fix(codex-link): recover Claude thinking-signature rejections the Hermes way

Claude rejects a turn whose latest assistant thinking blocks are not exactly
as produced ("cannot be modified" / "Invalid signature in thinking block").
Following Hermes (_manage_thinking_signatures: a dead-signature turn is
replayed without thinking), the bridge reads the first upstream piece and,
on exactly these messages, resends that one request once with reasoning
effort none before anything reaches Codex. Other errors are not resent.
The rejected request is kept as log/dusunme-400-*.json for diagnosis.

The overflow guard and layer filter still used a flat 2 bytes/token; they
now use the measured ratio. Ship IDE beta.48 (bridge version 27).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XeNgWNzPjzUKQuS32RUYXx

.gitea/workflows/ci.yml

@@ -56,7 +56,7 @@ jobs:
       - name: Desktop browser control lane (AITURK Chrome)
         run: '"$RUNNER_TEMP/ci-venv/bin/python" -m pytest tests/gateway/test_browser_control_local_lane.py tests/gateway/test_browser_control_broker.py tests/gateway/test_browser_control_broker_hardening.py tests/gateway/test_browser_control_cloud.py tests/tools/test_browser_extension_router.py tests/tools/test_browser_extension_router_wiring.py -q -p no:cacheprovider -rf'
       - name: Native watch and speech mux pytest regressions
-        run: '"$RUNNER_TEMP/ci-venv/bin/python" -m pytest apps/desktop/assets/codex-bridge/test_bekci_watch.py apps/desktop/assets/codex-bridge/test_native_speech_mux.py apps/desktop/assets/codex-bridge/test_live_limits.py -q -p no:cacheprovider -rf'
+        run: '"$RUNNER_TEMP/ci-venv/bin/python" -m pytest apps/desktop/assets/codex-bridge/test_bekci_watch.py apps/desktop/assets/codex-bridge/test_native_speech_mux.py apps/desktop/assets/codex-bridge/test_live_limits.py apps/desktop/assets/codex-bridge/test_dusunme_imzasi.py -q -p no:cacheprovider -rf'
       - name: Mobile and desktop startup regression tests
         run: |
           npm install --prefix "$RUNNER_TEMP/mobile-tests" --no-save --ignore-scripts --no-audit --no-fund --legacy-peer-deps vitest@4.1.10 ws@8.21.1

apps/desktop/assets/codex-bridge/README.md

@@ -27,6 +27,28 @@ Run: `python bridge.py --home <data dir> --hermes-home <agent home>`. The IDE wr
 
 Tests: `python -m unittest discover` in this directory.
 
+## Version 27: Claude thinking-signature rejection recovered the Hermes way (IDE beta.48)
+
+- Claude rejects a turn when the thinking blocks of the latest assistant message are not exactly as it produced them
+  ("thinking or redacted_thinking blocks in the latest assistant message cannot be modified", or "Invalid signature
+  in thinking block"). OmniRoute builds those blocks; seen 8 times in 48 h, always while Opus ran parallel sub-agents.
+- Hermes rule (agent/anthropic_message_convert.py, _manage_thinking_signatures): a turn whose signature is dead is
+  replayed without its thinking. The bridge reads the first upstream piece; on exactly these two messages it resends
+  that one request once with `reasoning.effort = none` (history unchanged) before anything reaches Codex. Any other
+  error is not resent. The rejected request is kept for diagnosis as `log/dusunme-400-*.json` (newest 5).
+- Needs the gateway to forward the provider reason (`detail`) in Responses stream errors (gateway 21 Oct 2026).
+- The overflow guard and layer filter used a flat 2 bytes/token; they now use the measured ratio from version 26.
+
+## Version 26: list downloaded once, context counted by measurement (IDE beta.47)
+
+- Model list: the upstream stamp (ETag) is kept. chatgpt.com repeats it in every answer (`x-models-etag`); the
+  bridge's own stamp made Codex re-download the ~17 MB merged list after every answer (6198 times on 20 October
+  2026). Added-model changes still arrive on Codex's own 5-minute list refresh.
+- Token estimate: measured on 673 real requests (bridge bytes against provider-reported input tokens). Claude
+  family and unknown names use 2.5 bytes/token (Claude lowest quartile 2.55); recognised non-Claude families use
+  3.4 (their lowest 5%). The old flat 2 bytes/token counted non-Claude chats about twice as large, summarized too
+  early and raised false "Son mesaj ve zorunlu talimatlar hedefe sigmiyor" errors.
+
 ## Version 25: subscription WebSocket stays original (IDE beta.46)
 
