and the encoder
+-- recognizes it by IDENTITY (json.lua:63) -- a write would have poisoned it
+-- process-wide AND the field would still have been swallowed.
 local function jsonrpc_result(id, result)
+    local copy = {}
+    if type(result) == "table" then
+        for k, v in pairs(result) do copy[k] = v end
+    end
+    -- "input_required" is set by the caller itself (MRTR); everything else
+    -- is "complete". An existing field is not overwritten.
+    if copy.resultType == nil then copy.resultType = "complete" end
+    return json.encode({ jsonrpc = JSONRPC, id = id, result = copy })
+end
+
+-- Legacy (2025-06-18) result encoder: the old server's jsonrpc_result did
+-- NOT inject resultType -- that field belongs to the 2026-07-28 result
+-- envelope (review finding #4: the legacy branches must keep the old wire
+-- shape; a strict old client that validates ping's empty result would
+-- break on a foreign resultType).
+local function jsonrpc_result_legacy(id, result)
     return json.encode({ jsonrpc = JSONRPC, id = id, result = result })
 end
 
-local function jsonrpc_error(id, code, message)
+local function jsonrpc_error(id, code, message, data)
+    -- `data` is optional and only set when present: all existing callers
+    -- pass three arguments and do not change.
+    -- Without this field, `data.supported` of the version check would fall
+    -- away on the way, and the rejection would be formally correct and
+    -- practically useless -- the client would not learn which version the
+    -- server speaks.
+    local err = { code = code, message = message }
+    if data ~= nil then err.data = data end
     return json.encode({
         jsonrpc = JSONRPC,
         id = id,
-        error = { code = code, message = message },
+        error = err,
     })
 end
 
+-- What this server can do. ONE computation for `initialize` and
+-- `server/discover`. Two places deriving the same thing are the disease
+-- this campaign found three times today (TRUST vs. roster,
+-- MCP_VERSION vs. M.SUPPORTED, the delta's stage boundary).
+function lmcp:_capabilities()
+    local caps = { tools = { listChanged = false } }
+    if self._force_resources_cap
+       or next(self.resources)
+       or self.resource_templates[1] then
+        caps.resources = { listChanged = true, subscribe = false }
+    end
+    if self._force_prompts_cap or next(self.prompts) then
+        caps.prompts = { listChanged = true }
+    end
+    if self._force_completions_cap or next(self.completions) then
+        -- Spec uses an empty object as the "supported" marker.
+        -- json.empty_object → {} (not [] from the empty-table gotcha).
+        caps.completions = json.empty_object
+    end
+    if self._force_logging_cap then
+        caps.logging = json.empty_object
+    end
+    return caps
+end
+
 -- Handle a single JSON-RPC request
 function lmcp:handle_request(req)
     local method = req.method
@@ -413,48 +550,103 @@ function lmcp:handle_request(req)
         if method == "notifications/roots/list_changed" then
             -- Invalidate cached roots for the session that sent this.
             if req._session_id then self._roots_cache[req._session_id] = nil end
+        elseif method == "notifications/cancelled" then
+            -- Issue #11 — flip cancel flag for the named request id,
+            -- but ONLY if the request is actually in-flight. Cancels
+            -- for unknown/already-completed ids drop silently (per Phase
+            -- 5 review fix #2 — prevents unbounded map growth).
+            local rid = (req.params or {}).requestId
+            if rid ~= nil then
+                local rid_str = tostring(rid)
+                local in_flight = false
+                -- Scan _ctx_by_co for a matching live request.
+                for _, c in pairs(_ctx_by_co) do
+                    if c.request_id ~= nil
+                       and tostring(c.request_id) == rid_str then
+                        in_flight = true; break
+                    end
+                end
+                if in_flight then
+                    self._cancelled_ids[rid_str] = true
+                end
+            end
         end
-        -- (Other client→server notifications: cancelled, message — no
-        -- action today; add side-effects here as needed.)
+        -- (Other client→server notifications drop silently.)
         return nil
     end
 
-    if method == "initialize" then
-        self._session_id = self._session_id or tostring(os.time())
+    -- Dual-protocol routing, transport-agnostic (review fix #1): the HTTP
+    -- path pre-sets _announced/_legacy in _dispatch_post (it can see the
+    -- MCP-Protocol-Version header); stdio reaches handle_request directly
+    -- with NO headers, so derive here from _meta only. Empty string counts
+    -- as "no version announced" (versions.check rule a, review fix #6).
+    if req._legacy == nil then
+        local meta_v = ((req.params or {})._meta or {})
+                           ["io.modelcontextprotocol/protocolVersion"]
+        if meta_v == "" then meta_v = nil end
+        req._announced = meta_v
+        req._legacy = (meta_v == nil or meta_v == "2025-06-18")
+    end
+
+    -- Legacy surface (2025-06-18), reachable only when the router announced
+    -- the old version or none at all (E2: no version = old). SEP-2575 removed
+    -- these from 2026-07-28; on the NEW path they fall through to the default
+    -- branch (-32601 + 404) exactly as the suite expects. Capabilities come
+    -- from the ONE `_capabilities()` computation, not a second copy.
+    if req._legacy and method == "initialize" then
         -- Capture client capabilities (MCP issue #9 — sampling needs to
         -- know if the client supports it before issuing a request).
         local p = req.params or {}
         self._client_caps = p.capabilities or {}
         self._client_info = p.clientInfo or {}
-        local caps = { tools = { listChanged = false } }
-        if self._force_resources_cap
-           or next(self.resources)
-           or self.resource_templates[1] then
-            caps.resources = { listChanged = true, subscribe = false }
-        end
-        if self._force_prompts_cap or next(self.prompts) then
-            caps.prompts = { listChanged = true }
-        end
-        if self._force_completions_cap or next(self.completions) then
-            -- Spec uses an empty object as the "supported" marker.
-            -- json.empty_object → {} (not [] from the empty-table gotcha).
-            caps.completions = json.empty_object
-        end
-        if self._force_logging_cap then
-            caps.logging = json.empty_object
-        end
-        return jsonrpc_result(id, {
+        return jsonrpc_result_legacy(id, {
             protocolVersion = MCP_VERSION,
-            capabilities = caps,
+            capabilities = self:_capabilities(),
             serverInfo = {
                 name = self.name,
                 version = self.version,
             },
         })
+    elseif req._legacy and method == "ping" then
+        return jsonrpc_result_legacy(id, json.empty_object)
+    elseif req._legacy and method == "logging/setLevel" then
+        local lvl = (req.params or {}).level
+        if type(lvl) ~= "string" or not LOG_LEVELS[lvl] then
+            return jsonrpc_error(id, -32602,
+                "level must be one of: debug, info, notice, warning, error, critical, alert, emergency")
+        end
+        self._log_level = lvl
+        return jsonrpc_result_legacy(id, json.empty_object)
+    end
 
-    elseif method == "ping" then
-        return jsonrpc_result(id, json.empty_object)
+    -- MCP 2026-07-28, /server/discover: "Servers MUST implement it."
+    -- Additive: a client of the old version never calls it. Placed BEFORE
+    -- `initialize` so it stays reachable should the latter fall away.
+    if method == "server/discover" then
+        -- Field names from the spec's example, not from the changelog: it is
+        -- `supportedVersions` (not protocolVersions), and serverInfo sits IN
+        -- `_meta`, not at the top. Both were guessed wrong in the draft.
+        local supported = {}
+        for i = 1, #versions.SUPPORTED do supported[i] = versions.SUPPORTED[i] end
+        local caps = self:_capabilities()
+        return jsonrpc_result(id, {
+            resultType        = "complete",
+            supportedVersions = supported,
+            -- An empty Lua table encodes as `[]`, not `{}` -- the spec
+            -- requires an object. (Finding @testdesigner, room #622.)
+            capabilities      = next(caps) and caps or json.empty_object,
+            _meta = {
+                ["io.modelcontextprotocol/serverInfo"] = {
+                    name = self.name,
+                    version = self.version,
+                },
+            },
+        })
 
+    -- SEP-2575: `initialize`, `ping` and `logging/setLevel` are gone from
+    -- spec 2026-07-28 and fall into the default branch here
+    -- (jsonrpc_error -32601, HTTP 404). Version negotiation moves into
+    -- the envelope (envelope check, last block).
     elseif method == "tools/list" then
         local tool_list = {}
         for _, t in pairs(self.tools) do
@@ -484,15 +676,59 @@ function lmcp:handle_request(req)
         if not tool then
             return jsonrpc_error(id, -32601, "Tool not found: " .. tostring(tool_name))
         end
-        -- ctx exposes the request's _meta (issue #13) and the session_id
-        -- (issue #9 — so handlers can call self:sample(ctx.session_id, …)).
-        -- Handlers that don't declare a second parameter ignore it (Lua
-        -- call discards extras).
-        local ctx = {
+        -- ctx exposes the request's _meta (issue #13), the session_id
+        -- (issue #9 — handlers can call self:sample(ctx.session_id, …)),
+        -- progress() and cancelled() (issue #11), and a `server` back-ref
+        -- (so lmcp.current_ctx() can find the right server instance
+        -- without a singleton). Handlers that don't declare a second
+        -- parameter ignore it (Lua call discards extras).
+        local rid_str = tostring(id)
+        local ptoken = (params._meta or {}).progressToken  -- nil if absent
+        local ctx
+        ctx = {
             _meta = params._meta,
             request_id = id,
             session_id = req._session_id,
+            server = self,
+            -- progress(p, total?, message?): emits notifications/progress
+            -- on session's notify_q. No-op if client didn't supply a
+            -- progressToken. Type-checks; rejects non-numeric progress.
+            progress = function(p, total, message)
+                if ptoken == nil then return false end
+                if type(p) ~= "number" then return false end
+                if total ~= nil and type(total) ~= "number" then return false end
+                local sess = self._sessions[req._session_id]
+                if not sess then return false end
+                local np = { progressToken = ptoken, progress = p }
+                if total   ~= nil then np.total   = total end
+                if message ~= nil then np.message = tostring(message) end
+                sess.notify_q[#sess.notify_q + 1] = {
+                    jsonrpc = JSONRPC, method = "notifications/progress",
+                    params = np,
+                }
+                return true
+            end,
+            -- cancelled(): true if a notifications/cancelled for this
+            -- request id has been received.
+            cancelled = function()
+                return self._cancelled_ids[rid_str] == true
+            end,
         }
+
+        -- Register on the currently-running coroutine so lmcp.current_ctx()
+        -- (and thus server.lua:run()'s auto-cancel) can find this ctx.
+        -- Pure-Lua handlers also get this registration; harmless.
+        local co = coroutine.running()
+        if co ~= nil then _ctx_by_co[co] = ctx end
+
+        -- Pre-handler cancellation short-circuit (Phase 5 review fix #9).
+        -- If a notifications/cancelled landed for this id before dispatch
+        -- reached here, skip the handler entirely. _finalise_dispatch
+        -- will see `not result` and suppress the response.
+        if self._cancelled_ids[rid_str] then
+            return nil
+        end
+
         local ok, result = pcall(tool.handler, arguments, ctx)
         if ok then
             local resp = { isError = false }
@@ -566,10 +802,15 @@ function lmcp:handle_request(req)
         end
         local entry, args = _resolve_resource(self, uri)
         if not entry then
-            -- -32002 is the MCP-conventional "Resource not found" code;
-            -- distinct from -32602 "Invalid params" so retry/UX logic
-            -- can tell a malformed URI from a missing one.
-            return jsonrpc_error(id, -32002, "Resource not found: " .. uri)
+            -- MCP 2026-07-28: "Implementations of this protocol version MUST
+            -- NOT emit these codes: -32002 — resource not found … replaced by
+            -- -32602." The earlier objection recorded here (‑32002 distinguishing
+            -- a missing from a malformed URI) is thereby settled: the spec has
+            -- deliberately merged both into -32602, and a receiver "MUST NOT
+            -- assume any specific meaning" for the old range. The difference
+            -- lives on in the message text, where it can no longer be confused
+            -- with a protocol promise.
+            return jsonrpc_error(id, -32602, "Resource not found: " .. uri)
         end
         local item, err = _read_resource(entry, args, uri)
         if not item then
@@ -597,7 +838,11 @@ function lmcp:handle_request(req)
         end
         local entry = self.prompts[name]
         if not entry then
-            return jsonrpc_error(id, -32002, "Prompt not found: " .. name)
+            -- Second occurrence of the same forbidden code. The suite does
+            -- NOT check it (it only asks resources/read) -- which is why it is
+            -- listed here: fixing only the tested one would make the counter
+            -- green and the server non-conformant.
+            return jsonrpc_error(id, -32602, "Prompt not found: " .. name)
         end
         local result, err = _get_prompt(entry, params.arguments)
         if not result then
@@ -605,15 +850,6 @@ function lmcp:handle_request(req)
         end
         return jsonrpc_result(id, result)
 
-    elseif method == "logging/setLevel" then
-        local lvl = (req.params or {}).level
-        if type(lvl) ~= "string" or not LOG_LEVELS[lvl] then
-            return jsonrpc_error(id, -32602,
-                "level must be one of: debug, info, notice, warning, error, critical, alert, emergency")
-        end
-        self._log_level = lvl
-        return jsonrpc_result(id, json.empty_object)
-
     elseif method == "completion/complete" then
         local params = req.params or {}
         local ref = params.ref or {}
@@ -831,7 +1067,7 @@ local function _check_auth(self, conn)
     if not self._auth_token then return true end
     if conn.method == "OPTIONS" then return true end
     local auth = conn.headers["authorization"] or ""
-    local token = auth:match("^Bearer%s+(.+)$")
+    local token = auth:match("^[Bb]earer%s+(.+)$")
     return token == self._auth_token
 end
 
@@ -887,6 +1123,40 @@ end
 
 -- ---- Dispatch a fully-parsed POST body ----
 
+-- Forward declarations: used by _dispatch_post, defined below.
+local _drive_handler_co
+local _finalise_dispatch
+
+-- MCP 2026-07-28 standardizes the HTTP status, not just the body. This one
+-- mapping decides it, so each branch does not choose it itself -- and so the
+-- following parts (header, _meta, version) get their 400s without anyone
+-- having to add to it again.
+--
+--   -32601  unknown method                -> 404  („MUST respond with 404 Not
+--                                                  Found and … -32601")
+--   -32020  HeaderMismatch               -> 400
+--   -32021  MissingRequiredClientCapability -> 400
+--   -32022  UnsupportedProtocolVersion   -> 400
+--   -32602  Invalid params               -> 400  (missing required _meta fields)
+local _error_status = {
+    [-32601] = "404 Not Found",
+    [-32020] = "400 Bad Request",
+    [-32021] = "400 Bad Request",
+    [-32022] = "400 Bad Request",
+    [-32602] = "400 Bad Request",
+}
+
+local function _status_for_error_code(encoded)
+    -- `encoded` is the finished JSON response. Decoding once is cheaper
+    -- and more honest than searching the string for numbers: a string
+    -- can contain `-32601` inside a message too.
+    local ok, obj = pcall(json.decode, encoded)
+    if not ok or type(obj) ~= "table" or type(obj.error) ~= "table" then
+        return "200 OK"
+    end
+    return _error_status[obj.error.code] or "200 OK"
+end
+
 local function _dispatch_post(self, conn)
     local body = conn.body
     if body == "" then
@@ -918,6 +1188,67 @@ local function _dispatch_post(self, conn)
         end
     end
 
