fix(codex): canonicalise the custom-endpoint issuer kind

The openai SDK appends a trailing slash to `client.base_url`, so the aux
adapter stamped `other:https://h/v1/` while the main transport stamped
`other:https://h/v1`. On custom Responses endpoints every aux call
(compression, flush_memories) therefore dropped all main-minted reasoning
items as "foreign".

`_classify_responses_issuer` now strips whitespace and trailing slashes and
lowercases scheme+netloc before stamping. The aux adapter also derives its
route flags from `classify_responses_route` — the single owner of the
codex/xai/github predicates — instead of an inline chatgpt.com host check,
and reuses the same flags for the effort clamp.

agent/auxiliary_client.py

@@ -1341,13 +1341,16 @@ class _CodexCompletionsAdapter:
             _chat_messages_to_responses_input,
             _classify_responses_issuer,
             _wire_model_identity,
+            classify_responses_route,
         )
         model = kwargs.get("model", self._model)
         wire_model = _wire_model_identity(model)
         host = str(getattr(self._client, "base_url", "") or "")
         is_xai = base_url_host_matches(host, "x.ai") or base_url_host_matches(host, "api.x.ai")
         is_copilot = base_url_host_matches(host, "githubcopilot.com")
-        is_github = is_copilot or base_url_host_matches(host, "models.github.ai")
+        # Same route classifier as the main transport, so the issuer stamp matches what it minted.
+        route = classify_responses_route(SimpleNamespace(provider=None, base_url=host))
+        is_github = route.is_github_responses
         # System → ``instructions``; the rest goes through the SINGLE shared chat→Responses
         # converter (a private loop here once let role="tool" leak into input[]; the shared one
         # encodes tool history as function_call/function_call_output).
@@ -1368,10 +1371,7 @@ class _CodexCompletionsAdapter:
         # Aux requests run their own model; stamp/filter reasoning provenance against it, not the main agent's.
         input_items = _chat_messages_to_responses_input(
             replay_messages, is_github_responses=is_copilot,
-            current_issuer_kind=_classify_responses_issuer(
-                is_xai_responses=is_xai, is_github_responses=is_github,
-                is_codex_backend=base_url_host_matches(host, "chatgpt.com"), base_url=host,
-            ),
+            current_issuer_kind=_classify_responses_issuer(base_url=host, **route._asdict()),
             current_issuer_model=wire_model, native_compaction_eligible=False,
         )
         resp_kwargs: Dict[str, Any] = {
@@ -1400,13 +1400,11 @@ class _CodexCompletionsAdapter:
             if isinstance(reasoning_cfg, dict) and reasoning_cfg.get("enabled") is not False:
                 # Truthy-only: Codex 400s on e.g. {"effort": null}, so falsy → default. Shared
                 # per-model clamp with the main transport ("max" is gpt-5.6-only; "minimal"/"ultra" rejected).
-                from agent.codex_responses_adapter import classify_responses_route
                 from agent.reasoning_effort import clamp_effort
                 from agent.transports.codex import _codex_efforts_for_route
-                is_codex_backend = classify_responses_route(SimpleNamespace(base_url=host)).is_codex_backend
                 effort = clamp_effort(
                     reasoning_cfg.get("effort") or "medium",
-                    _codex_efforts_for_route(model, host, is_codex_backend=is_codex_backend),
+                    _codex_efforts_for_route(model, host, is_codex_backend=route.is_codex_backend),
                 )
                 resp_kwargs["reasoning"] = {"effort": effort, "summary": "auto"}
                 resp_kwargs["include"] = ["reasoning.encrypted_content"]

agent/codex_responses_adapter.py

@@ -11,6 +11,7 @@ import unicodedata
 import uuid
 from types import SimpleNamespace
 from typing import Any, Callable, Dict, Iterator, List, NamedTuple, Optional, TypeGuard
+from urllib.parse import urlsplit, urlunsplit
 
 from agent.message_sanitization import deterministic_call_id
 from agent.prompt_builder import DEFAULT_AGENT_IDENTITY
@@ -27,7 +28,13 @@ def _classify_responses_issuer(
     for flag, kind in ((is_xai_responses, "xai_responses"), (is_github_responses, "github_responses"), (is_codex_backend, "codex_backend")):
         if flag:
             return kind
-    return f"other:{base_url}" if base_url else "other"
+    if not base_url:
+        return "other"
+    # The openai SDK appends a trailing slash to ``client.base_url`` and hosts are case-insensitive, so the
+    # aux adapter and the main transport must canonicalise the same endpoint to one kind or aux calls drop
+    # every main-minted blob.
+    parts = urlsplit(str(base_url).strip().rstrip("/"))
+    return f"other:{urlunsplit((parts.scheme.lower(), parts.netloc.lower(), parts.path, parts.query, parts.fragment))}"
 
 
 # Per-process throttle for the cross-issuer skip warning.

tests/agent/test_codex_responses_adapter.py

@@ -5,6 +5,7 @@ import pytest
 from agent.codex_responses_adapter import (
     _chat_content_to_responses_parts,
     _chat_messages_to_responses_input,
+    _classify_responses_issuer,
     _sanitize_replayed_fn_name,
     _format_responses_error,
     _normalize_codex_response,
@@ -610,6 +611,15 @@ def test_legacy_endpoint_stamped_item_without_model_replays_on_same_issuer():
     assert not any(i.get("type") == "reasoning" for i in foreign)
 
 
+def test_issuer_kind_is_canonical_across_trailing_slash_and_host_case():
+    # The openai SDK stores ``client.base_url`` with a trailing slash; the aux adapter and the main
+    # transport must agree on one issuer kind or aux calls drop every main-minted blob.
+    canonical = _classify_responses_issuer(base_url="https://h/v1")
+    assert _classify_responses_issuer(base_url="https://h/v1/") == canonical
+    assert _classify_responses_issuer(base_url=" HTTPS://H/v1 ") == canonical
+    assert _classify_responses_issuer(base_url="https://other/v1") != canonical
+
+
 def test_preflight_codex_api_kwargs_drops_oversized_message_id_end_to_end():
     kwargs = _preflight_codex_api_kwargs(
         {