 User decision (20 October 2026): "Codex orjinal aynen kalacak, sadece ek combo ve model ekleyeceğiz."

apps/desktop/assets/codex-bridge/bridge.py

@@ -42,7 +42,7 @@ import gecis_context
 import terminal_stream
 
 HERE = os.path.dirname(os.path.abspath(__file__))
-SURUM = "25"         # AITURK IDE'nin bekledigi kopru surumu (codex-link-config.ts CODEX_BRIDGE_VERSION)
+SURUM = "27"         # AITURK IDE'nin bekledigi kopru surumu (codex-link-config.ts CODEX_BRIDGE_VERSION)
 try:
     with open(os.path.abspath(__file__), "rb") as _f:
         KOD = hashlib.sha256(_f.read()).hexdigest()[:8]

apps/desktop/assets/codex-bridge/gecis_context.py

@@ -68,9 +68,29 @@ def encoded(value):
     return json.dumps(value, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
 
 
+# Bytes per token, measured on 673 real added-model requests (20 October 2026; bridge request bytes against the
+# provider's reported input tokens): Claude family 2.55 at the lowest quartile (median 2.58); every other family
+# 3.4 or more at the lowest 5% (Gemini, GPT, abliteration, Jail-Omni, bekci ~3.7-4.2). The old flat 2 bytes/token
+# counted non-Claude chats about twice as large, summarized too early and raised false "does not fit" errors.
+# Claude, profiles and anything unrecognised keep the safe 2.5; only a recognised non-Claude family gets 3.4.
+SAFE_BYTES_PER_TOKEN = 2.5
+NON_CLAUDE_BYTES_PER_TOKEN = 3.4
+CLAUDE_HINTS = ("claude", "anthropic", "opus", "sonnet", "haiku")
+NON_CLAUDE_HINTS = ("gemini", "gemma", "gpt", "grok", "deepseek", "qwen", "glm", "kimi", "llama", "mistral",
+                    "nemotron", "minimax", "seed-", "mimo", "abliterat")
+
+
+def bytes_per_token(model):
+    slug = str(model or "").lower()
+    if not slug or any(h in slug for h in CLAUDE_HINTS):
+        return SAFE_BYTES_PER_TOKEN
+    return NON_CLAUDE_BYTES_PER_TOKEN if any(h in slug for h in NON_CLAUDE_HINTS) else SAFE_BYTES_PER_TOKEN
+
+
 def tokens(value):
-    # Deliberately more conservative than the old 3.2 bytes/token estimate.
-    return (len(encoded(value).encode("utf-8")) + 1) // 2
+    """Conservative token estimate. A request (dict with "model") uses that model's measured ratio."""
+    model = value.get("model") if isinstance(value, dict) else None
+    return int(len(encoded(value).encode("utf-8")) / bytes_per_token(model)) + 1
 
 
 def input_budget(req, window, settings=None, automatic=False):