+    -- Dual-protocol routing (Phase 2). The announced version comes from the
+    -- MCP-Protocol-Version header OR _meta.protocolVersion (E2: either source
+    -- is authoritative). Both present and differing -> -32020, checked BEFORE
+    -- routing so a mixed announcement never reaches either protocol path.
+    -- An empty string counts as "no version" (versions.check rule a; review
+    -- fix #6) — it is treated as absent, exactly like a missing channel.
+    local header_v = conn.headers['mcp-protocol-version']
+    if header_v == "" then header_v = nil end
+    local meta_v = ((rpc_req.params or {})._meta or {})
+                       ["io.modelcontextprotocol/protocolVersion"]
+    if meta_v == "" then meta_v = nil end
+    if header_v and meta_v and header_v ~= meta_v then
+        return _build_http_response(
+            _error_status[-32020] or "400 Bad Request",
+            { ["Content-Type"] = "application/json",
+              ["Access-Control-Allow-Origin"] = "*" },
+            jsonrpc_error(rpc_req.id, -32020,
+                "Header MCP-Protocol-Version does not match _meta protocolVersion"),
+            nil)
+    end
+    local announced = header_v or meta_v
+    rpc_req._announced = announced
+
+    -- Route: legacy (nil or 2025-06-18) skips the envelope entirely (B5:
+    -- a legacy request with no Mcp-Method/Version headers must NOT get
+    -- -32020). New (2026-07-28) runs the envelope + version gate as before.
+    local legacy = (announced == nil or announced == "2025-06-18")
+    rpc_req._legacy = legacy
+
+    -- Version gate (keystone): `versions.check` rejects unknown versions
+    -- with -32022 + 400 and names data.supported (a copy of the list,
+    -- copy-protection unchanged) + data.requested (the requested version).
+    -- Allows nil/"" (contract rule a). Runs on the announced version (header
+    -- OR _meta, B6: header-only bogus version must still fire -32022).
+    if announced ~= nil and not legacy then
+        local allowed, err = versions.check(announced)
+        if not allowed then
+            err.data.requested = announced
+            return _build_http_response(
+                _error_status[err.code] or "400 Bad Request",
+                { ["Content-Type"] = "application/json",
+                  ["Access-Control-Allow-Origin"] = "*" },
+                jsonrpc_error(rpc_req.id, err.code,
+                              "Unsupported protocol version", err.data), nil)
+        end
+    end
+
+    -- Envelope check (2026-07-28): required headers + _meta + version BEFORE
+    -- method dispatch. Errors carry their JSON-RPC code and their HTTP status
+    -- from _error_status (-32020/-32602 -> 400). Legacy requests skip it.
+    if not legacy then
+        local e_ok, e_code, e_message = envelope.check(conn.headers, rpc_req)
+        if not e_ok then
+            return _build_http_response(
+                _error_status[e_code] or "400 Bad Request",
+                { ["Content-Type"] = "application/json",
+                  ["Access-Control-Allow-Origin"] = "*" },
+                jsonrpc_error(rpc_req.id, e_code, e_message), nil)
+        end
+    end
+
     -- Session resolution (deferred from header-parse time so we can detect
     -- `initialize`). Rules:
     --   - `initialize`: always mint a fresh session, ignoring any client sid
@@ -942,28 +1273,120 @@ local function _dispatch_post(self, conn)
     -- expose it to handler ctx (issue #9 — sampling needs to know which
     -- session to push the request onto).
     rpc_req._session_id = sess.id
+    -- Stash the JSON-RPC id on the conn so _finalise_dispatch can clear
+    -- the cancellation flag for this request after building the response
+    -- (issue #11). Notifications have nil id; that's fine — the
+    -- nil-guard in _finalise_dispatch keeps tostring(nil) out of the
+    -- cancel map.
+    conn.dispatch_id = rpc_req.id
 
-    -- Normal client request / notification. Dispatch via handle_request.
-    local response = self:handle_request(rpc_req)
-    if not response then
+    -- Concurrent handler dispatch (issue #20). Wrap the dispatch call in
+    -- a coroutine so any tool handler that goes through server.lua:run()
+    -- (which yields when polling its sentinel file) can return control to
+    -- the event loop while it waits. Other connections continue making
+    -- progress.
+    --
+    -- The coroutine resumes itself synchronously the first time. If it
+    -- completes without yielding (pure-Lua handlers, ping, etc.) the
+    -- response is built inline as before. If it yields, we park it in
+    -- self._pending_handlers and return nil — the conn enters
+    -- dispatching_async, the scheduler tick resumes when wake_at passes.
+    local co = coroutine.create(function()
+        return self:handle_request(rpc_req)
+    end)
+    return _drive_handler_co(self, conn, co)
+end
+
+-- Resume a handler coroutine until it completes or yields. On completion,
+-- build the deferred HTTP response (preserving the Accept-aware shape).
+-- On yield, register in self._pending_handlers and return nil — the conn
+-- is parked in dispatching_async until the scheduler resumes it.
+_drive_handler_co = function(self, conn, co)
+    local rok, ryield = coroutine.resume(co)
+    if coroutine.status(co) == "dead" then
+        return _finalise_dispatch(self, conn, rok, ryield, co)
+    end
+    -- Suspended. Parse the yield payload.
+    local wake_at = (type(ryield) == "table" and ryield.wake_at) or 0
+    self._pending_handlers[#self._pending_handlers + 1] = {
+        co = co, conn = conn, wake_at = wake_at,
+    }
+    conn.state = "dispatching_async"
+    return nil  -- no write_buf change; conn parks
+end
+
+-- Build the HTTP response for a completed dispatch. `rok` is the coroutine.resume
+-- success flag; `result` is the handler/dispatch return (a JSON-RPC string when
+-- rok=true; an error message when rok=false). Used by both the sync path
+-- (_dispatch_post tail) and the async resume path (_scheduler_tick).
+-- Also: clears cancellation flag and ctx-by-co registry entry for this
+-- request (issue #11 — single cleanup site per Phase 5 review fix #7).
+_finalise_dispatch = function(self, conn, rok, result, co)
+    local session_id = conn.session_id
+
+    -- Cleanup (always): drop the coroutine's ctx entry and any
+    -- cancellation flag for this request id.
+    if co ~= nil then _ctx_by_co[co] = nil end
+    local rid = conn.dispatch_id
+    local was_cancelled = false
+    if rid ~= nil then
+        local rid_str = tostring(rid)
+        if self._cancelled_ids[rid_str] then
+            was_cancelled = true
+            self._cancelled_ids[rid_str] = nil
+        end
+    end
+    -- Issue #11: cancelled requests get a -32800 JSON-RPC error response.
+    -- The MCP spec wording is "SHOULD NOT respond" (not MUST NOT). A silent
+    -- TCP-close would be cleaner but the spawned shell subprocess in
+    -- server.lua:run() inherits the socket FD via fork(), so the kernel
+    -- keeps the connection alive until that shell exits (i.e. the
+    -- underlying long-running command completes anyway). The error
+    -- response gives the client a structured signal and exits curl
+    -- immediately, which is the practical UX they want. JSON-RPC 2.0
+    -- code -32800 is the convention for "Request cancelled."
+    if was_cancelled then
+        return _build_http_response("200 OK",
+            { ["Content-Type"] = "application/json",
+              ["Access-Control-Allow-Origin"] = "*" },
+            jsonrpc_error(rid, -32800, "Request cancelled"),
+            session_id)
+    end
+
+    if not rok then
+        -- Internal dispatch error — surface as a JSON-RPC error response.
+        return _build_http_response("500 Internal Server Error",
+            { ["Content-Type"] = "application/json",
+              ["Access-Control-Allow-Origin"] = "*" },
+            jsonrpc_error(nil, -32603, "Internal error: " .. tostring(result)),
+            session_id)
+    end
+    if not result then
         -- Notification → 202 Accepted, no body.
         return _build_http_response("202 Accepted",
             { ["Content-Type"] = "application/json",
               ["Access-Control-Allow-Origin"] = "*" },
-            "", conn.session_id)
+            "", session_id)
     end
-
-    -- If client accepts SSE, respond as a single-event SSE stream.
-    -- Otherwise plain JSON body.
+    -- Accept-aware response shape (re-checked at finalise time; survives
+    -- parking because conn.headers is captured by the closure scope).
     local accept = conn.headers["accept"] or ""
-    if accept:find("text/event%-stream") then
-        local hdrs = _build_sse_headers(conn.session_id)
-        return hdrs .. _sse_event(response)
+    -- Errors are NEVER streamed. The spec requires the client to send an
+    -- Accept header with text/event-stream (streamable-http, "MUST include
+    -- an Accept header listing both"), so the SSE branch only hit exactly
+    -- the prescribed request shape -- and _build_sse_headers wires 200.
+    -- The status mapping was therefore unreachable for the normal case. An
+    -- error has nothing to stream: no notifications before it, nothing to
+    -- keep open.
+    local status = _status_for_error_code(result)
+    if status == "200 OK" and accept:find("text/event%-stream") then
+        local hdrs = _build_sse_headers(session_id)
+        return hdrs .. _sse_event(result)
     end
-    return _build_http_response("200 OK",
+    return _build_http_response(status,
         { ["Content-Type"] = "application/json",
           ["Access-Control-Allow-Origin"] = "*" },
-        response, conn.session_id)
+        result, session_id)
 end
 
 local function _dispatch_options(conn)
@@ -1072,55 +1495,32 @@ local function _conn_read(self, conn)
         if conn.method == "OPTIONS" then
             conn.write_buf = _dispatch_options(conn)
             conn.state = "writing"
-        elseif conn.method == "DELETE" then
-            -- Resolve session: no auto-issue for DELETE; unknown sid → 404.
-            if not conn.requested_sid then
-                conn.write_buf = _build_http_response("400 Bad Request",
-                    { ["Content-Type"] = "text/plain",
-                      ["Access-Control-Allow-Origin"] = "*" },
-                    "Mcp-Session-Id required for DELETE", nil)
-                conn.state = "writing"
-                return
-            end
-            conn.session_id = conn.requested_sid
-            conn.write_buf = _dispatch_delete(self, conn)
+        elseif conn.method == "GET" or conn.method == "DELETE" then
+            -- MCP 2026-07-28: „HTTP GET or DELETE to the MCP endpoint: respond
+            -- with 405 Method Not Allowed." Both carried the mechanics of the
+            -- old version — GET opened a separate SSE stream, DELETE ended
+            -- a session. Spec quote for both: „None of these mechanisms are
+            -- part of this revision."
+            conn.write_buf = _build_http_response("405 Method Not Allowed",
+                { ["Content-Type"] = "text/plain",
+                  ["Allow"] = "POST, OPTIONS",
+                  ["Access-Control-Allow-Origin"] = "*" },
+                "Method Not Allowed", nil)
             conn.state = "writing"
-        elseif conn.method == "GET" then
-            -- Resolve session (auto-issue if missing, 404 if unknown).
-            local sess, serr = _resolve_session(self, conn.requested_sid)
-            if not sess then
-                conn.write_buf = _build_http_response("404 Not Found",
-                    { ["Content-Type"] = "text/plain",
-                      ["Access-Control-Allow-Origin"] = "*" },
-                    "Session not found: " .. tostring(conn.requested_sid), nil)
-                conn.state = "writing"
-                return
-            end
-            conn.session_id = sess.id
-            -- Persistent SSE stream. Enforce one per session.
-            if sess.sse_conn and sess.sse_conn ~= conn then
-                conn.write_buf = _build_http_response("409 Conflict",
-                    { ["Content-Type"] = "text/plain",
-                      ["Access-Control-Allow-Origin"] = "*" },
-                    "Session already has an open SSE stream", nil)
-                conn.state = "writing"
-                return
-            end
-            sess.sse_conn = conn
-            local hdrs = _build_sse_headers(conn.session_id)
-            -- For backwards compat with the old SDK probe shape, emit a one-shot
-            -- 'endpoint' event so clients can self-discover; modern clients
-            -- ignore it and just await message events.
-            local endpoint_data = json.encode({
-                endpoint = "/mcp", sessionId = conn.session_id,
-            })
-            conn.write_buf = hdrs ..
-                "event: endpoint\r\ndata: " .. endpoint_data .. "\r\n\r\n"
-            conn.state = "sse_open"
-            conn.last_heart = os.time()
         elseif conn.method == "POST" then
-            conn.write_buf = _dispatch_post(self, conn)
-            conn.state = "writing"
+            -- _dispatch_post may return nil (issue #20) if the handler
+            -- coroutine yielded. In that case it set conn.state =
+            -- "dispatching_async" itself and parked the coroutine.
+            local resp = _dispatch_post(self, conn)
+            if resp then
+                conn.write_buf = resp
+                -- _finalise_dispatch sets conn.state = "closing" for
+                -- cancelled requests (issue #11); only override if not.
+                if conn.state ~= "closing" then
+                    conn.state = "writing"
+                end
+            end
+            -- else: conn already parked; scheduler tick will finalise.
         else
             conn.write_buf = _build_http_response("405 Method Not Allowed",
                 { ["Content-Type"] = "text/plain",
@@ -1195,6 +1595,73 @@ local function _heartbeat_tick(self)
     end
 end
 
+-- Issue #20 — scheduler tick. Resume any parked dispatch coroutine whose
+-- wake_at has passed. On completion, build the deferred response and
+-- queue it for write. If the connection died while the handler was
+-- parked, drop the coroutine.
+--
+-- gettime() is wall-clock (luasocket uses gettimeofday) — NOT monotonic.
+-- A large NTP step backwards could delay resumes; forwards could bunch
+-- them. Acceptable for the deployment fleet (chrony slews); revisit if
+-- a use case appears that needs CLOCK_MONOTONIC.
+local function _scheduler_tick(self)
+    if not self._pending_handlers[1] then return end
+    local socket = require("socket")
+    local now = socket.gettime()
+    local i = 1
+    while i <= #self._pending_handlers do
+        local p = self._pending_handlers[i]
+        if p.conn.state == "closing" then
+            -- Connection died mid-handler; drop the coroutine entirely
+            -- and free its ctx entry (issue #11 cleanup discipline).
+            _ctx_by_co[p.co] = nil
+            if p.conn.dispatch_id ~= nil then
+                self._cancelled_ids[tostring(p.conn.dispatch_id)] = nil
+            end
+            table.remove(self._pending_handlers, i)
+        elseif now >= p.wake_at then
+            -- Time to resume. Remove from pending BEFORE resume so a
+            -- re-yielding handler re-adds itself cleanly via _drive_handler_co.
+            table.remove(self._pending_handlers, i)
+            local rok, ryield = coroutine.resume(p.co)
+            if coroutine.status(p.co) == "dead" then
+                local resp = _finalise_dispatch(self, p.conn, rok, ryield, p.co)
+                p.conn.write_buf = (p.conn.write_buf or "") .. resp
+                -- _finalise_dispatch may set conn.state = "closing" for
+                -- cancelled requests; only transition to writing if it
+                -- didn't already pick the closing path.
+                if p.conn.state ~= "closing" then
+                    p.conn.state = "writing"
+                end
+            else
+                -- Yielded again — re-park.
+                local wake_at = (type(ryield) == "table" and ryield.wake_at) or 0
+                self._pending_handlers[#self._pending_handlers + 1] = {
+                    co = p.co, conn = p.conn, wake_at = wake_at,
+                }
+            end
+        else
+            i = i + 1
+        end
+    end
+end
+
+-- Returns the earliest pending wake_at as an offset from now, or nil if
+-- no handlers are parked. Used to tighten the select() timeout so the
+-- scheduler wakes on the right beat.
+local function _next_pending_delay(self)
+    if not self._pending_handlers[1] then return nil end
+    local socket = require("socket")
+    local now = socket.gettime()
+    local earliest = math.huge
+    for _, p in ipairs(self._pending_handlers) do
+        if p.wake_at < earliest then earliest = p.wake_at end
+    end
+    local d = earliest - now
+    if d < 0 then return 0 end
+    return d
+end
+
 -- ---- Public: server-initiated request (for sampling/roots/etc.) ----
 -- Enqueues a JSON-RPC request on the session's SSE stream. The callback
 -- fires when the client POSTs back the response (matched by id).
@@ -1322,7 +1789,14 @@ function lmcp:run()
             end
         end
 
