fix: improve Codex parity across provider, runtime, and workspace flows

core/framework/config.py

@@ -13,6 +13,7 @@
 from typing import Any
 
 from framework.graph.edge import DEFAULT_MAX_TOKENS
+from framework.llm.codex_backend import CODEX_API_BASE, build_codex_litellm_kwargs
 
 # ---------------------------------------------------------------------------
 # Low-level config file access
@@ -125,7 +126,6 @@ def get_worker_api_key() -> str | None:
                 return token
         except ImportError:
             pass
-
     api_key_env_var = worker_llm.get("api_key_env_var")
     if api_key_env_var:
         return os.environ.get(api_key_env_var)
@@ -141,7 +141,7 @@ def get_worker_api_base() -> str | None:
         return get_api_base()
 
     if worker_llm.get("use_codex_subscription"):
-        return "https://chatgpt.com/backend-api/codex"
+        return CODEX_API_BASE
     if worker_llm.get("use_kimi_code_subscription"):
         return "https://api.kimi.com/coding"
     if worker_llm.get("use_antigravity_subscription"):
@@ -169,23 +169,14 @@ def get_worker_llm_extra_kwargs() -> dict[str, Any]:
     if worker_llm.get("use_codex_subscription"):
         api_key = get_worker_api_key()
         if api_key:
-            headers: dict[str, str] = {
-                "Authorization": f"Bearer {api_key}",
-                "User-Agent": "CodexBar",
-            }
+            account_id = None
             try:
                 from framework.runner.runner import get_codex_account_id
 
                 account_id = get_codex_account_id()
-                if account_id:
-                    headers["ChatGPT-Account-Id"] = account_id
             except ImportError:
                 pass
-            return {
-                "extra_headers": headers,
-                "store": False,
-                "allowed_openai_params": ["store"],
-            }
+            return build_codex_litellm_kwargs(api_key, account_id=account_id)
     return {}
 
 
@@ -274,7 +265,6 @@ def get_api_key() -> str | None:
                 return token
         except ImportError:
             pass
-
     # Standard env-var path (covers ZAI Code and all API-key providers)
     api_key_env_var = llm.get("api_key_env_var")
     if api_key_env_var:
@@ -380,7 +370,7 @@ def get_api_base() -> str | None:
     llm = get_hive_config().get("llm", {})
     if llm.get("use_codex_subscription"):
         # Codex subscription routes through the ChatGPT backend, not api.openai.com.
-        return "https://chatgpt.com/backend-api/codex"
+        return CODEX_API_BASE
     if llm.get("use_kimi_code_subscription"):
         # Kimi Code uses an Anthropic-compatible endpoint (no /v1 suffix).
         return "https://api.kimi.com/coding"
@@ -415,23 +405,14 @@ def get_llm_extra_kwargs() -> dict[str, Any]:
     if llm.get("use_codex_subscription"):
         api_key = get_api_key()
         if api_key:
-            headers: dict[str, str] = {
-                "Authorization": f"Bearer {api_key}",
-                "User-Agent": "CodexBar",
-            }
+            account_id = None
             try:
                 from framework.runner.runner import get_codex_account_id
 
                 account_id = get_codex_account_id()
-                if account_id:
-                    headers["ChatGPT-Account-Id"] = account_id
             except ImportError:
                 pass
-            return {
-                "extra_headers": headers,
-                "store": False,
-                "allowed_openai_params": ["store"],
-            }
+            return build_codex_litellm_kwargs(api_key, account_id=account_id)
     return {}
 
 

core/framework/graph/conversation.py

@@ -351,13 +351,15 @@ def __init__(
     def system_prompt(self) -> str:
         return self._system_prompt
 
-    def update_system_prompt(self, new_prompt: str) -> None:
+    def update_system_prompt(self, new_prompt: str, output_keys: list[str] | None = None) -> None:
         """Update the system prompt.
 
         Used in continuous conversation mode at phase transitions to swap
         Layer 3 (focus) while preserving the conversation history.
         """
         self._system_prompt = new_prompt
+        if output_keys is not None:
+            self._output_keys = output_keys
         self._meta_persisted = False  # re-persist with new prompt
 
     def set_current_phase(self, phase_id: str) -> None:
@@ -771,7 +773,7 @@ async def compact(
             delete_before = recent_messages[0].seq if recent_messages else self._next_seq
             await self._store.delete_parts_before(delete_before)
             await self._store.write_part(summary_msg.seq, summary_msg.to_storage_dict())
-            await self._store.write_cursor({"next_seq": self._next_seq})
+            await self._write_cursor_update({"next_seq": self._next_seq})
 
         self._messages = [summary_msg] + recent_messages
         self._last_api_input_tokens = None  # reset; next LLM call will recalibrate
@@ -975,7 +977,7 @@ async def compact_preserving_structure(
             # Write kept structural messages (they may have been modified)
             for msg in kept_structural:
                 await self._store.write_part(msg.seq, msg.to_storage_dict())
-            await self._store.write_cursor({"next_seq": self._next_seq})
+            await self._write_cursor_update({"next_seq": self._next_seq})
 
         # Reassemble: reference + kept structural (in original order) + recent
         self._messages = [ref_msg] + kept_structural + recent_messages
@@ -1012,7 +1014,7 @@ async def clear(self) -> None:
         """Remove all messages, keep system prompt, preserve ``_next_seq``."""
         if self._store:
             await self._store.delete_parts_before(self._next_seq)
-            await self._store.write_cursor({"next_seq": self._next_seq})
+            await self._write_cursor_update({"next_seq": self._next_seq})
         self._messages.clear()
         self._last_api_input_tokens = None
 
@@ -1047,14 +1049,22 @@ def export_summary(self) -> str:
 
     # --- Persistence internals ---------------------------------------------
 
+    async def _write_cursor_update(self, data: dict[str, Any]) -> None:
+        """Merge cursor updates instead of clobbering existing crash-recovery state."""
+        if self._store is None:
+            return
+        cursor = await self._store.read_cursor() or {}
+        cursor.update(data)
+        await self._store.write_cursor(cursor)
+
     async def _persist(self, message: Message) -> None:
         """Write-through a single message.  No-op when store is None."""
         if self._store is None:
             return
         if not self._meta_persisted:
             await self._persist_meta()
         await self._store.write_part(message.seq, message.to_storage_dict())
-        await self._store.write_cursor({"next_seq": self._next_seq})
+        await self._write_cursor_update({"next_seq": self._next_seq})
 
     async def _persist_meta(self) -> None:
         """Lazily write conversation metadata to the store (called once)."""