apps/desktop/assets/codex-bridge/kopru.py

@@ -127,6 +127,54 @@ UA = CFG.get("user_agent", "AstraBridge/1.0 (codex)")
 # Bu metinleri iceren hatalar kalicidir; yeniden denemek yerine hemen Codex'e iletilir.
 NO_RETRY_PATTERNS = CFG.get("no_retry_patterns", ["All fusion panel models failed", "model not found", "does not exist", "insufficient_balance", "allowance exhausted"])
 LOGDIR = os.path.join(HOME, "log")
+
+# Hermes'in Anthropic imza hatalari (agent/anthropic_message_convert.py): "thinking blocks in the latest assistant
+# message cannot be modified" ve "Invalid signature in thinking block". Gecit ayrintiyi "detail" alaninda iletir.
+DUSUNME_IMZA_HATASI = re.compile(r"(thinking|redacted_thinking)[^\n]{0,160}(cannot be modified|must remain as they were)"
+                                 r"|invalid signature in thinking block", re.I)
+DUSUNME_ORNEK_SAYI = 5          # tesis icin saklanan son istek sayisi (yalniz bu bilgisayarda, log klasorunde)
+
+
+def dusunme_imzasi_bozuk(ilk):
+    """Ust kaynagin ilk parcasi Claude'un dusunme-imzasi reddi mi? (yalniz "error" olayinda bakilir)"""
+    metin = ilk.decode("utf-8", "replace") if isinstance(ilk, (bytes, bytearray)) else str(ilk or "")
+    return ('"error"' in metin or "event: error" in metin) and bool(DUSUNME_IMZA_HATASI.search(metin))
+
+
+def dusunme_ornegi_sakla(req, ilk):
+    """Teshis: reddedilen istegi yerel log klasorune yazar; en yeni DUSUNME_ORNEK_SAYI dosya kalir."""
+    try:
+        os.makedirs(LOGDIR, exist_ok=True)
+        ad = os.path.join(LOGDIR, time.strftime("dusunme-400-%Y%m%d-%H%M%S-") + f"{threading.get_ident() % 100000}.json")
+        with open(ad, "w", encoding="utf-8") as f:
+            json.dump({"hata": ilk.decode("utf-8", "replace")[:2000], "istek": req}, f, ensure_ascii=False)
+        eski = sorted(x for x in os.listdir(LOGDIR) if x.startswith("dusunme-400-"))
+        for x in eski[:-DUSUNME_ORNEK_SAYI]:
+            os.remove(os.path.join(LOGDIR, x))
+        return ad
+    except OSError:
+        return None
+
+
+class OnekliYanit:
+    """Okunmus ilk parcayi once veren yanit sarmalayicisi; geri kalani ustteki yanittan gelir."""
+    def __init__(self, yanit, onek):
+        self.yanit, self.onek = yanit, onek
+
+    def __getattr__(self, ad):
+        return getattr(self.yanit, ad)
+
+    def read1(self, n=-1):
+        if self.onek:
+            parca, self.onek = self.onek, b""
+            return parca
+        return self.yanit.read1(n)
+
+    def read(self, n=-1):
+        if self.onek:
+            parca, self.onek = self.onek, b""
+            return parca + (self.yanit.read(n) if n is None or n < 0 else b"")
+        return self.yanit.read(n)
 os.makedirs(LOGDIR, exist_ok=True)
 lock = threading.Lock()
 log_lock = threading.Lock()
@@ -373,8 +421,7 @@ def sigan_katmanlar(sira, req, tag):
         bilgi = IDE_BILGI.get(IDE_PREFIX + ad) or {}
         p = bilgi.get("pencere")
         if p and i < len(sira) - 1 and pencere.bekci_calisir(bilgi.get("kaynak"), p):
-            if tahmin is None:
-                tahmin = pencere.tahmini_token(req, 2.0)
+            tahmin = pencere.tahmini_token(req, gecis_context.bytes_per_token(ad))
             if tahmin > int(p * 0.90):
                 log(f"{tag} katman {ad} atlandi: istek sigmaz (tahmini {tahmin} token, pencere {p})")
                 continue
@@ -1628,7 +1675,6 @@ class H(http.server.BaseHTTPRequestHandler):
             out = raw; changed = 0
         raw_len = len(out)
         duz = out                   # sikistirilmamis govde: yedege bu yazilir
-        catalog_etag = 'W/"astrabridge-' + hashlib.sha256(out).hexdigest()[:24] + '"'
         enc_out = None
         if "gzip" in (self.headers.get("Accept-Encoding") or "").lower() and changed:
             import gzip