-        local ready_r, ready_w = socket.select(reads, writes, SELECT_TIMEOUT)
+        -- Tighten select timeout if a parked handler is due sooner.
+        -- Otherwise a 100ms tick adds 100ms latency to short shell-tool runs.
+        local select_timeout = SELECT_TIMEOUT
+        local next_pend = _next_pending_delay(self)
+        if next_pend and next_pend < select_timeout then
+            select_timeout = next_pend
+        end
+        local ready_r, ready_w = socket.select(reads, writes, select_timeout)
 
         for _, sock in ipairs(ready_r or {}) do
             if sock == server_sock then
@@ -1365,9 +1839,11 @@ function lmcp:run()
         -- Per-tick maintenance.
         _drain_notifications(self)
         _heartbeat_tick(self)
+        _scheduler_tick(self)  -- issue #20: resume due dispatch coroutines
 
         -- After draining, attempt immediate writes on conns whose write_buf
-        -- just got bytes (so list_changed / heartbeat appears within one tick).
+        -- just got bytes (so list_changed / heartbeat / async-completed
+        -- responses appear within one tick).
         for sock, conn in pairs(self._conns) do
             if conn.write_buf ~= "" and conn.state ~= "closing" then
                 pcall(_conn_write, conn)

server.lua

@@ -35,7 +35,51 @@ local function tmpname()
     end
 end
 
+-- Lazy-required luasocket — only needed in the coroutine path for
+-- gettime(). Avoids forcing luasocket as a hard dep at server.lua
+-- load time (callers like example_server already require it via lmcp).
+local _socket = nil
+local function gettime()
+    if not _socket then _socket = require("socket") end
+    return _socket.gettime()
+end
+
+-- Lazy access to the lmcp module for cross-module ctx lookup (issue #11).
+-- server.lua doesn't statically require lmcp (it's an example/runtime
+-- server, not the library); but lmcp must already be loaded when we run.
+-- Defensive: if the lookup fails for any reason, current_ctx returns nil
+-- and run() falls back to non-cancellable behaviour.
+local _lmcp_mod = nil
+local function current_ctx()
+    if _lmcp_mod == false then return nil end
+    if _lmcp_mod == nil then
+        local ok, mod = pcall(require, "lmcp")
+        _lmcp_mod = ok and mod or false
+        if _lmcp_mod == false then return nil end
+    end
+    return _lmcp_mod.current_ctx and _lmcp_mod.current_ctx() or nil
+end
+
+-- in_coroutine() — true if we're running inside an lmcp dispatch
+-- coroutine (issue #20). Handles both Lua 5.4 (coroutine.running →
+-- (co, isMain)) and LuaJIT 5.1 (coroutine.running → nil on main).
+local function in_coroutine()
+    local co, is_main = coroutine.running()
+    if co == nil then return false end       -- 5.1 / LuaJIT main
+    if is_main then return false end          -- 5.4 main thread
+    return true
+end
+
 local function sleep_ms(ms)
+    -- Coroutine-aware: yield with a wake deadline instead of busy-blocking.
+    -- The lmcp event loop services I/O for other connections while this
+    -- coroutine sleeps, then resumes it once the deadline elapses.
+    -- (Issue #20: gives concurrent tool dispatch without changing handler
+    -- source code — tools that go through run() get it for free.)
+    if in_coroutine() then
+        coroutine.yield({ wake_at = gettime() + (ms / 1000) })
+        return
+    end
     if WINDOWS then
         -- ping loopback: ~1s per -n count. For sub-second, use busy-wait.
         if ms < 500 then
@@ -78,6 +122,35 @@ local function run(cmd, timeout_sec)
     local out_file = base .. ".out"
     local done_file = base .. ".done"
 
+    -- Wall-clock deadline rather than an accumulated interval-counter:
+    -- when we're inside a dispatch coroutine (issue #20), the scheduler
+    -- may delay our resume by more than `interval`, so an accumulator
+    -- diverges from real elapsed. gettime() comparison stays honest in
+    -- both busy-poll and yield-resume modes.
+    --
+    -- Auto-cancellation (issue #11): if a ctx is available on the
+    -- running coroutine AND it has been cancelled, exit the polling
+    -- loop early. The interval is capped at 500ms when a ctx is
+    -- present so worst-case cancel latency is ~0.5s, not ~2s.
+    local started = gettime()
+    local cancelled = false
+    local function poll_loop()
+        local interval = WINDOWS and 100 or 50  -- ms
+        while gettime() - started < timeout_sec do
+            if file_exists(done_file) then return true end
+            local ctx = current_ctx()
+            if ctx and ctx.cancelled and ctx.cancelled() then
+                cancelled = true
+                return false
+            end
+            sleep_ms(interval)
+            if interval < 2000 then interval = math.floor(interval * 1.5) end
+            -- When cancellable, cap so we can respond to cancel quickly.
+            if ctx and interval > 500 then interval = 500 end
+        end
+        return false
+    end
+
     if WINDOWS then
         -- Write a batch wrapper that runs the command and signals completion
         local bat_file = base .. ".bat"
@@ -89,49 +162,51 @@ local function run(cmd, timeout_sec)
         bf:close()
         os.execute('start /B cmd /C "' .. bat_file .. '"')
 
-        -- Poll for sentinel
-        local elapsed = 0
-        local interval = 100  -- ms
-        while elapsed < timeout_sec * 1000 do
-            if file_exists(done_file) then break end
-            sleep_ms(interval)
-            elapsed = elapsed + interval
-            if interval < 2000 then interval = math.floor(interval * 1.5) end
-        end
-
+        local completed = poll_loop()
         local output = read_file(out_file)
         remove_silent(bat_file)
         remove_silent(out_file)
         remove_silent(done_file)
 
-        if elapsed >= timeout_sec * 1000 then
+        if not completed then
+            if cancelled then return "(cancelled)" end
             return output or ("Error: command timed out after " .. timeout_sec .. "s")
         end
         return output and output ~= "" and output or "(no output)"
     else
-        -- POSIX: use shell backgrounding + wait with timeout
-        -- sh -c '(cmd > out 2>&1; echo $? > done) &' then poll
+        -- POSIX: run in its OWN session/process group (setsid) so a
+        -- timeout or cancel can kill the WHOLE tree instead of orphaning
+        -- backgrounded children (the classic "shell timed out, children
+        -- kept thrashing" bug). $! is the setsid leader pid == pgid.
+        local pid_file = base .. ".pid"
         local sh_cmd = string.format(
             "(%s) > '%s' 2>&1; echo $? > '%s'",
             cmd, out_file, done_file
         )
-        os.execute("sh -c '" .. sh_cmd:gsub("'", "'\\''") .. "' &")
+        os.execute("setsid sh -c '" .. sh_cmd:gsub("'", "'\\''")
+            .. "' & echo $! > '" .. pid_file .. "'")
+        local pgid = (read_file(pid_file) or ""):match("(%d+)")
+        remove_silent(pid_file)
 
-        local elapsed = 0
-        local interval = 50  -- ms
-        while elapsed < timeout_sec * 1000 do
-            if file_exists(done_file) then break end
-            sleep_ms(interval)
-            elapsed = elapsed + interval
-            if interval < 2000 then interval = math.floor(interval * 1.5) end
+        local completed = poll_loop()
+
+        -- Timeout or cancel -> kill the entire process group. No orphans.
+        if not completed and pgid then
+            os.execute("kill -TERM -" .. pgid .. " 2>/dev/null")
+            sleep_ms(300)
+            os.execute("kill -KILL -" .. pgid .. " 2>/dev/null")
         end
 
         local output = read_file(out_file)
         remove_silent(out_file)
         remove_silent(done_file)
 
-        if elapsed >= timeout_sec * 1000 then
-            return output or ("Error: command timed out after " .. timeout_sec .. "s")
+        if not completed then
+            if cancelled then return "(cancelled -- process group killed)" end
+            return (output and output ~= "" and (output .. "\n") or "")
+                .. "Error: command timed out after " .. timeout_sec
+                .. "s -- the process group was KILLED (nothing is still running). "
+                .. "For a long-running command, re-run it with shell_bg."
         end
         return output and output ~= "" and output or "(no output)"
     end
@@ -142,6 +217,13 @@ end
 local server_name = os.getenv("LMCP_NAME") or (WINDOWS and "windows-tools" or "linux-tools")
 local server = lmcp.new(server_name, {
     port = tonumber(os.getenv("LMCP_PORT") or arg[1]) or 8080,
+    -- LMCP_HOST: bind interface (default 0.0.0.0). Hosts that need
+    -- single-interface binding (hertz: 192.168.88.18 only) set this.
+    host = os.getenv("LMCP_HOST"),
+    -- LMCP_CONF: path to a conf file with bearer-token entries
+    -- (e.g. /opt/herding/etc/hertz-tools.conf). Read by lmcp.lua's
+    -- read_conf; the `.godparticle` entry becomes the bearer token.
+    conf = os.getenv("LMCP_CONF"),
 })
 
 -- ---- Tools ----
@@ -218,7 +300,15 @@ server:tool("shell_bg",
             f:close()
             os.remove(pid_file)
         end
-        return string.format("launched pid=%s log=%s", pid, log)
+        -- register so list_jobs/kill_job can see and reap it (no more reboots)
+        if pid ~= "?" then
+            local reg = io.open("/tmp/lmcp-bg-jobs.tsv", "a")
+            if reg then
+                reg:write(pid.."\t"..log.."\t"..os.date("%Y-%m-%dT%H:%M:%S").."\t"..inner:gsub("[\t\n]"," ").."\n")
+                reg:close()
+            end
+        end
+        return string.format("launched pid=%s log=%s (kill with kill_job pid=%s)", pid, log, pid)
     end, {
         annotations = {
             title           = "Run shell (background)",
@@ -229,6 +319,42 @@ server:tool("shell_bg",
         },
     })
 
+server:tool("kill_job",
+    "Kill a runaway background job by PID. SIGKILLs the whole process group of a shell_bg/setsid job so no children survive. Use when a background job is thrashing a machine.",
+    { type = "object", properties = { pid = { type = "integer", description = "PID from shell_bg / list_jobs" } }, required = { "pid" } },
+    function(a)
+        if WINDOWS then return "Error: kill_job is Linux-only" end
+        local pid = tostring(a.pid or ""):match("(%d+)")
+        if not pid then return "Error: numeric pid required" end
+        os.execute("kill -KILL -"..pid.." 2>/dev/null; kill -KILL "..pid.." 2>/dev/null")
+        sleep_ms(200)
+        local alive = os.execute("kill -0 "..pid.." 2>/dev/null")
+        if alive == true or alive == 0 then return "pid "..pid.." may still be alive (uninterruptible?)" end
+        return "killed pid "..pid.." (process group)"
+    end,
+    { annotations = { title = "Kill background job", destructiveHint = true } })
+
+server:tool("list_jobs",
+    "List background jobs started via shell_bg and whether each is still running. Use to find runaway jobs to kill_job.",
+    { type = "object", properties = {} },
+    function()
+        if WINDOWS then return "Error: list_jobs is Linux-only" end
+        local reg = io.open("/tmp/lmcp-bg-jobs.tsv", "r")
+        if not reg then return "(no background jobs recorded)" end
+        local out = {}
+        for line in reg:lines() do
+            local pid, log, ts, cmd = line:match("^(%d+)\t([^\t]*)\t([^\t]*)\t(.*)$")
+            if pid then
+                local alive = os.execute("kill -0 "..pid.." 2>/dev/null")
+                local st = (alive == true or alive == 0) and "RUNNING" or "done"
+                table.insert(out, string.format("pid=%s [%s] %s log=%s\n    %s", pid, st, ts, log, (cmd or ""):sub(1,100)))
+            end
+        end
+        reg:close()
+        return #out>0 and table.concat(out, "\n") or "(no background jobs recorded)"
+    end,
+    { annotations = { title = "List background jobs", readOnlyHint = true } })
+
 server:tool("read_file", "Read a file.", {
     type = "object",
     properties = { path = { type = "string" } },
@@ -1008,6 +1134,46 @@ if WINDOWS then
     })
 end
 
+-- ---- host-local tool plugins (issue #22) ----
+-- Load every .lua file in LMCP_TOOLS_DIR (default /opt/lmcp/tools.d on POSIX,
+-- %ProgramData%\lmcp\tools.d on Windows). Each file is invoked as a function
+-- receiving the configured `server` instance and the `run` helper:
+--
+--     local server, run = ...
+--     server:tool("my_local_tool", "...", {...}, function(a) return run(...) end)
+--
+-- This is the standard plugin pattern (nginx conf.d/, systemd-tmpfiles.d, …).
+-- Hosts can ship their own tools alongside the packaged generics without
+-- forking the upstream server.lua.
+local plugin_dir = os.getenv("LMCP_TOOLS_DIR")
+    or (WINDOWS and (os.getenv("ProgramData") or "C:\\ProgramData") .. "\\lmcp\\tools.d"
+                 or "/opt/lmcp/tools.d")
+local list_cmd = WINDOWS
+    and ('dir /b "' .. plugin_dir .. '\\*.lua" 2>nul')
+    or  ('ls -1 "'  .. plugin_dir ..  '"/*.lua 2>/dev/null')
+local lh = io.popen(list_cmd)
+if lh then
+    for path in lh:lines() do
+        -- On Windows `dir /b` emits bare filenames; prefix the dir.
+        local full = path:match("[/\\]") and path
+                  or (plugin_dir .. (WINDOWS and "\\" or "/") .. path)
+        local chunk, err = loadfile(full)
+        if chunk then
+            local ok, perr = pcall(chunk, server, run)
+            if ok then
+                io.stderr:write("lmcp: loaded plugin " .. full .. "\n")
+            else
+                io.stderr:write("lmcp: plugin " .. full .. " errored: "
+                    .. tostring(perr) .. "\n")
+            end
+        else
+            io.stderr:write("lmcp: plugin " .. full .. " load error: "
+                .. tostring(err) .. "\n")
+        end
+    end
+    lh:close()
+end
+
 local transport = os.getenv("LMCP_TRANSPORT") or "http"
 if transport == "stdio" then
     if os.getenv("LMCP_PORT") then

