tep 0.11.2 → 0.11.4

data/test/test_llm_gateway.rb

@@ -51,8 +51,8 @@ class TestLlmGateway < TepTest
         0
       end
       def on_stream_end(req, out, stats)
-        model = Tep::Json.get_str(req.raw_body, "model")
-        extra = "{" + Tep::Json.encode_pair_str("request_id", "req-1") + "}"
+        model = SpinelKit::Json.get_str(req.raw_body, "model")
+        extra = "{" + SpinelKit::Json.encode_pair_str("request_id", "req-1") + "}"
         EVENTS.inference(model, 0, stats.chunk_count, 1000000, extra)
         0
       end

data/test/test_logger.rb

@@ -1,13 +1,13 @@
 require_relative "helper"
 
-# Tep::Logger -- levelled logger with stderr / file output.
+# SpinelKit::Log -- levelled logger with stderr / file output.
 class TestLogger < TepTest
   TMP_LOG = "/tmp/tep_logger_test_#{$$}.log"
 
   app_source <<~RB
     require 'sinatra'
 
-    LOGGER = Tep::Logger.new
+    LOGGER = SpinelKit::Log.new
     LOGGER.set_level("debug")
     LOGGER.to_file("#{TMP_LOG}")
 

data/test/test_openai_server.rb

@@ -89,6 +89,8 @@ class TestOpenAIServer < TepTest
     assert_equal "200", res.code
     body = JSON.parse(res.body)
     assert_equal "text_completion", body["object"]
+    # Backend leaves Completion#id default -> "cmpl-tep" (back-compat).
+    assert_equal "cmpl-tep", body["id"]
     assert_equal "echo-1", body["model"]
     assert_equal "echoed 3 tokens t=1.0 p=1.0", body["choices"][0]["text"]
     assert_equal "stop", body["choices"][0]["finish_reason"]
@@ -131,6 +133,9 @@ class TestOpenAIServerEvents < TepTest
         c.text              = "echoed " + token_ids.length.to_s + " tokens"
         c.prompt_tokens     = token_ids.length
         c.completion_tokens = sampling.max_tokens
+        # Backend-minted per-request id (#209): must surface as the
+        # response `id` AND the inference event's request_id.
+        c.id                = "cmpl-evt-" + token_ids.length.to_s
         c
       end
     end
@@ -168,6 +173,8 @@ class TestOpenAIServerEvents < TepTest
     res = post("/v1/completions",
                "{\"model\":\"echo-1\",\"prompt\":[1,2,3,4],\"max_tokens\":7}")
     assert_equal "200", res.code
+    # Backend-minted id (#209) surfaces as the response `id`.
+    assert_equal "cmpl-evt-4", JSON.parse(res.body)["id"]
 