@@ -1637,8 +1683,9 @@ class H(http.server.BaseHTTPRequestHandler):
         for k, v in resp.headers.items():
             kl = k.lower()
             if kl in RESP_HOP or kl in ("date", "set-cookie", "cf-ray"): continue
-            if kl in ("etag", "x-models-etag") and changed:
-                v = catalog_etag
+            # The upstream list stamp is kept (20 October 2026). chatgpt.com sends the same stamp in every answer
+            # (x-models-etag); a different stamp here made Codex re-download this ~17 MB list after every answer
+            # (6198 times in one day). Added-model changes still arrive on Codex's own 5-minute list refresh.
             hdrs_out.append((k, v))
         if changed:
             # Yedek her zaman sikistirilmamis govdeyle tutulur (istemci gzip istese de).
@@ -1959,7 +2006,9 @@ class H(http.server.BaseHTTPRequestHandler):
         if not hedef_pencere or not pencere.bekci_calisir(pencere_kaynagi, hedef_pencere):
             hedef_pencere = 32768
         if CFG.get("tasma_bekcisi", True) and not ozet_tetik:
-            tahmin = pencere.tahmini_token(req, 2.0)
+            # Olcume dayali oran (21 Ekim 2026): duz 2 bayt/token Claude disi sohbetleri iki kat buyuk sayip
+            # sahte "sigmiyor" veriyordu; beta.47'de gecis_context duzeltilmis, bu ikinci yer kalmisti.
+            tahmin = pencere.tahmini_token(req, gecis_context.bytes_per_token(slug))
             sigdi = tahmin <= int(hedef_pencere * 0.90)
             if rewriter is not None:
                 rewriter.dolu = tahmin >= hedef_pencere * 0.7
@@ -2048,7 +2097,9 @@ class H(http.server.BaseHTTPRequestHandler):
         # have started (timeout, 504, cut stream) resending it runs and bills it twice (28 Sep 2026);
         # only failures before the work starts (no connection, 500/502/503/529) are retried.
         combo = "/" not in target
-        for attempt in range(1, RETRIES + 1):
+        denemeler = list(range(1, RETRIES + 1))
+        dusunme_tekrar = False
+        for attempt in denemeler:
             r = urllib.request.Request(route["url"], data=data, headers=hdrs, method="POST")
             try:
                 resp = urllib.request.urlopen(r, timeout=UPSTREAM_IDLE_TIMEOUT)
@@ -2092,6 +2143,27 @@ class H(http.server.BaseHTTPRequestHandler):
                     break
                 log(f"{tag} deneme {attempt}: {e}")
                 time.sleep(min(2 * attempt, 8)); continue
+            if "text/event-stream" in str(resp.headers.get("Content-Type", "")):
+                try:
+                    ilk = resp.read1(4096)
+                except (OSError, http.client.HTTPException):
+                    ilk = b""
+                if not dusunme_tekrar and dusunme_imzasi_bozuk(ilk):
+                    # Hermes: imzasi olen son tur dusunmesiz yeniden gonderilir. Bir kez; istemciye bir sey yazilmadi.
+                    dusunme_tekrar = True
+                    ornek = dusunme_ornegi_sakla(req, ilk)
+                    log(f"{tag} Claude dusunme imzasi reddi: istek bir kez dusunme kapali yeniden gonderiliyor"
+                        f"{' (ornek ' + os.path.basename(ornek) + ')' if ornek else ''}")
+                    try:
+                        resp.close()
+                    except Exception:
+                        pass
+                    req["reasoning"] = {"effort": "none"}
+                    req.pop("include", None)
+                    data = json.dumps(req, ensure_ascii=False).encode("utf-8")
+                    denemeler.append(attempt)
+                    continue
+                resp = OnekliYanit(resp, ilk)
             # Only the resolved exact voice target owns this sideband protocol.
             # Other models retain their original response/Jev/stream pipeline.
             if jev_denetim.izleme_mi(target):