tests/dual_protocol_acceptance.py

@@ -0,0 +1,201 @@
+#!/usr/bin/env python3
+"""Acceptance test for lmcp's dual protocol: does a shell tool echo "hello world"
+over BOTH the 2025-06-18 and the 2026-07-28 surface, on BOTH transports?
+
+WHY THIS SHAPE
+--------------
+The two protocols are not symmetric, and a test that treats them as such would
+pass for the wrong reason:
+
+  * 2025-06-18 has `initialize`. 2026-07-28 does NOT -- SEP-2575 removed it, and
+    on the new path it deliberately falls through to -32601. So the old flow is
+    initialize -> tools/call and the new flow is tools/call alone.
+  * The envelope (headers Mcp-Method / Mcp-Name / MCP-Protocol-Version plus
+    params._meta) is checked in the HTTP dispatcher only. Over stdio there are
+    no headers at all, so there the version is announced through _meta only.
+
+CONTROLS
+--------
+A green result means nothing unless the same harness can produce a red one, so:
+
+  RED-1  an invented version (1999-01-01) must be REJECTED, not served. If the
+         server answers "hello world" to that, the version gate is decorative.
+  RED-2  the same stdio/old-protocol case run against the code from BEFORE the
+         review fix (4ac9296) must FAIL with -32601. That is the echo guard: it
+         proves this harness can see the very defect the fix was for. Without
+         it, a green run only proves the harness is polite.
+
+Usage: lmcp-dual-test.py <lmcp-dir> [http-port]
+"""
+import json
+import os
+import subprocess
+import sys
+import time
+import urllib.error
+import urllib.request
+
+DIR = sys.argv[1] if len(sys.argv) > 1 else "/opt/lmcp"
+PORT = int(sys.argv[2]) if len(sys.argv) > 2 else 8099
+ALT, NEU = "2025-06-18", "2026-07-28"
+BEFEHL = "echo hello world"
+ERWARTET = "hello world"
+
+ergebnisse = []
+
+
+def merke(name, ok, detail=""):
+    ergebnisse.append((name, ok, detail))
+    print(f"  [{'ok ' if ok else '!! '}] {name}" + (f"  — {detail}" if detail else ""))
+
+
+def env_fuer(transport):
+    e = dict(os.environ)
+    e["LMCP_TRANSPORT"] = transport
+    e["LMCP_TOOLS_DIR"] = "/tmp/leer"      # keine Flotten-Werkzeuge laden
+    e["LMCP_PORT"] = str(PORT)
+    return e
+
+
+# --------------------------------------------------------------------------
+# stdio: eine Zeile rein, eine Zeile raus. Jede Anfrage bekommt einen frischen
+# Prozess -- so kann kein Zustand aus dem vorigen Aufruf ein Ergebnis tragen.
+# --------------------------------------------------------------------------
+def stdio_ruf(verzeichnis, anfragen):
+    p = subprocess.run(
+        ["lua5.4", os.path.join(verzeichnis, "server.lua")],
+        input="\n".join(json.dumps(a) for a in anfragen) + "\n",
+        capture_output=True, text=True, timeout=60,
+        env=env_fuer("stdio"), cwd=verzeichnis)
+    antworten = []
+    for zeile in (p.stdout or "").splitlines():
+        zeile = zeile.strip()
+        if not zeile:
+            continue
+        try:
+            antworten.append(json.loads(zeile))
+        except json.JSONDecodeError:
+            pass
+    return antworten, (p.stderr or "")
+
+
+def http_ruf(koerper, kopf=None):
+    daten = json.dumps(koerper).encode()
+    h = {"Content-Type": "application/json"}
+    h.update(kopf or {})
+    req = urllib.request.Request(f"http://127.0.0.1:{PORT}/mcp", daten, h)
+    try:
+        with urllib.request.urlopen(req, timeout=30) as r:
+            return r.status, json.loads(r.read().decode())
+    except urllib.error.HTTPError as e:
+        roh = e.read().decode("utf-8", "replace")
+        try:
+            return e.code, json.loads(roh)
+        except json.JSONDecodeError:
+            return e.code, {"raw": roh[:200]}
+
+
+def ruf_alt(n):
+    return {"jsonrpc": "2.0", "id": n, "method": "tools/call",
+            "params": {"name": "shell", "arguments": {"command": BEFEHL}}}
+
+
+def ruf_neu(n, fassung=NEU):
+    return {"jsonrpc": "2.0", "id": n, "method": "tools/call",
+            "params": {"name": "shell", "arguments": {"command": BEFEHL},
+                       "_meta": {
+                           "io.modelcontextprotocol/protocolVersion": fassung,
+                           "io.modelcontextprotocol/clientCapabilities": {}}}}
+
+
+def kopf_neu(methode, name=None, fassung=NEU):
+    h = {"MCP-Protocol-Version": fassung, "Mcp-Method": methode}
+    if name:
+        h["Mcp-Name"] = name
+    return h
+
+
+def text_aus(antwort):
+    """Der Werkzeugtext, egal ob als content[] oder als blanker String."""
+    r = (antwort or {}).get("result") or {}
+    c = r.get("content")
+    if isinstance(c, list):
+        return " ".join(str(t.get("text", "")) for t in c if isinstance(t, dict))
+    return json.dumps(r)
+
+
+def fehler_von(antwort):
+    e = (antwort or {}).get("error") or {}
+    return e.get("code"), e.get("message")
+
+
+# ==========================================================================
+print(f"\n=== lmcp aus {DIR} ===")
+
+# --- 1+2: HTTP -------------------------------------------------------------
+srv = subprocess.Popen(["lua5.4", os.path.join(DIR, "server.lua")],
+                       env=env_fuer("http"), cwd=DIR,
+                       stdout=subprocess.DEVNULL, stderr=subprocess.PIPE, text=True)
+time.sleep(3)
+try:
+    if srv.poll() is not None:
+        merke("HTTP-Server startet", False, (srv.stderr.read() or "")[:200])
+    else:
+        merke("HTTP-Server startet", True, f"Port {PORT}")
+
+        # 1) ALT ueber HTTP: initialize, dann tools/call. Keine Fassungs-Kopfzeile
+        #    -- "keine Fassung" MUSS als alt gelten (E2).
+        st, a = http_ruf({"jsonrpc": "2.0", "id": 1, "method": "initialize",
+                          "params": {"protocolVersion": ALT, "capabilities": {},
+                                     "clientInfo": {"name": "test", "version": "1"}}})
+        pv = ((a.get("result") or {}).get("protocolVersion"))
+        merke("1a HTTP/alt: initialize", pv is not None, f"HTTP {st}, protocolVersion={pv}")
+
+        st, a = http_ruf(ruf_alt(2))
+        t = text_aus(a)
+        merke("1b HTTP/alt: shell echo", ERWARTET in t, f"HTTP {st}, {t.strip()[:60]!r}")
+
+        # 2) NEU ueber HTTP: kein initialize (SEP-2575), voller Umschlag.
+        st, a = http_ruf(ruf_neu(3), kopf_neu("tools/call", "shell"))
+        t = text_aus(a)
+        merke("2  HTTP/neu: shell echo", ERWARTET in t, f"HTTP {st}, {t.strip()[:60]!r}")
+
+        # RED-1: erfundene Fassung muss abgewiesen werden.
+        st, a = http_ruf(ruf_neu(4, "1999-01-01"), kopf_neu("tools/call", "shell", "1999-01-01"))
+        code, msg = fehler_von(a)
+        abgewiesen = code is not None and ERWARTET not in text_aus(a)
+        merke("R1 HTTP: erfundene Fassung wird abgewiesen", abgewiesen,
+              f"HTTP {st}, code={code} {str(msg)[:50]!r}")
+finally:
+    srv.terminate()
+    try:
+        srv.wait(timeout=10)
+    except subprocess.TimeoutExpired:
+        srv.kill()
+
+# --- 3+4: stdio ------------------------------------------------------------
+# 3) ALT: initialize + tools/call in EINEM Prozess, wie ein echter Klient.
+antw, err = stdio_ruf(DIR, [
+    {"jsonrpc": "2.0", "id": 1, "method": "initialize",
+     "params": {"protocolVersion": ALT, "capabilities": {},
+                "clientInfo": {"name": "test", "version": "1"}}},
+    ruf_alt(2)])
+nach_id = {a.get("id"): a for a in antw}
+c1, m1 = fehler_von(nach_id.get(1))
+merke("3a stdio/alt: initialize", c1 is None and nach_id.get(1) is not None,
+      f"code={c1} {str(m1)[:60]!r}" if c1 else "ohne Fehler")
+t = text_aus(nach_id.get(2))
+merke("3b stdio/alt: shell echo", ERWARTET in t, f"{t.strip()[:60]!r}")
+
+# 4) NEU ueber stdio: kein Umschlag (nur HTTP prueft ihn), Fassung im _meta.
+antw, err = stdio_ruf(DIR, [ruf_neu(1)])
+nach_id = {a.get("id"): a for a in antw}
+t = text_aus(nach_id.get(1))
+merke("4  stdio/neu: shell echo", ERWARTET in t, f"{t.strip()[:60]!r}")
+
+# ==========================================================================
+schlecht = [n for n, ok, _ in ergebnisse if not ok]
+print(f"\n  {len(ergebnisse) - len(schlecht)} von {len(ergebnisse)} bestanden")
+if schlecht:
+    print("  offen: " + ", ".join(schlecht))
+sys.exit(1 if schlecht else 0)

tests/konformitaet_2026_07_28.py

@@ -0,0 +1,234 @@
+#!/usr/bin/env python3
+"""Conformance suite for MCP 2026-07-28, cut from the specification itself.
+
+Sources fetched 2026-08-09 and quoted per check:
+  /specification/2026-07-28/                        (index)
+  /specification/2026-07-28/basic                   (messages, _meta, error codes)
+  /specification/2026-07-28/basic/versioning        (version negotiation)
+  /specification/2026-07-28/server/discover         (DiscoverResult shape)
+  /specification/2026-07-28/basic/transports/streamable-http  (headers, statuses)
+
+NOT derived from the changelog or from any second-hand delta: two field names in
+the earlier draft were guessed wrong (`protocolVersions` instead of
+`supportedVersions`, `serverInfo` at top level instead of inside `_meta`).
+
+The spec normalises BOTH the JSON-RPC error code and the HTTP status, so both
+are checked. A check that only looks at the body would pass a server that
+answers 200 where the spec demands 400.
+
+Expected result on an unmigrated server: almost everything OFFEN. That is the
+measurement — the distance to the target — not a failure of the suite.
+
+    python3 konf2026.py <pfad/zu/lmcp.lua>
+"""
+import json
+import os
+import socket
+import subprocess
+import sys
+import time
+import urllib.error
+import urllib.request
+
+ZIEL = "2026-07-28"
+KLON = (os.path.dirname(os.path.abspath(sys.argv[1])) if len(sys.argv) > 1
+        else "/home/mfritsche/src/lmcp-work")
+
+_offen = 0
+def pruefe(punkt, was, ok, beleg=None, bekommen=None):
+    global _offen
+    if not ok:
+        _offen += 1
+    print(f"  [{'ok    ' if ok else 'OFFEN '}] {punkt:<7} {was}")
+    if not ok and bekommen is not None:
+        print(f"            bekommen: {bekommen}")
+    if not ok and beleg:
+        print(f"            Spec: {beleg}")
+
+
+META = {"io.modelcontextprotocol/protocolVersion": ZIEL,
+        "io.modelcontextprotocol/clientCapabilities": {},
+        "io.modelcontextprotocol/clientInfo": {"name": "konf", "version": "1"}}
+
+
+def freier_port():
+    s = socket.socket(); s.bind(("127.0.0.1", 0)); p = s.getsockname()[1]; s.close(); return p
+
+
+def ruf(port, token, methode, params=None, kopf=None, meta=META, art="POST"):
+    """Gibt (jsonrpc-objekt, http-status) zurueck. Beides zaehlt."""
+    p = dict(params or {})
+    if meta is not None:
+        p["_meta"] = meta
+    rumpf = {"jsonrpc": "2.0", "id": 1, "method": methode}
+    if p:
+        rumpf["params"] = p
+    k = {"Content-Type": "application/json",
+         "Accept": "application/json, text/event-stream",
+         "Authorization": "Bearer " + token,
+         "MCP-Protocol-Version": ZIEL,
+         "Mcp-Method": methode}
+    if kopf is not None:
+        for name, wert in kopf.items():
+            if wert is None:
+                k.pop(name, None)
+            else:
+                k[name] = wert
+    req = urllib.request.Request(f"http://127.0.0.1:{port}/mcp",
+                                 json.dumps(rumpf).encode(), k, method=art)
+    try:
+        a = urllib.request.urlopen(req, timeout=20)
+        roh, status = a.read().decode(), a.status
+    except urllib.error.HTTPError as e:
+        roh, status = e.read().decode(), e.code
+    except Exception as e:
+        return {"_transport": f"{type(e).__name__}: {e}"}, 0
+    for z in roh.splitlines():
+        if z.startswith("data: "):
+            roh = z[6:]; break
+    try:
+        return json.loads(roh), status
+    except Exception:
+        return {"_roh": roh[:160]}, status
+
+
+def code(a):
+    return (a.get("error") or {}).get("code")
+
+
+port, token = freier_port(), "konf-2026-07-28"
+umg = dict(os.environ, LMCP_PORT=str(port), LMCP_TOKEN=token, LMCP_HOST="127.0.0.1",
+           LUA_PATH=f"{KLON}/?.lua;;")
+srv = subprocess.Popen(["lua5.4", f"{KLON}/server.lua"], cwd=KLON, env=umg,
+                       stdout=subprocess.DEVNULL, stderr=subprocess.PIPE)
+try:
+    for _ in range(60):
+        time.sleep(0.5)
+        try:
+            socket.create_connection(("127.0.0.1", port), 1).close(); break
+        except OSError:
+            if srv.poll() is not None:
+                print("  Server startete nicht:", (srv.stderr.read() or b"").decode()[:300])
+                sys.exit(2)
+    else:
+        print("  Server kam nicht hoch"); sys.exit(2)
+
+    print(f"  MCP {ZIEL} — Konformitaet von {KLON}\n")
+
+    # ---- server/discover ---------------------------------------------------
+    a, st = ruf(port, token, "server/discover")
+    r = a.get("result") or {}
+    pruefe("discover", "server/discover wird beantwortet", "result" in a,
+           "server/discover: „Servers MUST implement it.\"", json.dumps(a)[:130])
+    pruefe("discover", "result.supportedVersions ist eine nicht-leere Liste",
+           isinstance(r.get("supportedVersions"), list) and bool(r.get("supportedVersions")),
+           "DiscoverResult: supportedVersions", repr(r.get("supportedVersions")))
+    pruefe("discover", "result.capabilities ist ein Objekt",
+           isinstance(r.get("capabilities"), dict), None, repr(r.get("capabilities")))
+    pruefe("discover", "_meta traegt io.modelcontextprotocol/serverInfo",
+           isinstance((r.get("_meta") or {}).get("io.modelcontextprotocol/serverInfo"), dict),
+           "„Servers SHOULD include this field.\"", repr((r.get("_meta") or {}))[:90])
+    pruefe("discover", f"supportedVersions enthaelt {ZIEL}",
+           ZIEL in (r.get("supportedVersions") or []), None,
+           repr(r.get("supportedVersions")))
+
+    # ---- resultType auf JEDEM Ergebnis -------------------------------------
+    for m in ("server/discover", "tools/list"):
+        a, _ = ruf(port, token, m)
+        rt = (a.get("result") or {}).get("resultType")
+        pruefe("result", f"{m}: result.resultType == \"complete\"", rt == "complete",
+               "„The result MUST include a resultType field.\"", repr(rt))
+
+    # ---- _meta ist Pflicht --------------------------------------------------
+    a, st = ruf(port, token, "tools/list", meta=None)
+    pruefe("_meta", "fehlendes _meta -> -32602", code(a) == -32602,
+           "„A request missing any required field is malformed; the server MUST "
+           "reject it with -32602.\"", f"code={code(a)}")
+    pruefe("_meta", "fehlendes _meta -> HTTP 400", st == 400,
+           "„On HTTP, the response status MUST be 400 Bad Request.\"", f"HTTP {st}")
+
+    ohne_caps = {k: v for k, v in META.items()
+                 if k != "io.modelcontextprotocol/clientCapabilities"}
+    a, st = ruf(port, token, "tools/list", meta=ohne_caps)
+    pruefe("_meta", "fehlende clientCapabilities -> -32602", code(a) == -32602,
+           "clientCapabilities: Required = Yes", f"code={code(a)}")
+
+    # ---- Fassungsaushandlung ------------------------------------------------
+    schlecht = dict(META, **{"io.modelcontextprotocol/protocolVersion": "1900-01-01"})
+    a, st = ruf(port, token, "tools/list", meta=schlecht,
+                kopf={"MCP-Protocol-Version": "1900-01-01"})
+    pruefe("version", "unbekannte Fassung -> -32022", code(a) == -32022,
+           "„it MUST respond with an UnsupportedProtocolVersionError\"", f"code={code(a)}")
+    pruefe("version", "unbekannte Fassung -> HTTP 400", st == 400,
+           "„MUST respond with 400 Bad Request and an UnsupportedProtocolVersionError\"",
+           f"HTTP {st}")
+    d = (a.get("error") or {}).get("data") or {}
+    pruefe("version", "Fehler nennt data.supported", isinstance(d.get("supported"), list),
+           None, repr(d)[:90])
+    pruefe("version", "Fehler nennt data.requested == angefragte Fassung",
+           d.get("requested") == "1900-01-01",
+           "Beispiel der Spec: data.requested: \"1900-01-01\"", repr(d.get("requested")))
+
+    # ---- Koepfe -------------------------------------------------------------
+    a, st = ruf(port, token, "tools/list", kopf={"Mcp-Method": None})
+    pruefe("header", "fehlender Mcp-Method -> -32020", code(a) == -32020,
+           "„A required standard header … is missing.\" -> HeaderMismatch", f"code={code(a)}")
+    pruefe("header", "fehlender Mcp-Method -> HTTP 400", st == 400,
+           "„servers MUST return HTTP status 400 Bad Request\"", f"HTTP {st}")
+
+    a, st = ruf(port, token, "tools/list", kopf={"Mcp-Method": "prompts/list"})
+    pruefe("header", "Mcp-Method != method -> -32020", code(a) == -32020,
+           "„values specified in the headers do not match … MUST reject\"", f"code={code(a)}")
+
+    a, st = ruf(port, token, "tools/list", kopf={"MCP-Protocol-Version": None})
+    pruefe("header", "fehlender MCP-Protocol-Version -> -32020", code(a) == -32020,
+           "„Every POST request … MUST include an MCP-Protocol-Version header.\"",
+           f"code={code(a)}")
+
+    a, st = ruf(port, token, "tools/list", kopf={"MCP-Protocol-Version": "2025-06-18"})
+    pruefe("header", "Kopf-Fassung != _meta-Fassung -> -32020", code(a) == -32020,
+           "„If the values do not match, the server MUST reject … HeaderMismatch\"",
+           f"code={code(a)}")
+
+    # ---- unbekannte Methode -------------------------------------------------
+    a, st = ruf(port, token, "gibt/esnicht")
+    pruefe("method", "unbekannte Methode -> -32601", code(a) == -32601, None, f"code={code(a)}")
+    pruefe("method", "unbekannte Methode -> HTTP 404", st == 404,
+           "„it MUST respond with 404 Not Found and a JSON-RPC error with code -32601\"",
+           f"HTTP {st}")
+
+    # ---- gestrichene Methoden ----------------------------------------------
+    for m in ("initialize", "ping", "logging/setLevel"):
+        a, _ = ruf(port, token, m)
+        pruefe("entfernt", f"{m} existiert nicht mehr (-32601)", code(a) == -32601,
+               "SEP-2575: entfernt", f"code={code(a)}")
+
+    # ---- verbotene Fehlercodes ---------------------------------------------
+    a, _ = ruf(port, token, "resources/read", {"uri": "file:///gibtsnicht"},
+               kopf={"Mcp-Name": "file:///gibtsnicht"})
+    pruefe("codes", "unbekannte Ressource -> -32602, NICHT -32002", code(a) == -32602,
+           "„Implementations of this protocol version MUST NOT emit … -32002\"",
+           f"code={code(a)}")
+
+    # ---- GET/DELETE auf dem Endpunkt ---------------------------------------
+    for art in ("GET", "DELETE"):
+        try:
+            req = urllib.request.Request(f"http://127.0.0.1:{port}/mcp", method=art,
+                                         headers={"Authorization": "Bearer " + token})
+            st = urllib.request.urlopen(req, timeout=10).status
+        except urllib.error.HTTPError as e:
+            st = e.code
+        except Exception:
+            st = 0
+        pruefe("transport", f"HTTP {art} auf dem Endpunkt -> 405", st == 405,
+               "„respond with 405 Method Not Allowed\"", f"HTTP {st}")
+
+    print()
+    print(f"  {_offen} Punkte OFFEN")
+    sys.exit(1 if _offen else 0)
+finally:
+    srv.terminate()
+    try:
+        srv.wait(5)
+    except Exception:
+        srv.kill()