     lines2 = File.readlines(EVENTS_PATH).map { |l| JSON.parse(l) }
     # #136: inference events are kind:"eval"+name:"request"; per-request
@@ -182,7 +189,9 @@ class TestOpenAIServerEvents < TepTest
     assert_equal 7,        extra["completion_tokens"]
     assert_kind_of Integer, extra["latency_us"]
     assert extra["latency_us"] >= 0
-    assert_equal "cmpl-tep", extra["request_id"]
+    # request_id tracks the backend-minted Completion#id, same value as
+    # the response `id` above (the request_id == response id invariant).
+    assert_equal "cmpl-evt-4", extra["request_id"]
     assert_match(/\Auser:/, extra["principal_id"])
   end
 end
@@ -596,3 +605,65 @@ class TestOpenAIEmbeddings < TepTest
     assert_equal "invalid_request_error", body["error"]["type"]
   end
 end
+
+# IDs-only backend (toy#30 convergence): a backend with no detokenizer
+# returns the generated token IDs in Completion#token_ids. The
+# CompletionsHandler then emits choices[0].ids (text stays ""), honors
+# Completion#finish_reason, and ModelsHandler reflects Backend#model_owner
+# + a created stamp. This is the exact surface toy's serve path adopts to
+# drop its hand-rolled handlers.
+class TestOpenAIServerIdsBackend < TepTest
+  app_source <<~RB
+    require 'sinatra'
+
+    class IdsBackend < Tep::Llm::OpenAI::Backend
+      def list_models
+        ["toy-1"]
+      end
+      def model_owner
+        "toy"
+      end
+      def generate_from_tokens(model, token_ids, sampling)
+        c = Tep::Llm::OpenAI::Completion.new
+        # Echo input IDs +1000 as the "generated" IDs so the test can
+        # assert the ids field round-trips; a real backend decodes.
+        ids = [0]; ids.delete_at(0)
+        i = 0
+        while i < token_ids.length
+          ids.push(token_ids[i] + 1000)
+          i = i + 1
+        end
+        c.token_ids         = ids
+        c.prompt_tokens     = token_ids.length
+        c.completion_tokens = ids.length
+        c.finish_reason     = "length"
+        c
+      end
+    end
+
+    Tep::Llm::OpenAI::Server.use(IdsBackend.new)
+    Tep::Llm::OpenAI::Server.serve!
+  RB
+
+  def test_completions_emit_ids_field
+    res = post("/v1/completions",
+               "{\"model\":\"toy-1\",\"prompt\":[10,20,30],\"max_tokens\":3}")
+    assert_equal "200", res.code
+    body = JSON.parse(res.body)
+    assert_equal "text_completion", body["object"]
+    # Generated IDs surface as choices[0].ids (input + 1000); text is "".
+    assert_equal [1010, 1020, 1030], body["choices"][0]["ids"]
+    assert_equal "", body["choices"][0]["text"]
+    assert_equal "length", body["choices"][0]["finish_reason"]
+    assert_equal 3, body["usage"]["prompt_tokens"]
+    assert_equal 3, body["usage"]["completion_tokens"]
+  end
+
+  def test_models_reflects_backend_owner_and_created
+    body = JSON.parse(get("/v1/models").body)
+    m = body["data"][0]
+    assert_equal "toy-1", m["id"]
+    assert_equal "toy",   m["owned_by"]
+    assert_kind_of Integer, m["created"]
+  end
+end

data/test/test_password.rb

@@ -7,14 +7,14 @@ class TestPassword < TepTest
 
     post '/hash' do
       res.headers["Content-Type"] = "text/plain"
-      pwd = Tep::Json.get_str(req.raw_body, "password")
+      pwd = SpinelKit::Json.get_str(req.raw_body, "password")
       Tep::Password.hash(pwd)
     end
 
     post '/verify' do
       res.headers["Content-Type"] = "text/plain"
-      pwd = Tep::Json.get_str(req.raw_body, "password")
-      hash = Tep::Json.get_str(req.raw_body, "hash")
+      pwd = SpinelKit::Json.get_str(req.raw_body, "password")
+      hash = SpinelKit::Json.get_str(req.raw_body, "hash")
       Tep::Password.verify(pwd, hash) ? "ok" : "bad"
     end
 

data/test/test_real_world.rb

@@ -245,7 +245,12 @@ class TestRealWorld < TepTest
       # With a successful login + cookie jar...
       uri = URI("http://127.0.0.1:#{@port}/login")
       net = Net::HTTP.new(uri.host, uri.port)
-      r_login = net.post(uri.path, "user=alice&password=hunter2")
+      # Explicit form Content-Type: this direct net.post bypasses the
+      # harness req() helper, and Ruby 4.0's Net::HTTP no longer auto-sets
+      # it for a bodied request (3.x did), so without it the login form
+      # isn't parsed -> 401 instead of 302.
+      r_login = net.post(uri.path, "user=alice&password=hunter2",
+                         {"Content-Type" => "application/x-www-form-urlencoded"})
       assert_equal "302", r_login.code
       cookie = r_login["Set-Cookie"]
       refute_nil cookie