apps/desktop/assets/codex-bridge/test_access_programs.py

@@ -33,7 +33,7 @@ class AccessPrograms(unittest.TestCase):
             "fix_tool_ids": lambda req: 0,
             "drop_foreign_items": lambda req: 0,
             "SseRewriter": Mock(),
-            "log": lambda message: None, "pencere": __import__("pencere"), "IDE_BILGI": {}, "yarim": __import__("yarim"), "jev_denetim": __import__("jev_denetim"), "internet": __import__("internet"), "is_hatti": __import__("is_hatti"), "GORULEN_ISTEMCI": set(), "arac_semalarini_duzelt": lambda req: 0, "nokta_mesaji_mi": lambda req: False, "EXTRA_MODELS": [],
+            "log": lambda message: None, "pencere": __import__("pencere"), "gecis_context": __import__("gecis_context"), "dusunme_imzasi_bozuk": __import__("kopru").dusunme_imzasi_bozuk, "dusunme_ornegi_sakla": __import__("kopru").dusunme_ornegi_sakla, "OnekliYanit": __import__("kopru").OnekliYanit, "IDE_BILGI": {}, "yarim": __import__("yarim"), "jev_denetim": __import__("jev_denetim"), "internet": __import__("internet"), "is_hatti": __import__("is_hatti"), "GORULEN_ISTEMCI": set(), "arac_semalarini_duzelt": lambda req: 0, "nokta_mesaji_mi": lambda req: False, "EXTRA_MODELS": [],
             "DEBUG": False, "UA": "test", "RETRIES": 1, "UPSTREAM_IDLE_TIMEOUT": 3900,
         }
         exec(compile(ast.Module(body=taraf + [flatten, claude_sonu, responses, ozgun], type_ignores=[]), "bridge.py", "exec"), self.env)