tests/phase_a_acceptance.lua

@@ -0,0 +1,113 @@
+-- Abnahmetest fuer lmcp Phase A (nur beobachten, nie ablehnen).
+--
+-- Vom Vertrag geschrieben, NICHT von der Implementierung: der Autor dieses
+-- Tests hat den zu pruefenden Code nicht gesehen. Genau daran sind die
+-- letzten vier Runden gescheitert - der Implementierer hat seine eigenen
+-- Hausaufgaben korrigiert, und der Test prueft dann verlaesslich das, was
+-- der Code ohnehin tut.
+--
+-- VERTRAG
+--   Eine einzelne, in sich geschlossene Lua-5.4-Datei stellt eine Tabelle M
+--   bereit mit:
+--     M.report(version, peer, ua)  -- Produktionseinstieg
+--     M.sink(version, peer, ua)    -- ueberschreibbar, wird beim ERSTEN
+--                                  -- Auftreten eines Tripels gerufen
+--   * jedes verschiedene Tripel (version, peer, ua) genau EINMAL
+--   * version nil oder "" -> nichts
+--   * der Entprellungszustand ist ein privates Upvalue: KEINE globale
+--     Variable, KEIN Parameter, den der Aufrufer mitgeben muss
+--   * gedeckelt bei 50 verschiedenen Tripeln; danach nichts mehr
+--   * M.report lehnt NIE etwas ab und liefert immer nil
+--
+-- Aufruf:  lua5.4 phase_a_acceptance.lua <zu-pruefende-datei.lua>
+
+local pfad = arg and arg[1]
+if not pfad then
+    io.stderr:write("usage: lua5.4 phase_a_acceptance.lua <impl.lua>\n")
+    os.exit(2)
+end
+
+local vorher_global = {}
+for k in pairs(_G) do vorher_global[k] = true end
+
+local lade = assert(loadfile(pfad))
+local M = lade()
+if type(M) ~= "table" then
+    io.stderr:write("FAIL: die Datei liefert keine Tabelle zurueck\n")
+    os.exit(1)
+end
+
+local fehler = 0
+local function pruefe(name, bedingung, zusatz)
+    if bedingung then
+        print(("  [ok    ] %s"):format(name))
+    else
+        fehler = fehler + 1
+        print(("  [FAIL  ] %s%s"):format(name, zusatz and ("  — " .. zusatz) or ""))
+    end
+end
+
+-- Faenger einhaengen
+local gesehen = {}
+M.sink = function(v, p, u)
+    gesehen[#gesehen + 1] = { v = v, p = p, u = u }
+end
+local function zuruecksetzen() gesehen = {} end
+
+pruefe("M.report existiert und ist aufrufbar", type(M.report) == "function")
+if type(M.report) ~= "function" then os.exit(1) end
+
+-- 1. nil und Leerstring melden nichts
+zuruecksetzen()
+M.report(nil, "peerA", "uaA")
+M.report("", "peerA", "uaA")
+pruefe("nil und \"\" melden nichts", #gesehen == 0,
+       ("es kamen %d Meldungen"):format(#gesehen))
+
+-- 2. erstes Tripel meldet genau einmal
+zuruecksetzen()
+M.report("2025-06-18", "peerA", "uaA")
+pruefe("erstes Tripel meldet einmal", #gesehen == 1)
+
+-- 3. DER FALL, DER DREI RUNDEN LANG DURCHRUTSCHTE:
+--    Entprellung am ECHTEN Einstieg, ohne dass der Aufrufer Zustand mitgibt.
+zuruecksetzen()
+for _ = 1, 5 do M.report("2025-06-18", "peerA", "uaA") end
+pruefe("fuenf gleiche Tripel -> keine weitere Meldung", #gesehen == 0,
+       ("es kamen %d Meldungen; Zustand ueberlebt den Aufruf nicht")
+       :format(#gesehen))
+
+-- 4. gleiche Fassung, andere Gegenstelle -> meldet wieder
+zuruecksetzen()
+M.report("2025-06-18", "peerB", "uaA")
+pruefe("gleiche Fassung, andere Gegenstelle -> Meldung", #gesehen == 1,
+       "Entprellung nur auf die Fassung wuerde andere Klienten maskieren")
+
+-- 5. gleiche Fassung und Gegenstelle, anderer User-Agent -> meldet wieder
+zuruecksetzen()
+M.report("2025-06-18", "peerB", "uaZ")
+pruefe("anderer User-Agent -> Meldung", #gesehen == 1)
+
+-- 6. liefert immer nil, lehnt nie ab
+local r1 = M.report("2026-07-28", "peerC", "uaC")
+local r2 = M.report("2026-07-28", "peerC", "uaC")
+pruefe("M.report liefert nil (lehnt nie ab)", r1 == nil and r2 == nil)
+
+-- 7. Deckel bei 50
+zuruecksetzen()
+for i = 1, 80 do M.report("v" .. i, "peerD", "uaD") end
+pruefe("Deckel greift bei 50", #gesehen <= 50,
+       ("es kamen %d Meldungen"):format(#gesehen))
+pruefe("Deckel wirft nicht zu frueh", #gesehen >= 40,
+       ("nur %d Meldungen vor dem Deckel"):format(#gesehen))
+
+-- 8. kein globaler Zustand
+local neue = {}
+for k in pairs(_G) do
+    if not vorher_global[k] then neue[#neue + 1] = k end
+end
+pruefe("keine neuen globalen Variablen", #neue == 0,
+       "hinzugekommen: " .. table.concat(neue, ", "))
+
+print(("  %d Pruefungen fehlgeschlagen"):format(fehler))
+os.exit(fehler == 0 and 0 or 1)