data/test/test_shutdown.rb

@@ -0,0 +1,40 @@
+require_relative "helper"
+
+# #188 regression guard for the #186 fix.
+#
+# A server that does NOT mount the OpenAI events surface (the common case)
+# must not SIGSEGV on SIGTERM. The #186 bug: App#initialize never set
+# @openai_events, so Tep.on_shutdown's unconditional `openai_events.enabled?`
+# was a null-receiver deref -- a hard SIGSEGV under Spinel (exit 139) on a
+# clean `kill -TERM`. #186 (0.11.2) initialises @openai_events to a disabled
+# default, so on_shutdown is a safe no-op for apps that never call
+# Tep::Llm::OpenAI::Server.serve!.
+#
+# The events-mounted path is already covered by
+# test_openai_server#test_sigterm_emits_run_end; this is the no-events path
+# that the original bug actually hit. We assert the exit is NOT a SEGV.
+# Whether the clean exit is 143 (signal-terminated) or 0 (graceful return)
+# is the build/timing-sensitive residual tracked in #188 and is acceptable
+# here -- only a SIGSEGV is the regression.
+class TestShutdownNoEvents < TepTest
+  app_source <<~RB
+    require 'sinatra'
+
+    get '/ping' do
+      "pong"
+    end
+  RB
+
+  def test_sigterm_on_no_events_app_does_not_segv
+    assert_equal "pong", get("/ping").body, "server should be live before SIGTERM"
+
+    status = TepHarness.terminate_status(@port)
+    refute_nil status, "spawned server not found / never reaped"
+
+    segv = Signal.list["SEGV"] # 11 -> exit 139
+    crashed = status.signaled? && status.termsig == segv
+    refute crashed,
+           "no-events app SIGSEGV'd on SIGTERM (the #186 regression) -- " \
+           "termsig=#{status.termsig.inspect} exitstatus=#{status.exitstatus.inspect}"
+  end
+end

metadata

@@ -1,14 +1,13 @@
 --- !ruby/object:Gem::Specification
 name: tep
 version: !ruby/object:Gem::Version
-  version: 0.11.2
+  version: 0.11.4
 platform: ruby
 authors:
 - Ori Pekelman
-autorequire:
 bindir: bin
 cert_chain: []
-date: 2026-06-01 00:00:00.000000000 Z
+date: 1980-01-02 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: prism
@@ -88,6 +87,11 @@ files:
 - examples/qdrant/README.md
 - examples/sinatra_style.rb
 - examples/websocket_echo.rb
+- lib/spinel_kit/hex.rb
+- lib/spinel_kit/json.rb
+- lib/spinel_kit/json_decoder.rb
+- lib/spinel_kit/log.rb
+- lib/spinel_kit/url.rb
 - lib/tep.rb
 - lib/tep/agent_delegation.rb
 - lib/tep/app.rb
@@ -107,11 +111,9 @@ files:
 - lib/tep/http.rb
 - lib/tep/identity.rb
 - lib/tep/job.rb
-- lib/tep/json.rb
 - lib/tep/jwt.rb
 - lib/tep/live_view.rb
 - lib/tep/llm.rb
-- lib/tep/logger.rb
 - lib/tep/mcp.rb
 - lib/tep/multipart.rb
 - lib/tep/net.rb
@@ -137,7 +139,6 @@ files:
 - lib/tep/streamer.rb
 - lib/tep/tep_pg.c
 - lib/tep/tep_sqlite.c
-- lib/tep/url.rb
 - lib/tep/version.rb
 - lib/tep/websocket.rb
 - lib/tep/websocket/connection.rb
@@ -215,6 +216,7 @@ files:
 - test/test_server_scheduled.rb
 - test/test_sessions.rb
 - test/test_shell.rb
+- test/test_shutdown.rb
 - test/test_sqlite.rb
 - test/test_sqlite_cached.rb
 - test/test_static.rb
@@ -257,8 +259,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.4.20
-signing_key:
+rubygems_version: 4.0.3
 specification_version: 4
 summary: A Sinatra-flavoured web framework that compiles to a native binary via Spinel
 test_files: []