apps/desktop/assets/codex-bridge/test_combo_tekrar.py

@@ -40,7 +40,7 @@ class ComboNotResent(unittest.TestCase):
             "fix_tool_ids": lambda req: 0, "drop_foreign_items": lambda req: 0,
             "SseRewriter": Mock(), "log": lambda message: None,
             "DEBUG": False, "UA": "test", "RETRIES": 6, "UPSTREAM_IDLE_TIMEOUT": 3600,
-            "NO_RETRY_PATTERNS": [], "pencere": __import__("pencere"), "IDE_BILGI": {}, "yarim": __import__("yarim"), "jev_denetim": __import__("jev_denetim"), "internet": __import__("internet"), "is_hatti": __import__("is_hatti"), "GORULEN_ISTEMCI": set(), "arac_semalarini_duzelt": lambda req: 0, "nokta_mesaji_mi": lambda req: False, "EXTRA_MODELS": [],
+            "NO_RETRY_PATTERNS": [], "pencere": __import__("pencere"), "gecis_context": __import__("gecis_context"), "dusunme_imzasi_bozuk": __import__("kopru").dusunme_imzasi_bozuk, "dusunme_ornegi_sakla": __import__("kopru").dusunme_ornegi_sakla, "OnekliYanit": __import__("kopru").OnekliYanit, "IDE_BILGI": {}, "yarim": __import__("yarim"), "jev_denetim": __import__("jev_denetim"), "internet": __import__("internet"), "is_hatti": __import__("is_hatti"), "GORULEN_ISTEMCI": set(), "arac_semalarini_duzelt": lambda req: 0, "nokta_mesaji_mi": lambda req: False, "EXTRA_MODELS": [],
         }
         exec(compile(ast.Module(body=[pick("ek_model_mu"), pick("yonlendirilen"), pick("flatten_request"), pick("claude_sonu_duzelt"),
                                       pick("is_timeout"), pick("kopma_mi"), responses],

apps/desktop/assets/codex-bridge/test_denetim_ek.py

@@ -40,6 +40,11 @@ def method(name):
 
 
 def build(nodes, env):
+    # Names the real module defines at import time (beta.48: measured token ratio, thinking-signature resend).
+    env.setdefault("gecis_context", __import__("gecis_context"))
+    env.setdefault("dusunme_imzasi_bozuk", __import__("kopru").dusunme_imzasi_bozuk)
+    env.setdefault("dusunme_ornegi_sakla", __import__("kopru").dusunme_ornegi_sakla)
+    env.setdefault("OnekliYanit", __import__("kopru").OnekliYanit)
     exec(compile(ast.Module(body=nodes, type_ignores=[]), "bridge.py", "exec"), env)
     return env
 
@@ -483,6 +488,19 @@ class Liste(unittest.TestCase):
         self.assertIn(("Content-Encoding", "gzip"), fake.sent)
         env["save_last_models"].assert_called_once()
 
+    def test_the_list_keeps_the_upstream_stamp_so_codex_does_not_refetch_after_every_answer(self):
+        """chatgpt.com repeats its list stamp (x-models-etag) in every answer; a changed stamp here made Codex
+        download the whole list again after each answer (6198 times on 20 October 2026)."""
+        env = self.env(EXTRA_MODELS=[{"slug": "ide/a", "display_name": "Aaa"}])
+        fake = Sahte({"chatgpt-account-id": "hesap"}, "GET")
+        resp = types.SimpleNamespace(read=lambda: json.dumps(self.UPSTREAM).encode(),
+                                     headers={"ETag": 'W/"40b107fc"', "x-models-etag": 'W/"40b107fc"'})
+        env["_models"](fake, resp, "[t]")
+        sent = dict(fake.sent)
+        self.assertEqual((sent["ETag"], sent["x-models-etag"]), ('W/"40b107fc"', 'W/"40b107fc"'))
+        self.assertIn("ide/a", [m["slug"] for m in json.loads(fake.wfile.getvalue())["models"]])
+        self.assertEqual(dict(env["LAST_MODELS"]["hesap"]["headers"])["ETag"], 'W/"40b107fc"')
+
     def test_no_network_still_serves_the_last_list(self):
         cached = {"headers": [("Content-Type", "application/json")], "body": b'{"models": []}', "at": "x"}
         env = self.env(LAST_MODELS={"hesap": cached})