tests/phase_b_acceptance.lua

@@ -0,0 +1,195 @@
+-- Abnahmetest fuer Phase B der Angleichung an MCP 2026-07-28.
+--
+-- Phase A war beobachtend: feststellen, welche Fassungen ueberhaupt verlangt
+-- werden. Phase B setzt durch. Gemessen am 2026-08-08 an hertz-tools:8080
+-- antwortet lmcp mit HTTP 200 auf JEDE Fassungsangabe - auch auf 1999-01-01,
+-- eine Fassung, die es nie gab. Es liest den Kopf schlicht nicht.
+--
+-- VERTRAG. Eine in sich geschlossene Lua-5.4-Datei, Tabelle M:
+--
+--   M.SUPPORTED        Liste der Fassungen, die dieser Server spricht.
+--   M.check(version)   -> true, nil                  wenn zulaessig
+--                      -> false, <fehlertabelle>     sonst
+--
+--   1. version == nil oder ""  -> zulaessig. Ein fehlender Kopf ist erlaubt;
+--      der Server nimmt dann seine Grundfassung an. Das ist kein Sonderfall
+--      aus Bequemlichkeit: die sitzungslose Abkuerzung schickt ihn oft nicht,
+--      und sie traegt den meisten Verkehr.
+--   2. Genaue Uebereinstimmung mit einem Eintrag in M.SUPPORTED -> zulaessig.
+--   3. Alles andere -> false plus Fehlertabelle mit code == -32022 und einer
+--      data.supported-Liste, die GENAU M.SUPPORTED entspricht. Ohne die Liste
+--      kann die Gegenstelle nicht nachverhandeln, sie kann nur aufgeben.
+--   4. Wirft nie. Zahl, Tabelle, Wahrheitswert, Funktion - alles beantwortet
+--      sie mit false, nicht mit einem Laufzeitfehler. Ein Server, der an
+--      einem fremden Kopf stirbt, ist schlechter als einer, der ihn ignoriert.
+--   5. Aendert M.SUPPORTED nicht. Ein Aufruf darf die Liste des naechsten
+--      nicht verschieben.
+--   6. Keine neue globale Variable.
+--
+-- Aufruf: lua5.4 phase_b_acceptance.lua <impl.lua>
+
+local impl_path = arg and arg[1]
+if not impl_path then
+    io.stderr:write("usage: lua5.4 phase_b_acceptance.lua <impl.lua>\n")
+    os.exit(2)
+end
+
+-- Globale Variablen VOR dem Laden festhalten, damit Regel 6 pruefbar ist.
+local vorher = {}
+for k in pairs(_G) do vorher[k] = true end
+
+local chunk, lerr = loadfile(impl_path)
+if not chunk then
+    io.stderr:write("kann " .. impl_path .. " nicht laden: " .. tostring(lerr) .. "\n")
+    os.exit(2)
+end
+local ok_load, M = pcall(chunk)
+if not ok_load then
+    io.stderr:write("Laden warf: " .. tostring(M) .. "\n")
+    os.exit(2)
+end
+
+local fehler = 0
+local function pruefe(name, bedingung, detail)
+    if bedingung then
+        print(string.format("[ok    ] %s", name))
+    else
+        fehler = fehler + 1
+        print(string.format("[FEHLER] %s%s", name, detail and ("  -> " .. tostring(detail)) or ""))
+    end
+end
+
+-- 0. Form
+pruefe("M ist eine Tabelle", type(M) == "table", type(M))
+if type(M) ~= "table" then print(fehler .. " Pruefungen fehlgeschlagen"); os.exit(1) end
+pruefe("M.check ist aufrufbar", type(M.check) == "function", type(M.check))
+pruefe("M.SUPPORTED ist eine nicht-leere Liste",
+       type(M.SUPPORTED) == "table" and #M.SUPPORTED >= 1, type(M.SUPPORTED))
+if type(M.check) ~= "function" or type(M.SUPPORTED) ~= "table" then
+    print(fehler .. " Pruefungen fehlgeschlagen"); os.exit(1)
+end
+
+local function ruf(v)
+    local ok, a, b = pcall(M.check, v)
+    return ok, a, b
+end
+
+-- 1. fehlender Kopf
+local ok, zulaessig = ruf(nil)
+pruefe("nil ist zulaessig", ok and zulaessig == true, ok and tostring(zulaessig) or "warf")
+ok, zulaessig = ruf("")
+pruefe("leerer String ist zulaessig", ok and zulaessig == true, ok and tostring(zulaessig) or "warf")
+
+-- 2. bekannte Fassung
+local bekannt = M.SUPPORTED[1]
+ok, zulaessig = ruf(bekannt)
+pruefe("bekannte Fassung " .. tostring(bekannt) .. " ist zulaessig",
+       ok and zulaessig == true, ok and tostring(zulaessig) or "warf")
+
+-- 3. unbekannte Fassung -> -32022 samt Liste
+local ok3, zul3, err3 = ruf("1999-01-01")
+pruefe("unbekannte Fassung wird abgelehnt", ok3 and zul3 == false,
+       ok3 and tostring(zul3) or "warf")
+pruefe("Ablehnung traegt code -32022",
+       ok3 and type(err3) == "table" and err3.code == -32022,
+       ok3 and type(err3) == "table" and tostring(err3.code) or type(err3))
+local liste_ok = false
+if ok3 and type(err3) == "table" and type(err3.data) == "table"
+   and type(err3.data.supported) == "table" then
+    liste_ok = (#err3.data.supported == #M.SUPPORTED)
+    for i = 1, #M.SUPPORTED do
+        if err3.data.supported[i] ~= M.SUPPORTED[i] then liste_ok = false end
+    end
+end
+pruefe("Ablehnung nennt genau M.SUPPORTED", liste_ok)
+
+-- 3b. die Fassung, auf die wir zuwandern, ist noch NICHT zulaessig.
+-- Wer 2026-07-28 durchwinkt, bevor er sie spricht, hat den Fehler nur verschoben.
+local ok3b, zul3b, err3b = ruf("2026-07-28")
+local spricht_neu = false
+for i = 1, #M.SUPPORTED do if M.SUPPORTED[i] == "2026-07-28" then spricht_neu = true end end
+if spricht_neu then
+    pruefe("2026-07-28 steht in SUPPORTED und wird angenommen", ok3b and zul3b == true)
+else
+    pruefe("2026-07-28 wird abgelehnt, solange sie nicht in SUPPORTED steht",
+           ok3b and zul3b == false and type(err3b) == "table" and err3b.code == -32022,
+           ok3b and tostring(zul3b) or "warf")
+end
+
+-- 3c. VERANKERUNG: M.SUPPORTED gegen den LAUFENDEN Server.
+--
+-- Ohne diese Pruefung misst der Test die Liste nur an sich selbst -- und ein
+-- Modul, das eine Fassung beansprucht, die der Server nicht spricht, besteht
+-- ihn glatt. Genau das ist am 2026-08-08 passiert: {"2025-06-18","2025-11-25"}
+-- ergab 14/14, obwohl lmcp nur 2025-06-18 meldet.
+--
+-- Gefragt wird per `initialize`; die Antwort nennt genau EINE protocolVersion,
+-- naemlich die, die dieser Server spricht. M.SUPPORTED muss exakt daraus
+-- bestehen.
+--
+-- Kein Uebersprung, wenn der Server fehlt: eine Abnahme, die ihre zentrale
+-- Eigenschaft nicht pruefen kann, ist keine Abnahme.
+local probe_url   = os.getenv("LMCP_PROBE_URL")
+local probe_token = os.getenv("LMCP_PROBE_TOKEN")
+pruefe("LMCP_PROBE_URL ist gesetzt (ohne Server keine Verankerung)",
+       probe_url ~= nil and probe_url ~= "", tostring(probe_url))
+
+if probe_url and probe_url ~= "" then
+    local rumpf = '{"jsonrpc":"2.0","id":1,"method":"initialize","params":' ..
+                  '{"protocolVersion":"2025-06-18","capabilities":{},' ..
+                  '"clientInfo":{"name":"phase-b-acceptance","version":"1"}}}'
+    local befehl = "curl -s -m 15 -X POST"
+        .. " -H 'Content-Type: application/json'"
+        .. " -H 'Accept: application/json, text/event-stream'"
+    if probe_token and probe_token ~= "" then
+        befehl = befehl .. " -H 'Authorization: Bearer " .. probe_token .. "'"
+    end
+    befehl = befehl .. " -d '" .. rumpf .. "' '" .. probe_url .. "' 2>/dev/null"
+
+    local p = io.popen(befehl)
+    local antwort = p and p:read("*a") or ""
+    if p then p:close() end
+
+    local gemessen = antwort:match('"protocolVersion"%s*:%s*"([^"]+)"')
+    pruefe("Server nennt eine protocolVersion", gemessen ~= nil,
+           (#antwort > 0) and antwort:sub(1, 70) or "keine Antwort")
+
+    if gemessen then
+        local passt = (#M.SUPPORTED == 1) and (M.SUPPORTED[1] == gemessen)
+        pruefe("M.SUPPORTED entspricht GENAU dem, was der Server spricht ("
+               .. gemessen .. ")", passt, table.concat(M.SUPPORTED, ","))
+    end
+end
+
+-- 4. wirft nie
+local fremde = { 42, true, false, {}, print, 0/0 }
+local alle_still, welcher = true, nil
+for _, v in ipairs({42, true, {}, print}) do
+    local okx, zulx = pcall(M.check, v)
+    if not okx or zulx ~= false then alle_still = false; welcher = tostring(v) end
+end
+pruefe("fremde Typen ergeben false statt Laufzeitfehler", alle_still, welcher)
+
+-- 5. SUPPORTED bleibt unangetastet
+local kopie = {}
+for i, v in ipairs(M.SUPPORTED) do kopie[i] = v end
+ruf("1999-01-01"); ruf(bekannt); ruf(nil)
+local unveraendert = (#kopie == #M.SUPPORTED)
+for i = 1, #kopie do if kopie[i] ~= M.SUPPORTED[i] then unveraendert = false end end
+pruefe("M.SUPPORTED wird durch Aufrufe nicht veraendert", unveraendert)
+
+-- 5b. wiederholte Ablehnung bleibt gleich (kein verbrauchbarer Zustand)
+local _, _, e1 = ruf("1999-01-01")
+local _, _, e2 = ruf("1999-01-01")
+pruefe("zweite Ablehnung ist so vollstaendig wie die erste",
+       type(e1) == "table" and type(e2) == "table"
+       and e1.code == e2.code
+       and type(e2.data) == "table" and type(e2.data.supported) == "table")
+
+-- 6. keine neuen Globalen
+local neu = {}
+for k in pairs(_G) do if not vorher[k] then neu[#neu + 1] = tostring(k) end end
+pruefe("keine neuen globalen Variablen", #neu == 0, table.concat(neu, ","))
+
+print(fehler .. " Pruefungen fehlgeschlagen")
+os.exit(fehler == 0 and 0 or 1)

tests/test_tool_allow.lua

@@ -0,0 +1,124 @@
+-- Ausfuehrbare Zusicherung fuer LMCP_TOOL_ALLOW.
+--
+-- Hintergrund: eine lmcp-Instanz konnte ihren Werkzeugsatz nur ERWEITERN.
+-- tools.d-Dateien fuegen hinzu; der Grundstock aus server.lua bringt shell,
+-- write_file und Verwandte mit, und eine Plugin-Datei kann nichts wegnehmen.
+-- Am 2026-08-08 hatte damit jeder Agent mit dem Raum-Token eine Wurzelschale
+-- im Raum-Container -- nachgewiesen: uid=0(root), Schreibzugriff auf
+-- room.jsonl. Das ist keine Einbruchsluecke (eine Sicherheitsdomaene), aber
+-- es macht jede Aussage ueber Rollentrennung unbelegbar.
+--
+-- Geprueft wird an der REGISTRIERUNG, nicht nachtraeglich loeschend: was nicht
+-- auf der Liste steht, entsteht gar nicht -- fuer Built-ins wie fuer Plugins,
+-- heute wie fuer alles, was spaeter dazukommt.
+--
+-- Aufruf: lua5.4 tests/test_tool_allow.lua
+
+local hier = arg[0]:match('(.*/)') or './'
+-- VORNE anhaengen, nicht hinten. Sonst gewinnt die INSTALLIERTE Fassung unter
+-- /usr/share/lua/5.4/lmcp.lua, und der Test prueft nicht den Baum, in dem er
+-- liegt -- gemessen am 2026-08-08: der Test gab rot, obwohl der Code stimmte.
+package.path = hier .. '../?.lua;' .. package.path
+
+local fehler = 0
+local function pruefe(name, bedingung, detail)
+    if bedingung then
+        print(string.format("[ok    ] %s", name))
+    else
+        fehler = fehler + 1
+        print(string.format("[FEHLER] %s%s", name, detail and ("  -> " .. tostring(detail)) or ""))
+    end
+end
+
+local function namen(server)
+    local t = {}
+    for n in pairs(server.tools) do t[#t + 1] = n end
+    table.sort(t)
+    return t
+end
+
+local function enthaelt(liste, wert)
+    for _, v in ipairs(liste) do if v == wert then return true end end
+    return false
+end
+
+-- lmcp frisch laden, damit die Umgebungsvariable beim Anlegen gilt.
+local function frisch()
+    package.loaded['lmcp'] = nil
+    return require('lmcp')
+end
+
+local leer = { type = "object" }
+local function nichts() return "x" end
+
+-- 1. Ohne die Variable aendert sich nichts (Rueckwaertsvertraeglichkeit).
+-- Nur im ELTERNLAUF: im Kind ist die Liste gesetzt, dort waere die Aussage
+-- falsch und der Test wuerde sich selbst widerlegen.
+if os.getenv("LMCP_TOOL_ALLOW") == nil then
+    local lmcp = frisch()
+    local s = lmcp.new("probe-offen", { port = 0 })
+    s:tool("room_say", "d", leer, nichts)
+    s:tool("shell",    "d", leer, nichts)
+    local n = namen(s)
+    pruefe("ohne LMCP_TOOL_ALLOW bleibt alles registriert",
+           enthaelt(n, "room_say") and enthaelt(n, "shell"), table.concat(n, ","))
+end
+
+-- Ab hier mit Liste. lmcp liest sie beim Anlegen der Instanz, also muss sie
+-- VOR lmcp.new() in der Umgebung stehen -- in Lua nur ueber einen Kindprozess
+-- setzbar, deshalb startet der Test sich selbst neu.
+if os.getenv("LMCP_TOOL_ALLOW") == nil then
+    local eigen = arg[0]
+    local rc = os.execute(
+        'LMCP_TOOL_ALLOW="room_say,room_read,lease_acquire" lua5.4 "' .. eigen .. '" --kind')
+    local ok = (rc == true or rc == 0)
+    pruefe("Teillauf mit gesetzter Liste besteht", ok, tostring(rc))
+    print(fehler .. " Pruefungen fehlgeschlagen")
+    os.exit(fehler == 0 and 0 or 1)
+end
+
+-- --- Kindlauf: LMCP_TOOL_ALLOW ist gesetzt -----------------------------------
+do
+    local lmcp = frisch()
+    local s = lmcp.new("probe-eng", { port = 0 })
+
+    -- erlaubt
+    s:tool("room_say",      "d", leer, nichts)
+    s:tool("room_read",     "d", leer, nichts)
+    s:tool("lease_acquire", "d", leer, nichts)
+    -- nicht erlaubt: genau die, die die Wurzelschale ausmachten
+    s:tool("shell",         "d", leer, nichts)
+    s:tool("shell_bg",      "d", leer, nichts)
+    s:tool("write_file",    "d", leer, nichts)
+    s:tool("edit_file",     "d", leer, nichts)
+    s:tool("read_file",     "d", leer, nichts)
+
+    local n = namen(s)
+    pruefe("erlaubte Werkzeuge sind da",
+           enthaelt(n, "room_say") and enthaelt(n, "room_read") and enthaelt(n, "lease_acquire"),
+           table.concat(n, ","))
+    pruefe("shell ist NICHT registriert",      not enthaelt(n, "shell"))
+    pruefe("shell_bg ist NICHT registriert",   not enthaelt(n, "shell_bg"))
+    pruefe("write_file ist NICHT registriert", not enthaelt(n, "write_file"))
+    pruefe("edit_file ist NICHT registriert",  not enthaelt(n, "edit_file"))
+    pruefe("read_file ist NICHT registriert",  not enthaelt(n, "read_file"))
+    pruefe("genau drei Werkzeuge uebrig", #n == 3, table.concat(n, ","))
+
+    -- Der Kern: `tools/list` und `tools/call` lesen DASSELBE Register. Ein
+    -- nicht registriertes Werkzeug ist also nicht bloss unsichtbar, es ist
+    -- nicht rufbar. Waere es nur aus der Liste gefiltert, bliebe es erreichbar.
+    pruefe("verweigertes Werkzeug ist auch nicht aufrufbar",
+           s.tools["shell"] == nil)
+
+    -- Die Registrierung darf nicht werfen: ein Plugin, das ein gesperrtes
+    -- Werkzeug anbietet, soll weiterlaufen, nicht abstuerzen.
+    local ok = pcall(function() s:tool("shell", "d", leer, nichts) end)
+    pruefe("Registrierung eines gesperrten Werkzeugs wirft nicht", ok)
+
+    -- Verkettung muss erhalten bleiben (tool() gibt self zurueck).
+    local zurueck = s:tool("shell", "d", leer, nichts)
+    pruefe("tool() liefert weiterhin self (verkettbar)", zurueck == s)
+end
+
+print(fehler .. " Pruefungen fehlgeschlagen")
+os.exit(fehler == 0 and 0 or 1)

tools.d/boltzmann.lua

@@ -0,0 +1,30 @@
+-- boltzmann-tools plugin: `stash` — fleet persistent-memory CLI wrapper.
+-- Receives (server, run). Runs the stash CLI inside the memory Incus container's
+-- stash-stash-1 docker container (which has NO shell — /stash is invoked directly
+-- as argv; docker exec handles that, no `sh -c` inside the container).
+local server, run = ...
+
+server:tool("stash",
+    "Fleet persistent memory (stash knowledge graph on the memory container). "
+    .. "`command` = a stash subcommand + its args as ONE string. "
+    .. "Examples: command:=\"recall escher plug AIN\"  |  command:=\"facts\"  |  "
+    .. "command:=\"remember 'the NAS is at 192.168.88.10'\"  |  command:=\"namespace list\". "
+    .. "Wrap multi-word text in single quotes. Read subcommands: recall, facts, namespace list, "
+    .. "goal list, context show. Write: remember, forget, consolidate run.",
+    { type = "object", properties = {
+        command = { type = "string",
+            description = "stash subcommand + args, e.g. \"recall <query>\" or \"remember '<text>'\"" },
+    }, required = { "command" } },
+    function(a)
+        local c = tostring(a.command or ""):gsub("^%s+", ""):gsub("%s+$", "")
+        if c == "" then return "Error: command required (e.g. command:=\"recall <query>\")" end
+        return run("incus exec memory -- docker exec stash-stash-1 /stash " .. c, 60)
+    end,
+    { annotations = {
+        title           = "Stash fleet memory",
+        readOnlyHint    = false,
+        destructiveHint = false,
+        idempotentHint  = false,
+        openWorldHint   = true,
+    } }
+)

