feat: deepseek

traderai/agent.py

@@ -20,7 +20,8 @@ from traderai.tools import ToolRegistry
 from traderai.version import __version__
 
 
-SYSTEM_PROMPT = """You are TraderAI, a local assistant for UEX marketplace work.
+SYSTEM_PROMPT = """You are TraderAI, a sharp Star Citizen marketplace copilot for UEX work.
+Sound like a competent player who knows the game and the market. Be natural, direct, and helpful. Avoid corporate filler, robotic phrasing, and meta notes.
 Use tools when the user asks about UEX data, open/current listings, active negotiations, unread notifications, messages, offers, or posting ads.
 Use continual plan tools when the user asks for multi-day or recurring marketplace work, such as finding several parts, watching for deals, tracking candidates, or coordinating negotiations over time.
 UEX credentials are configured server-side when available. Never ask the user to provide UEX_SECRET_KEY or UEX_BEARER_TOKEN in chat; call the authenticated UEX tool and only mention credential configuration if the tool returns an authentication error.
@@ -29,11 +30,13 @@ When the user asks for history, trends, changes over time, or past prices, prefe
 When you need missing Star Citizen knowledge to answer accurately, use Star Citizen Wiki tools during your reasoning instead of guessing.
 Use SCMDB tools when the user asks about Star Citizen missions/contracts, mission rewards, payouts, reputation gains, item rewards, blueprint rewards, or hauling mission cargo. Prefer SCMDB live data unless the user asks for PTU or a specific game version.
 Use Star Citizen Wiki tools for general game knowledge, ships and vehicles, store availability, purchase locations, ship prices, manufacturers, locations, and page summaries from starcitizen.tools.
+Use Wikelo ship project tools when the user asks for Wikelo ship requirements, Wikelo build materials, or what items are needed for a Wikelo ship project.
 Use Cornerstone tools when the user asks where an item is sold, which shops carry an item, item store locations, in-game item base prices, or Universal Item Finder data.
 When drafting UEX marketplace item posts that need images, use Cornerstone media tools or draft_marketplace_listing_with_cornerstone_image so the pending listing can include UEX image_data sourced from Cornerstone.
 Prefer open and current UEX marketplace information. Do not use historical sale data, completed sale records, or sale/average-history information unless the user explicitly asks for historical sales.
 Treat UEX marketplace prices as in-game aUEC/UEC credits, never real-world dollars, unless the user explicitly says otherwise.
 For marketplace writes, draft the exact pending action and tell the user what will be sent; never claim it was sent until approval succeeds.
+When drafting negotiation messages or marketplace replies, write like a real player would. Keep messages human, concise, and purposeful. Never include internal notes like "Tone note".
 For continual plans, never invent an unknown parts checklist. If the required items cannot be derived from provided details or tools, create the plan in a needs-input state and say what item list is missing.
 When a scheduled wake job fires, always write a concise Inbox-ready result that says what you checked, the key findings, and the suggested next action.
 Keep prices, listing ids, slugs, users, and UEX status codes precise. If data is missing, say what you need next."""
@@ -64,7 +67,7 @@ class OllamaAgent:
         self.thread_messages: dict[str, list[dict[str, Any]]] = {}
 
     async def health(self) -> dict[str, Any]:
-        if self.provider == "openai":
+        if self._is_openai_compatible_provider():
             return await self._openai_health()
         if self.provider == "codex":
             return await self._codex_health()
@@ -119,7 +122,7 @@ class OllamaAgent:
         last_tool_results: list[dict[str, Any]] = []
         image_scope = self.tools.chat_image_scope(normalized_images) if hasattr(self.tools, "chat_image_scope") else nullcontext()
         with image_scope:
-            for _ in range(10):
+            for _ in self._tool_rounds():
                 try:
                     response = await self._chat_once(
                         prompt_text,
@@ -155,7 +158,7 @@ class OllamaAgent:
                     result = await self.tools.execute(name, arguments)
                     last_tool_results.append({"tool": name, "result": result})
                     messages.append({"role": "tool", "tool_name": name, "tool_call_id": call.get("id"), "content": json.dumps(result)})
-        fallback = "I hit the tool-call limit while working on that. Try narrowing the request or approve any pending action first."
+        fallback = self._tool_round_limit_message()
         messages.append({"role": "assistant", "content": fallback})
         if self.memory:
             self.memory.add_conversation("assistant", fallback, resolved_thread_id)
@@ -187,7 +190,7 @@ class OllamaAgent:
         last_tool_results: list[dict[str, Any]] = []
         image_scope = self.tools.chat_image_scope(normalized_images) if hasattr(self.tools, "chat_image_scope") else nullcontext()
         with image_scope:
-            for _ in range(10):
+            for _ in self._tool_rounds():
                 assistant_message: dict[str, Any] = {"role": "assistant", "content": ""}
                 tool_calls: list[dict[str, Any]] = []
 