apps/desktop/assets/codex-bridge/test_dusunme_imzasi.py

@@ -0,0 +1,145 @@
+"""Claude thinking-signature rejection: one clean resend without thinking (Hermes rule), nothing else retried.
+
+Hermes (agent/anthropic_message_convert.py, _manage_thinking_signatures): when the latest turn's thinking
+signature is dead, the turn is replayed without it. The bridge cannot touch the blocks OmniRoute builds, so
+it resends that one request once with thinking off before anything reaches Codex. Real sockets, no network.
+"""
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+import http.client
+import json
+import os
+import threading
+
+import pytest
+
+import gecis_context
+import kopru
+import pencere
+
+IMZA = ("[400]: messages.11.content.1: `thinking` or `redacted_thinking` blocks in the latest assistant message "
+        "cannot be modified. These blocks must remain as they were in the original response.")
+
+
+def hata_olayi(detail):
+    body = {"type": "error", "error": {"type": "upstream_error", "message": "upstream 400", "detail": detail}}
+    return b"event: error\ndata: " + json.dumps(body).encode() + b"\n\n"
+
+
+TAMAM = (b'event: response.output_text.delta\ndata: {"type":"response.output_text.delta","delta":"TAMAM"}\n\n'
+         b'event: response.completed\ndata: {"type":"response.completed","response":{"id":"r","status":"completed","output":[]}}\n\n')
+
+
+class Ust:
+    """Fake gateway: answers each request with the next scripted SSE body and records the requests."""
+    def __init__(self, cevaplar):
+        self.istekler, self.cevaplar = [], list(cevaplar)
+        sahip = self
+
+        class H(BaseHTTPRequestHandler):
+            def log_message(self, *a):
+                pass
+
+            def do_POST(self):
+                sahip.istekler.append(json.loads(self.rfile.read(int(self.headers["Content-Length"]))))
+                govde = sahip.cevaplar.pop(0) if sahip.cevaplar else TAMAM
+                self.send_response(200)
+                self.send_header("Content-Type", "text/event-stream")
+                self.send_header("Content-Length", str(len(govde)))
+                self.end_headers()
+                self.wfile.write(govde)
+
+        self.server = ThreadingHTTPServer(("127.0.0.1", 0), H)
+        self.server.daemon_threads = True
+        threading.Thread(target=self.server.serve_forever, daemon=True).start()
+        self.url = f"http://127.0.0.1:{self.server.server_port}/v1/responses"
+
+    def kapat(self):
+        self.server.shutdown()
+        self.server.server_close()
+
+
+@pytest.fixture
+def kurulum(tmp_path, monkeypatch):
+    def kur(cevaplar, slug="ide/claude-claude-opus-5-5", hedef="claude-claude-opus-5-5", bilgi=None):
+        ust = Ust(cevaplar)
+        monkeypatch.setattr(kopru, "WATCH_OWNER", None)
+        monkeypatch.setattr(kopru, "LOGDIR", str(tmp_path))
+        monkeypatch.setattr(kopru.jev_denetim, "anahtar", lambda: "")
+        monkeypatch.setitem(kopru.MODEL_MAP, slug, ("aiturk-ide", hedef, None))
+        monkeypatch.setitem(kopru.ROUTES, "aiturk-ide", {"url": ust.url, "key": "uye", "flatten_tools": False})
+        if bilgi:
+            monkeypatch.setitem(kopru.IDE_BILGI, slug, bilgi)
+        sunucu = kopru.TS(("127.0.0.1", 0), kopru.H)
+        threading.Thread(target=sunucu.serve_forever, daemon=True).start()
+
+        def iste(girdi=None):
+            b = http.client.HTTPConnection("127.0.0.1", sunucu.server_port, timeout=10)
+            istek = {"model": slug, "stream": True, "reasoning": {"effort": "medium", "summary": "detailed"},
+                     "include": ["reasoning.encrypted_content"],
+                     "input": girdi or [{"type": "message", "role": "user", "content": [{"type": "input_text", "text": "selam"}]}]}
+            b.request("POST", "/v1/responses", json.dumps(istek), {"Content-Type": "application/json"})
+            y = b.getresponse()
+            veri = y.read()
+            b.close()
+            return y.status, veri
+
+        kur.kapanacak = [ust.kapat, sunucu.shutdown, sunucu.server_close]
+        return ust, iste
+    yield kur
+    for k in getattr(kur, "kapanacak", []):
+        k()
+
+
+def test_both_hermes_signature_messages_are_recognised_only_in_an_error_event():
+    assert kopru.dusunme_imzasi_bozuk(hata_olayi(IMZA))
+    assert kopru.dusunme_imzasi_bozuk(hata_olayi("[400]: messages.3.content.0: Invalid signature in thinking block"))
+    assert not kopru.dusunme_imzasi_bozuk(hata_olayi("upstream 400"))
+    assert not kopru.dusunme_imzasi_bozuk(b'data: {"type":"response.output_text.delta","delta":"thinking blocks cannot be modified"}\n\n')
+
+
+def test_signature_rejection_is_resent_once_without_thinking_and_codex_sees_no_error(kurulum, tmp_path):
+    ust, iste = kurulum([hata_olayi(IMZA), TAMAM])
+    durum, veri = iste()
+    assert durum == 200 and b"TAMAM" in veri and b"error" not in veri
+    assert len(ust.istekler) == 2
+    assert ust.istekler[0]["reasoning"]["effort"] == "medium" and "include" in ust.istekler[0]
+    assert ust.istekler[1]["reasoning"] == {"effort": "none"} and "include" not in ust.istekler[1]
+    assert ust.istekler[1]["input"] == ust.istekler[0]["input"]                 # history itself untouched
+    ornekler = [x for x in os.listdir(tmp_path) if x.startswith("dusunme-400-")]
+    assert len(ornekler) == 1
+    kayit = json.loads((tmp_path / ornekler[0]).read_text(encoding="utf-8"))
+    assert "cannot be modified" in kayit["hata"] and kayit["istek"]["input"] == ust.istekler[0]["input"]
+
+
+def test_any_other_upstream_error_is_not_resent(kurulum):
+    ust, iste = kurulum([hata_olayi("[400]: messages.0: some other provider error")])
+    durum, veri = iste()
+    assert len(ust.istekler) == 1
+    assert b"error" in veri or b"failed" in veri
+
+
+def test_a_second_signature_rejection_is_not_resent_again(kurulum):
+    ust, iste = kurulum([hata_olayi(IMZA), hata_olayi(IMZA), TAMAM])
+    iste()
+    assert len(ust.istekler) == 2
+
+
+def test_a_normal_stream_keeps_its_first_piece(kurulum):
+    ust, iste = kurulum([TAMAM])
+    durum, veri = iste()
+    assert durum == 200 and b"TAMAM" in veri and b"response.completed" in veri
+    assert len(ust.istekler) == 1 and ust.istekler[0]["reasoning"]["effort"] == "medium"
+
+
+def test_overflow_guard_counts_with_the_measured_ratio(kurulum):
+    """~100k real tokens for a 128k Gemini: the old 2 bytes/token guard refused it before sending."""
+    metin = "x" * 330_000
+    girdi = [{"type": "message", "role": "user", "content": [{"type": "input_text", "text": metin}]}]
+    slug = "ide/agy-gemini-3.8-flash-low"
+    assert pencere.tahmini_token({"input": girdi}, 2.0) > int(128000 * 0.90)                    # old: refused
+    assert pencere.tahmini_token({"input": girdi}, gecis_context.bytes_per_token(slug)) < int(128000 * 0.90)
+    ust, iste = kurulum([TAMAM], slug=slug, hedef="agy-gemini-3.8-flash-low",
+                        bilgi={"pencere": 128000, "kaynak": "saglayici-katalogu"})
+    durum, veri = iste(girdi)
+    assert len(ust.istekler) == 1 and b"TAMAM" in veri
+    assert "tahmini_token(req, 2.0)" not in open(kopru.__file__, encoding="utf-8").read()