tools.d/hertz.lua

@@ -0,0 +1,292 @@
+-- /opt/lmcp/tools.d/hertz.lua — hertz-specific tool registrations.
+--
+-- Invoked by the packaged server.lua's tools.d scan (lmcp ≥ v1.2.0, issue #22).
+-- Receives (server, run) — the configured lmcp instance and the
+-- coroutine-aware run() helper. We add hertz-only tools on top of the
+-- packaged generics (shell, read_file, write_file, edit_file, list_dir,
+-- search_files, fetch, web_search, shell_bg).
+--
+-- Previously these all lived in /opt/lmcp/server.lua (a copy-paste fork of
+-- the packaged server.lua). That pattern drifted on every release; now the
+-- packaged file stays canonical and only the genuine hertz-specifics live
+-- here.
+
+local server, run = ...
+
+-- Local helpers — single-host scoped, not worth lifting into the packaged lib.
+local function read_file_raw(path)
+    local f = io.open(path, 'r')
+    if not f then return nil end
+    local c = f:read('*a'); f:close(); return c
+end
+
+local function farad(cmd, timeout)
+    return run("incus exec farad -- sh -c " .. string.format("%q", cmd),
+        timeout or 15)
+end
+
+-- ---- LXD tools ----
+
+server:tool("incus_exec", "Execute a command inside an Incus container on hertz.", {
+    type = "object",
+    properties = {
+        container = { type = "string", description = "Container name" },
+        command   = { type = "string", description = "Command to execute" },
+        timeout   = { type = "integer", default = 30 },
+    },
+    required = { "container", "command" },
+}, function(a)
+    return run(string.format("incus exec %s -- sh -c %q",
+        a.container:gsub("[^%w%-]", ""), a.command), a.timeout or 30)
+end)
+
+server:tool("incus_list", "List all Incus containers on hertz.",
+    { type = "object" },
+    function() return run("incus list -c ns4 -f csv", 10) end)
+
+-- ---- Fritz!Box tools ----
+
+server:tool("fritz", "Execute Fritz!Box TR-064 command (info, hosts, wan, "
+    .. "wol <mac>, reconnect, reboot, reboot-repeater <ip>, routes, route-add, "
+    .. "route-del, services, actions, call).",
+    {
+        type = "object",
+        properties = {
+            command = { type = "string", description = "fritz subcommand and args" },
+        },
+        required = { "command" },
+    },
+    function(a) return run("sudo /root/.local/bin/fritz " .. a.command, 15) end)
+
+-- ---- Network tools ----
+
+server:tool("ping_host", "Check if a host is reachable (1 ICMP ping, 2s timeout).", {
+    type = "object",
+    properties = { host = { type = "string" } },
+    required = { "host" },
+}, function(a)
+    local host = a.host:gsub("[^%w%.%-:]", "")
+    return run("ping -c1 -W2 " .. host, 5)
+end)
+
+server:tool("network_status",
+    "Check reachability of all infrastructure hosts and MCP endpoints.",
+    { type = "object" },
+    function()
+        local script = [[
+hosts="hertz:localhost boltzmann:boltzmann tesla:tesla data:192.168.88.30 broglie:192.168.88.160 higgs:10.170.16.10"
+for entry in $hosts; do
+    name="${entry%%:*}"
+    ip="${entry#*:}"
+    if ping -c1 -W2 "$ip" >/dev/null 2>&1; then
+        status="UP"
+        if [ "$name" != "data" ]; then
+            if timeout 3 bash -c "echo >/dev/tcp/${ip}/8080" 2>/dev/null; then
+                status="UP (MCP ok)"
+            else
+                status="UP (no MCP)"
+            fi
+        fi
+    else
+        status="DOWN"
+    fi
+    printf "%-12s %-20s %s\n" "$name" "$ip" "$status"
+done
+]]
+        return run(script, 30)
+    end)
+
+server:tool("wol_and_wait",
+    "Wake the data server via Fritz!Box WoL and wait until it (or broglie MCP) is reachable.",
+    {
+        type = "object",
+        properties = {
+            mac          = { type = "string",  default = "", description = "MAC address (auto-detected if empty)" },
+            wait_for_mcp = { type = "boolean", default = true, description = "Wait for broglie MCP instead of just ping" },
+            timeout      = { type = "integer", default = 120 },
+        },
+    },
+    function(a)
+        local mac = a.mac
+        if not mac or mac == "" then
+            local hosts_out = run("sudo /root/.local/bin/fritz hosts 2>/dev/null | grep -i 'data\\|192.168.88.30'", 10)
+            mac = hosts_out and hosts_out:match("(%x%x:%x%x:%x%x:%x%x:%x%x:%x%x)") or ""
+            if mac == "" then return "Error: could not detect MAC for data. Provide it manually." end
+        end
+        run("sudo /root/.local/bin/fritz wol " .. mac:gsub("[^%x:]", ""), 10)
+
+        local timeout = a.timeout or 120
+        local check
+        if a.wait_for_mcp == false then
+            check = "ping -c1 -W2 192.168.88.30"
+        else
+            check = "timeout 3 bash -c 'echo >/dev/tcp/192.168.88.160/8080'"
+        end
+        local target = (a.wait_for_mcp == false) and "data" or "broglie:8080"
+
+        local elapsed = 0
+        while elapsed < timeout do
+            os.execute("sleep 5")
+            elapsed = elapsed + 5
+            local rc = os.execute(check)
+            if rc == true or rc == 0 then
+                return string.format("WoL sent to %s. %s reachable after %ds.",
+                    mac, target, elapsed)
+            end
+        end
+        return string.format("WoL sent to %s but %s not reachable after %ds.",
+            mac, target, timeout)
+    end)
+
+-- ---- Proxmox fallback ----
+
+server:tool("pct_exec",
+    "Execute a command in a Proxmox CT on data (SSH fallback when broglie is down).",
+    {
+        type = "object",
+        properties = {
+            ctid    = { type = "string", description = "Container ID (e.g. 108, 165)" },
+            command = { type = "string" },
+            timeout = { type = "integer", default = 30 },
+        },
+        required = { "ctid", "command" },
+    },
+    function(a)
+        local ctid = a.ctid:gsub("[^%d]", "")
+        return run(string.format("ssh -o ConnectTimeout=5 root@data 'pct exec %s -- sh -c %q'",
+            ctid, a.command), a.timeout or 30)
+    end)
+
+-- ---- Home Assistant ----
+
+server:tool("ha_cli",
+    "Run Home Assistant CLI command inside farad (e.g. 'core restart', 'backups list', 'info').",
+    {
+        type = "object",
+        properties = {
+            command  = { type = "string", description = "ha subcommand, e.g. 'core restart', 'supervisor info', 'backups list'" },
+            raw_json = { type = "boolean", default = false, description = "Return raw JSON output" },
+        },
+        required = { "command" },
+    },
+    function(a)
+        local flags = a.raw_json and " --raw-json" or ""
+        return farad("ha " .. a.command .. flags)
+    end)
+
+server:tool("ha_api",
+    "Call Home Assistant Core REST API. Requires token in /opt/lmcp/ha_token on hertz.",
+    {
+        type = "object",
+        properties = {
+            method   = { type = "string", default = "GET" },
+            endpoint = { type = "string", description = "/api/states, /api/services/climate/set_temperature, etc." },
+            body     = { type = "string", description = "JSON body for POST" },
+        },
+        required = { "endpoint" },
+    },
+    function(a)
+        local token = read_file_raw("/opt/lmcp/ha_token")
+        if not token or token == "" then
+            return "Error: no HA token. Create a long-lived token in HA UI → Profile → "
+                .. "Long-Lived Access Tokens, then: echo '<token>' | sudo tee /opt/lmcp/ha_token"
+        end
+        token = token:gsub("%s+", "")
+        local method = (a.method or "GET"):upper()
+        local cmd = string.format(
+            "curl -sf -X %s -H 'Authorization: Bearer %s' -H 'Content-Type: application/json'",
+            method, token)
+        if a.body and a.body ~= "" then
+            cmd = cmd .. " -d " .. string.format("'%s'", a.body:gsub("'", "'\\''"))
+        end
+        cmd = cmd .. " http://localhost:8123" .. a.endpoint
+        return farad(cmd, 15)
+    end)
+
+-- ---- MQTT (Eurotronic thermostats + general) ----
+
+server:tool("mqtt_pub", "Publish an MQTT message (via Mosquitto in farad).", {
+    type = "object",
+    properties = {
+        topic   = { type = "string" },
+        message = { type = "string" },
+        retain  = { type = "boolean", default = false },
+    },
+    required = { "topic", "message" },
+}, function(a)
+    local retain = a.retain and "-r " or ""
+    return farad(string.format("mosquitto_pub %s-t '%s' -m '%s'",
+        retain, a.topic:gsub("'", "'\\''"), a.message:gsub("'", "'\\''")))
+end)
+
+server:tool("mqtt_sub",
+    "Subscribe to MQTT topics and collect messages (via Mosquitto in farad).",
+    {
+        type = "object",
+        properties = {
+            topic   = { type = "string",  description = "Topic pattern, e.g. 'eurotronic/#' or '#'" },
+            count   = { type = "integer", default = 10, description = "Number of messages to collect" },
+            timeout = { type = "integer", default = 5,  description = "Seconds to wait" },
+            verbose = { type = "boolean", default = true, description = "Show topics with messages" },
+        },
+        required = { "topic" },
+    },
+    function(a)
+        local v = a.verbose ~= false and "-v " or ""
+        return farad(string.format("mosquitto_sub %s-t '%s' -C %d -W %d",
+            v, a.topic:gsub("'", "'\\''"), a.count or 10, a.timeout or 5),
+            (a.timeout or 5) + 5)
+    end)
+
+-- ---- Mediagrab (kids show downloader) ----
+
+server:tool("mediagrab",
+    "Manage kids show downloads in doppler container. Commands: list, weekly, "
+    .. "archive, test <url>, add '<json>'",
+    {
+        type = "object",
+        properties = {
+            command = { type = "string", description = "Command: list, weekly, archive, 'test <url>', 'add <json>'" },
+        },
+        required = { "command" },
+    },
+    function(a)
+        return run("incus exec doppler -- python3 /opt/mediagrab/mediagrab.py "
+            .. a.command, 120)
+    end)
+
+-- ---- Fleet wake (pipi: wake-only, no power-off) ----
+server:tool("wake_fleet",
+    "Wake a fleet NUC (pve1..pve4) via Fritz!Box Wake-on-LAN. Powers a node ON only; it cannot power anything off. Node boots in ~30-60s.",
+    {
+        type = "object",
+        properties = {
+            node = { type = "string", description = "Node to wake: '1'..'4' or 'pve1'..'pve4'" },
+        },
+        required = { "node" },
+    },
+    function(a)
+        local node = tostring(a.node or ""):gsub("[^%w]", "")
+        if not node:match("^p?v?e?[1-4]$") then
+            return "Error: node must be 1-4 or pve1-pve4 (got: " .. tostring(a.node) .. ")"
+        end
+        return run("sudo /root/.local/bin/wake-pve " .. node .. " 2>&1", 15)
+    end)
+
+-- ---- apropos: lean facade for stash recall (read-only memory) ----
+server:tool("apropos",
+    "Search shared fleet memory (stash) for facts about the fleet, projects, decisions, and preferences. query = 2-6 words on the topic; limit = max results (default 3). Read-only.",
+    {
+        type = "object",
+        properties = {
+            query = { type = "string",  description = "2-6 words describing what to recall" },
+            limit = { type = "integer", description = "max results, default 3" },
+        },
+        required = { "query" },
+    },
+    function(a)
+        local q = tostring(a.query or ""):gsub("[^%w%s%-%.]", " "):gsub("%s+", " ")
+        if q:gsub("%s","") == "" then return "Error: query required" end
+        local lim = tonumber(a.limit) or 3
+        return run("python3 /opt/lmcp/helpers/stash_recall.py '" .. q .. "' " .. lim, 30)
+    end)

versions.lua

@@ -0,0 +1,53 @@
+local M = {}
+-- versions.lua
+-- Module for version checking according to LMCP contract
+
+-- Both versions are supported. The old version (2025-06-18) is retained
+-- for backward compatibility, while the new version (2026-07-28) is added.
+-- Dual protocol (old + new) is a later phase.
+M.SUPPORTED = {'2025-06-18', '2026-07-28'}
+
+--- Checks whether a version is supported.
+-- @param version The version to check (string or nil)
+-- @return boolean true if supported, otherwise false
+-- @return table|nil error details if not supported
+function M.check(version)
+    -- (a) version == nil or '' -> true, nil
+    if version == nil or version == '' then
+        return true, nil
+    end
+
+    -- (b) exact match in M.SUPPORTED -> true, nil
+    for _, supported_version in ipairs(M.SUPPORTED) do
+        if version == supported_version then
+            return true, nil
+        end
+    end
+
+    -- (c) everything else -> false, { code = -32022, data = { supported = <copy> } }
+    --
+    -- COPY, not M.SUPPORTED itself. Previously the module table leaked out:
+    -- a caller that keeps the error object and err.data.supported[1]
+    -- overwrites, changes the module's list for EVERY subsequent
+    -- call -- demonstrated on 2026-08-09: after
+    -- e.data.supported[1] = 'CAPTURED' check('2025-06-18') is false and
+    -- check('CAPTURED') is true. Contract rule 5 ('A call must not shift the list of
+    -- the next') was thereby broken, and test 5 did not see it
+    -- because it never writes over the returned reference.
+    -- As long as no one called the module, this was theoretical. Since it is attached to
+    -- initialize, it is reachable.
+    local copy = {}
+    for i = 1, #M.SUPPORTED do
+        copy[i] = M.SUPPORTED[i]
+    end
+    return false, {
+        code = -32022,
+        data = {
+            supported = copy
+        }
+    }
+end
+
+-- No new globals, no side effects, M.SUPPORTED is not modified
+
+return M

windows/README.md

@@ -0,0 +1,57 @@
+# lmcp Windows MSI build
+
+This directory contains the WiX manifest and packaging files for the
+Windows MSI build of lmcp.
+
+## Recommended: cross-build on Linux (one command)
+
+```sh
+./build-msi.sh /path/to/output/dir
+```
+
+Downloads Lua 5.4 Win64 binaries from LuaBinaries, cross-compiles
+LuaSocket via `mingw-w64`, stages `pkg/lua/`, and runs `wixl` to
+produce `lmcp-<version>.msi`. No Windows VM required.
+
+Prereqs on a Debian/Ubuntu builder:
+```sh
+sudo apt install wixl unzip gcc-mingw-w64-x86-64 \
+                 binutils-mingw-w64-x86-64 mingw-w64-x86-64-dev curl
+```
+
+Version comes from `lmcp.wxs` `Version="…"`. Bump that before
+building a release.
+
+## Alternative: build on Windows via WiX toolset
+
+```cmd
+sync.sh                                  REM see "tracked vs. generated"
+REM ensure pkg/lua/ has the runtime — see below
+candle.exe lmcp.wxs
+light.exe lmcp.wixobj -o lmcp-1.x.y.msi
+```
+
+## What's tracked vs. generated
+
+- **Tracked** (edit in git):
+  - `lmcp.wxs` — WiX MSI manifest
+  - `sync.sh` — copies root .lua sources → `pkg/`
+  - `README.md` — this file
+  - `pkg/install_service.bat` — Windows service installer
+  - `pkg/start.bat` — manual launcher
+
+- **Generated / external** (gitignored):
+  - `pkg/lmcp.lua`, `pkg/server.lua`, `pkg/json.lua` — produced by
+    `sync.sh`. Never edit directly; edit the root files and re-sync.
+  - `pkg/lua/` — the Lua + LuaSocket runtime drop-in. Download
+    separately and place here. Suggested source: the lua-binaries
+    project (https://github.com/rjpcomputing/luaforwindows) or a
+    similar pre-built bundle. The MSI expects `pkg/lua/lua.exe`,
+    `pkg/lua/lua54.dll`, and the `pkg/lua/socket/` + `pkg/lua/mime/`
+    subdirectories per the manifest.
+
+## Issue history
+
+Issue #18 (closed in v1.1.0) introduced this workflow after the
+`pkg/` lua sources had silently drifted ~6 months out of date,
+missing every feature added since April 2026.