apps/desktop/assets/codex-bridge/test_gecis_context.py

@@ -195,8 +195,9 @@ class PreparationService(unittest.TestCase):
 
     def test_output_reserve_can_require_preparation_below_seventy_percent(self):
         req = {"model": "ide/test", "max_output_tokens": 24000,
-               "input": [{"role": "user", "content": "x" * 20000}, {"role": "user", "content": "latest"}]}
+               "input": [{"role": "user", "content": "x" * 30000}, {"role": "user", "content": "latest"}]}
         self.assertLess(gc.tokens(req), int(32768 * .7))
+        self.assertGreater(gc.tokens(req), 32768 - 24000)
         result = self.service._prepare_transfer({"request": req})["request"]
         self.assertLessEqual(gc.tokens(result), 32768 - 24000)
         self.assertTrue(self.server.gelen)
@@ -446,5 +447,35 @@ class SocketMigration(fixtures.Ayirici):
         self.assertEqual(len(self.kopru.gelen), 1)
 
 
+class TokenEstimate(unittest.TestCase):
+    """Measured bytes/token (20 October 2026): Claude ~2.55, other families >= 3.4. Unknown stays safe."""
+
+    def test_ratio_follows_the_measured_model_family_and_unknown_stays_safe(self):
+        for slug, ratio in (("ide/claude-claude-haiku-5-5", 2.5), ("ide/claude-claude-opus-5-5", 2.5),
+                            ("ide/nanogpt-Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled", 2.5),
+                            ("ide/Jail-Omni", 2.5), ("ide/Takim-Hizli", 2.5), ("", 2.5), (None, 2.5),
+                            ("ide/agy-gemini-3.8-flash-low", 3.4), ("gpt-6.1-sol", 3.4), ("ide/deepseek-deepseek-flash", 3.4),
+                            ("ide/abliteration-abliterated-model-large-v2", 3.4), ("ide/grok-cli-grok-4.7", 3.4)):
+            self.assertEqual(gc.bytes_per_token(slug), ratio, slug)
+
+    def test_estimate_never_counts_fewer_tokens_than_the_measured_lowest_ratio(self):
+        text = "Türkçe ve English karışık bir metin, kod: def f(x): return x * 2. " * 4000
+        size = len(gc.encoded({"model": "ide/claude-claude-opus-5-5", "input": text}).encode("utf-8"))
+        self.assertGreaterEqual(gc.tokens({"model": "ide/claude-claude-opus-5-5", "input": text}), size / 2.55)
+        size = len(gc.encoded({"model": "ide/agy-gemini-3.8-flash-low", "input": text}).encode("utf-8"))
+        self.assertGreaterEqual(gc.tokens({"model": "ide/agy-gemini-3.8-flash-low", "input": text}), size / 3.41)
+        self.assertGreater(gc.tokens([text]), gc.tokens({"model": "gpt-6.1-sol", "input": text}))  # no model: safe
+
+    def test_a_chat_that_really_fits_a_non_claude_target_is_not_refused(self):
+        """Before: 2 bytes/token counted this ~75k-token chat as ~130k and refused it for a 128k model."""
+        turn = [{"type": "message", "role": "user", "content": [{"type": "input_text", "text": "soru " + "x" * 64000}]},
+                {"type": "message", "role": "assistant", "content": [{"type": "output_text", "text": "cevap " + "y" * 64000}]}]
+        req = {"model": "ide/agy-gemini-3.8-flash-low", "input": turn * 2}
+        self.assertLess(gc.tokens(req), int(128000 * 0.70))
+        self.assertGreater(len(gc.encoded(req).encode("utf-8")) // 2, int(128000 * 0.70))   # old estimate refused it
+        prepared = gc.Preparer().prepare(req, 128000, lambda text, budget: self.fail("must not summarize"))
+        self.assertEqual(prepared["input"], req["input"])
+
+
 if __name__ == "__main__":
     unittest.main()

apps/desktop/assets/codex-bridge/test_karar_asamalari.py

@@ -336,7 +336,7 @@ class TurBasi(Ev):
 class KatmanPenceresi(unittest.TestCase):
     def suzgec(self, bilgi):
         logs = []
-        env = {'IDE_PREFIX': 'ide/', 'IDE_BILGI': bilgi, 'pencere': pencere, 'log': logs.append}
+        env = {'IDE_PREFIX': 'ide/', 'IDE_BILGI': bilgi, 'pencere': pencere, 'log': logs.append, 'gecis_context': __import__('gecis_context')}
         exec(compile(ast.Module(body=[fonksiyon('sigan_katmanlar')], type_ignores=[]), 'kopru.py', 'exec'), env)
         return env['sigan_katmanlar'], logs
 

apps/desktop/electron/codex-link-config.ts

@@ -27,7 +27,7 @@ export const CODEX_LINK_REFRESH_URL = `${CODEX_LINK_ORIGIN}/aiturk/codex-baglama
 export const CODEX_LINK_MARKER = '# AITURK IDE Codex Baglama'
 // Must equal SURUM in assets/codex-bridge/bridge.py; a running bridge that reports
 // another version is restarted so an updated IDE never keeps an old bridge alive.
-export const CODEX_BRIDGE_VERSION = '25'
+export const CODEX_BRIDGE_VERSION = '27'
 export const CODEX_LINK_LINE = `openai_base_url = "${CODEX_LINK_ORIGIN}/backend-api/codex"  ${CODEX_LINK_MARKER}`
 
 export const CODEX_LINK_RUN_KEY = 'HKCU\\Software\\Microsoft\\Windows\\CurrentVersion\\Run'

apps/desktop/package.json

@@ -2,7 +2,7 @@
   "name": "aiturk-ide",
   "productName": "AITURK IDE",
   "private": true,
-  "version": "1.0.0-beta.46",
+  "version": "1.0.0-beta.48",
   "description": "TurkServis resmi yapay zekâ geliştirme ortamı. Hermes Agent tabanlıdır.",
   "author": "AITURK / TurkServis",
   "repository": {

package-lock.json

@@ -65,7 +65,7 @@
     },
     "apps/desktop": {
       "name": "aiturk-ide",
-      "version": "1.0.0-beta.46",
+      "version": "1.0.0-beta.48",
       "license": "MIT",
       "dependencies": {
         "@assistant-ui/core": "0.2.23",