windows/build-msi.sh

@@ -0,0 +1,100 @@
+#!/bin/sh
+# windows/build-msi.sh — produce lmcp-<ver>.msi on Linux via wixl.
+#
+# This is the first-time-discovered cross-build path: download Lua 5.4
+# Win64 binaries from LuaBinaries, cross-compile LuaSocket with mingw-w64,
+# stage windows/pkg/lua/, then invoke wixl on the WiX manifest.
+#
+# Avoids the VM106-clone + WiX-on-Windows path entirely. ~1 minute on a
+# warm cache; ~3-5 minutes cold (downloads ~700 KB + cross-compiles).
+#
+# Prereqs (apt install on Debian aarch64):
+#   apt-get install -y wixl unzip gcc-mingw-w64-x86-64 binutils-mingw-w64-x86-64 \
+#                      mingw-w64-x86-64-dev
+#
+# Usage: ./build-msi.sh [output_dir]
+#   Output: $output_dir/lmcp-<ver>.msi (default: $PWD)
+#
+# Version comes from windows/lmcp.wxs Version="…" attribute.
+set -eu
+
+here=$(dirname "$(readlink -f "$0")")
+root=$(cd "$here/.." && pwd)
+out_dir=${1:-$PWD}
+work=$(mktemp -d /tmp/lmcp-msi-XXXXXX)
+trap "rm -rf $work" EXIT
+
+# Versions — bump as upstream releases.
+LUA_VER=5.4.2
+LUASOCKET_VER=3.1.0
+
+# Pull current lmcp version from the WiX manifest.
+lmcp_ver=$(sed -n 's/.*Version="\([^"]*\)".*/\1/p' "$here/lmcp.wxs" | head -1)
+[ -n "$lmcp_ver" ] || { echo "build-msi.sh: cannot parse Version from lmcp.wxs" >&2; exit 1; }
+
+echo "build-msi.sh: lmcp $lmcp_ver, lua $LUA_VER, luasocket $LUASOCKET_VER"
+
+echo "==> 1/5 sync lmcp .lua sources into pkg/"
+"$here/sync.sh"
+
+echo "==> 2/5 fetch lua $LUA_VER win64 binaries + dev library"
+cd "$work"
+curl -sSLf -o lua-bin.zip \
+    "https://downloads.sourceforge.net/project/luabinaries/${LUA_VER}/Tools%20Executables/lua-${LUA_VER}_Win64_bin.zip"
+curl -sSLf -o lua-lib.zip \
+    "https://downloads.sourceforge.net/project/luabinaries/${LUA_VER}/Windows%20Libraries/Dynamic/lua-${LUA_VER}_Win64_dllw6_lib.zip"
+mkdir -p luabin lualib include/lua/54 include/lua54 bin/lua/54 bin/lua54 lib/lua/54 lib/lua54
+unzip -q -o lua-bin.zip -d luabin
+unzip -q -o lua-lib.zip -d lualib
+cp lualib/include/*.h include/lua/54/
+cp lualib/include/*.h include/lua54/
+cp lualib/liblua54.a lib/lua/54/
+cp lualib/liblua54.a lib/lua54/
+cp lualib/lua54.dll bin/lua/54/
+cp lualib/lua54.dll bin/lua54/
+
+echo "==> 3/5 cross-compile LuaSocket $LUASOCKET_VER for win64"
+curl -sSLf -o luasocket.tar.gz \
+    "https://github.com/lunarmodules/luasocket/archive/refs/tags/v${LUASOCKET_VER}.tar.gz"
+tar xzf luasocket.tar.gz
+cd "luasocket-${LUASOCKET_VER}"
+make -s PLAT=mingw \
+    CC=x86_64-w64-mingw32-gcc \
+    LD=x86_64-w64-mingw32-gcc \
+    LUAV=54 \
+    LUAINC_mingw_base="$work/include" \
+    LUALIB_mingw_base="$work/bin" \
+    > /dev/null
+
+echo "==> 4/5 stage pkg/lua/"
+pkg_lua="$here/pkg/lua"
+rm -rf "$pkg_lua"
+mkdir -p "$pkg_lua/socket" "$pkg_lua/mime"
+
+# WiX manifest expects "lua.exe" (not "lua54.exe").
+cp "$work/luabin/lua54.exe"        "$pkg_lua/lua.exe"
+cp "$work/luabin/lua54.dll"        "$pkg_lua/lua54.dll"
+cp src/socket.lua                  "$pkg_lua/"
+cp src/mime.lua                    "$pkg_lua/"
+cp src/ltn12.lua                   "$pkg_lua/"
+cp src/socket-3.0.0.dll            "$pkg_lua/socket/core.dll"
+cp src/ftp.lua                     "$pkg_lua/socket/"
+cp src/headers.lua                 "$pkg_lua/socket/"
+cp src/http.lua                    "$pkg_lua/socket/"
+cp src/smtp.lua                    "$pkg_lua/socket/"
+cp src/tp.lua                      "$pkg_lua/socket/"
+cp src/url.lua                     "$pkg_lua/socket/"
+cp src/mime-1.0.3.dll              "$pkg_lua/mime/core.dll"
+
+echo "==> 5/5 wixl: produce MSI"
+# wixl wants forward slashes; rewrite Windows-style backslashes in Source=.
+wxs_tmp="$work/lmcp.wxs"
+sed 's|Source="pkg\\|Source="pkg/|g; s|\\\([a-zA-Z]\)|/\1|g' "$here/lmcp.wxs" > "$wxs_tmp"
+mkdir -p "$out_dir"
+out_msi="$out_dir/lmcp-${lmcp_ver}.msi"
+(cd "$here" && wixl -v "$wxs_tmp" -o "$out_msi")
+
+echo ""
+echo "==> done: $out_msi"
+ls -la "$out_msi"
+sha256sum "$out_msi"

windows/lmcp.wxs

@@ -0,0 +1,119 @@
+<?xml version="1.0" encoding="utf-8"?>
+<Wix xmlns="http://schemas.microsoft.com/wix/2006/wi">
+  <!-- Bump Version on every release. See windows/README.md. -->
+  <Product Id="*"
+           Name="lmcp — Lua MCP Server"
+           Language="1033"
+           Version="1.1.0"
+           Manufacturer="QAP'LA Project"
+           UpgradeCode="A7F3E2D1-4B5C-6D7E-8F9A-0B1C2D3E4F5A">
+
+    <Package InstallerVersion="200"
+             Compressed="yes"
+             InstallScope="perMachine"
+             Description="Lightweight MCP server in Lua. 2MB RSS."
+             Comments="Zero-dependency MCP server." />
+
+    <MediaTemplate EmbedCab="yes" />
+
+    <MajorUpgrade DowngradeErrorMessage="A newer version is already installed." />
+
+    <Directory Id="TARGETDIR" Name="SourceDir">
+      <Directory Id="ProgramFiles64Folder">
+        <Directory Id="INSTALLFOLDER" Name="lmcp">
+          <Directory Id="LUA_DIR" Name="lua">
+            <Directory Id="SOCKET_DIR" Name="socket" />
+            <Directory Id="MIME_DIR" Name="mime" />
+          </Directory>
+        </Directory>
+      </Directory>
+    </Directory>
+
+    <!-- lmcp application files -->
+    <DirectoryRef Id="INSTALLFOLDER">
+      <Component Id="JsonLua" Guid="B1A2C3D4-E5F6-7890-ABCD-EF1234567890">
+        <File Id="json.lua" Source="pkg\json.lua" KeyPath="yes" />
+      </Component>
+      <Component Id="LmcpLua" Guid="B1A2C3D4-E5F6-7890-ABCD-EF1234567891">
+        <File Id="lmcp.lua" Source="pkg\lmcp.lua" KeyPath="yes" />
+      </Component>
+      <Component Id="ServerLua" Guid="B1A2C3D4-E5F6-7890-ABCD-EF1234567892">
+        <File Id="server.lua" Source="pkg\server.lua" KeyPath="yes" />
+      </Component>
+      <Component Id="StartBat" Guid="B1A2C3D4-E5F6-7890-ABCD-EF1234567893">
+        <File Id="start.bat" Source="pkg\start.bat" KeyPath="yes" />
+      </Component>
+      <Component Id="InstallService" Guid="B1A2C3D4-E5F6-7890-ABCD-EF1234567894">
+        <File Id="install_service.bat" Source="pkg\install_service.bat" KeyPath="yes" />
+      </Component>
+    </DirectoryRef>
+
+    <!-- Lua runtime -->
+    <DirectoryRef Id="LUA_DIR">
+      <Component Id="LuaExe" Guid="C2B3D4E5-F6A7-8901-BCDE-F12345678900">
+        <File Id="lua.exe" Source="pkg\lua\lua.exe" KeyPath="yes" />
+      </Component>
+      <Component Id="LuaDll" Guid="C2B3D4E5-F6A7-8901-BCDE-F12345678901">
+        <File Id="lua54.dll" Source="pkg\lua\lua54.dll" KeyPath="yes" />
+      </Component>
+      <Component Id="SocketLua" Guid="C2B3D4E5-F6A7-8901-BCDE-F12345678902">
+        <File Id="socket.lua" Source="pkg\lua\socket.lua" KeyPath="yes" />
+      </Component>
+      <Component Id="MimeLua" Guid="C2B3D4E5-F6A7-8901-BCDE-F12345678903">
+        <File Id="mime.lua" Source="pkg\lua\mime.lua" KeyPath="yes" />
+      </Component>
+      <Component Id="Ltn12Lua" Guid="C2B3D4E5-F6A7-8901-BCDE-F12345678904">
+        <File Id="ltn12.lua" Source="pkg\lua\ltn12.lua" KeyPath="yes" />
+      </Component>
+    </DirectoryRef>
+
+    <!-- LuaSocket native DLLs -->
+    <DirectoryRef Id="SOCKET_DIR">
+      <Component Id="SocketCoreDll" Guid="D3C4E5F6-A7B8-9012-CDEF-123456789010">
+        <File Id="socket_core.dll" Name="core.dll" Source="pkg\lua\socket\core.dll" KeyPath="yes" />
+      </Component>
+      <Component Id="SocketFtp" Guid="D3C4E5F6-A7B8-9012-CDEF-123456789011">
+        <File Id="ftp.lua" Source="pkg\lua\socket\ftp.lua" KeyPath="yes" />
+      </Component>
+      <Component Id="SocketHeaders" Guid="D3C4E5F6-A7B8-9012-CDEF-123456789012">
+        <File Id="headers.lua" Source="pkg\lua\socket\headers.lua" KeyPath="yes" />
+      </Component>
+      <Component Id="SocketHttp" Guid="D3C4E5F6-A7B8-9012-CDEF-123456789013">
+        <File Id="http.lua" Source="pkg\lua\socket\http.lua" KeyPath="yes" />
+      </Component>
+      <Component Id="SocketTp" Guid="D3C4E5F6-A7B8-9012-CDEF-123456789014">
+        <File Id="tp.lua" Source="pkg\lua\socket\tp.lua" KeyPath="yes" />
+      </Component>
+      <Component Id="SocketUrl" Guid="D3C4E5F6-A7B8-9012-CDEF-123456789015">
+        <File Id="url.lua" Source="pkg\lua\socket\url.lua" KeyPath="yes" />
+      </Component>
+    </DirectoryRef>
+
+    <DirectoryRef Id="MIME_DIR">
+      <Component Id="MimeCoreDll" Guid="E4D5F6A7-B8C9-0123-DEFA-234567890120">
+        <File Id="mime_core.dll" Name="core.dll" Source="pkg\lua\mime\core.dll" KeyPath="yes" />
+      </Component>
+    </DirectoryRef>
+
+    <Feature Id="MainFeature" Title="lmcp Server" Level="1">
+      <ComponentRef Id="JsonLua" />
+      <ComponentRef Id="LmcpLua" />
+      <ComponentRef Id="ServerLua" />
+      <ComponentRef Id="StartBat" />
+      <ComponentRef Id="InstallService" />
+      <ComponentRef Id="LuaExe" />
+      <ComponentRef Id="LuaDll" />
+      <ComponentRef Id="SocketLua" />
+      <ComponentRef Id="MimeLua" />
+      <ComponentRef Id="Ltn12Lua" />
+      <ComponentRef Id="SocketCoreDll" />
+      <ComponentRef Id="SocketFtp" />
+      <ComponentRef Id="SocketHeaders" />
+      <ComponentRef Id="SocketHttp" />
+      <ComponentRef Id="SocketTp" />
+      <ComponentRef Id="SocketUrl" />
+      <ComponentRef Id="MimeCoreDll" />
+    </Feature>
+
+  </Product>
+</Wix>

windows/pkg/install_service.bat

@@ -0,0 +1,24 @@
+@echo off
+REM Install lmcp as a Windows service using NSSM (Non-Sucking Service Manager)
+REM Download nssm from https://nssm.cc if not present
+
+if not exist "%~dp0nssm.exe" (
+    echo ERROR: nssm.exe not found in %~dp0
+    echo Download from https://nssm.cc and place nssm.exe here.
+    exit /b 1
+)
+
+set INSTALL_DIR=%~dp0
+set SERVICE_NAME=lmcp
+
+echo Installing lmcp as Windows service...
+%INSTALL_DIR%nssm.exe install %SERVICE_NAME% "%INSTALL_DIR%lua\lua.exe" "%INSTALL_DIR%server.lua"
+%INSTALL_DIR%nssm.exe set %SERVICE_NAME% AppDirectory "%INSTALL_DIR%"
+%INSTALL_DIR%nssm.exe set %SERVICE_NAME% AppEnvironmentExtra "LMCP_PORT=8080"
+%INSTALL_DIR%nssm.exe set %SERVICE_NAME% DisplayName "lmcp MCP Server"
+%INSTALL_DIR%nssm.exe set %SERVICE_NAME% Description "Lightweight MCP server in Lua"
+%INSTALL_DIR%nssm.exe set %SERVICE_NAME% Start SERVICE_AUTO_START
+%INSTALL_DIR%nssm.exe start %SERVICE_NAME%
+
+echo Done. Service '%SERVICE_NAME%' installed and started.
+echo Check: sc query %SERVICE_NAME%

windows/pkg/start.bat

@@ -0,0 +1,7 @@
+@echo off
+REM lmcp — Lua MCP Server
+REM Start the server on port 8080 (or LMCP_PORT if set)
+cd /d "%~dp0"
+if not defined LMCP_PORT set LMCP_PORT=8080
+echo Starting lmcp on port %LMCP_PORT%...
+lua\lua.exe server.lua

windows/sync.sh

@@ -0,0 +1,25 @@
+#!/bin/sh
+# windows/sync.sh — refresh windows/pkg/ from root .lua sources (issue #18).
+#
+# Run BEFORE invoking the WiX build so the MSI bundles whatever is in
+# master. The .lua files in windows/pkg/ are regenerated on every run
+# and are gitignored — never edit them directly.
+#
+# Idempotent: re-running just re-copies. Safe to call from a Makefile,
+# a CI step, or by hand before `candle.exe + light.exe`.
+
+set -eu
+
+here=$(dirname "$(readlink -f "$0")")
+root=$(cd "$here/.." && pwd)
+
+for f in lmcp.lua server.lua json.lua; do
+    if [ ! -f "$root/$f" ]; then
+        echo "windows/sync.sh: missing source $root/$f" >&2
+        exit 1
+    fi
+    cp "$root/$f" "$here/pkg/$f"
+    echo "  synced $f"
+done
+
+echo "windows/sync.sh: done — pkg/ matches root .lua at $(date +%Y-%m-%dT%H:%M:%S)"