@@ -198,7 +201,15 @@ class OllamaAgent:
                         previous_interaction=previous_interaction,
                         thread_id=resolved_thread_id,
                     ):
+                        if event.get("type") == "reasoning":
+                            reasoning_chunk = event.get("content") or ""
+                            if reasoning_chunk:
+                                assistant_message["reasoning_content"] = assistant_message.get("reasoning_content", "") + reasoning_chunk
+                                yield {"type": "reasoning", "content": reasoning_chunk}
+                            continue
                         message = event.get("message") or {}
+                        if message.get("reasoning_content"):
+                            assistant_message["reasoning_content"] = message.get("reasoning_content")
                         chunk = message.get("content") or ""
                         if chunk:
                             assistant_message["content"] += chunk
@@ -247,7 +258,7 @@ class OllamaAgent:
                     messages.append({"role": "tool", "tool_name": name, "tool_call_id": call.get("id"), "content": json.dumps(result)})
 
                 yield {"type": "status", "message": "Writing response"}
-        fallback = "I hit the tool-call limit while working on that. Try narrowing the request or approve any pending action first."
+        fallback = self._tool_round_limit_message()
         messages.append({"role": "assistant", "content": fallback})
         if self.memory:
             self.memory.add_conversation("assistant", fallback, resolved_thread_id)
@@ -260,7 +271,7 @@ class OllamaAgent:
         previous_interaction = self.memory.last_interaction("wake") if self.memory else None
         messages.append({"role": "user", "content": wake_message})
         last_tool_results: list[dict[str, Any]] = []
-        for _ in range(10):
+        for _ in self._tool_rounds():
             try:
                 response = await self._chat_once(
                     wake_message,
@@ -298,7 +309,7 @@ class OllamaAgent:
                 result = await self.tools.execute(name, arguments)
                 last_tool_results.append({"tool": name, "result": result})
                 messages.append({"role": "tool", "tool_name": name, "tool_call_id": call.get("id"), "content": json.dumps(result)})
-        content = "I hit the tool-call limit while running this scheduled wake job. Check the job prompt or pending approvals."
+        content = self._wake_tool_round_limit_message()
         messages.append({"role": "assistant", "content": content})
         if self.memory:
             self.memory.add_conversation("system", wake_message, "wake")
@@ -312,7 +323,7 @@ class OllamaAgent:
         previous_interaction: dict[str, Any] | None = None,
         thread_id: str | None = DEFAULT_THREAD_ID,
     ) -> dict[str, Any]:
-        if self.provider == "openai":
+        if self._is_openai_compatible_provider():
             return await self._openai_chat(
                 query,
                 messages,
@@ -340,7 +351,7 @@ class OllamaAgent:
         previous_interaction: dict[str, Any] | None = None,
         thread_id: str | None = DEFAULT_THREAD_ID,
     ) -> AsyncIterator[dict[str, Any]]:
-        if self.provider == "openai":
+        if self._is_openai_compatible_provider():
             async for event in self._openai_chat_stream(
                 query,
                 messages,
@@ -441,8 +452,8 @@ class OllamaAgent:
                         thread_id=thread_id,
                     ),
                     "tools": self.tools.schemas,
-                    "reasoning_effort": self.reasoning_effort,
                     "stream": False,
+                    **self._openai_request_options(stream=False),
                 },
             )
         response.raise_for_status()
@@ -452,6 +463,7 @@ class OllamaAgent:
         return {
             "message": {
                 "role": message.get("role", "assistant"),
+                "reasoning_content": message.get("reasoning_content") or "",
                 "content": message.get("content") or "",
                 "tool_calls": message.get("tool_calls") or [],
             }
@@ -479,8 +491,8 @@ class OllamaAgent:
                         thread_id=thread_id,
                     ),
                     "tools": self.tools.schemas,
-                    "reasoning_effort": self.reasoning_effort,
                     "stream": True,
+                    **self._openai_request_options(stream=True),
                 },
             ) as response:
                 response.raise_for_status()
@@ -493,8 +505,15 @@ class OllamaAgent:
                     if payload == "[DONE]":
                         break
                     event = json.loads(payload)
+                    if event.get("usage"):
+                        metrics = self._cloud_usage_metrics(event["usage"])
+                        if metrics:
+                            yield {"type": "metrics", **metrics}
                     choice = (event.get("choices") or [{}])[0]
                     delta = choice.get("delta") or {}
+                    reasoning_content = delta.get("reasoning_content") or ""
+                    if reasoning_content:
+                        yield {"type": "reasoning", "content": reasoning_content}
                     content = delta.get("content") or ""
                     if content:
                         yield {"message": {"role": "assistant", "content": content}}
@@ -502,9 +521,11 @@ class OllamaAgent:
                         self._merge_openai_tool_call(tool_calls, tool_call)
                     finish_reason = choice.get("finish_reason")
                     if finish_reason:
+                        message = choice.get("message") or {}
                         yield {
                             "message": {
                                 "role": "assistant",
+                                "reasoning_content": message.get("reasoning_content") or "",
                                 "content": "",
                                 "tool_calls": self._ordered_tool_calls(tool_calls),
                             },
@@ -574,18 +595,19 @@ class OllamaAgent:
                 continue
             attached_image_count = len(message.get("images") or [])
             break
-        context = self._runtime_context(
+        stable_context, volatile_context = self._runtime_context_parts(
             query,
             previous_interaction=previous_interaction,
             thread_id=thread_id,
             attached_image_count=attached_image_count,
         )
-        if not context:
+        contexts = [part for part in ("\n".join(stable_context), "\n".join(volatile_context)) if part]
+        if not contexts:
             return messages
-        return [messages[0], {"role": "system", "content": context}, *messages[1:]]
+        return [messages[0], *({"role": "system", "content": context} for context in contexts), *messages[1:]]
 
     async def _openai_health(self) -> dict[str, Any]:
-        return await self._cloud_health("openai")
+        return await self._cloud_health(self.provider if self._is_openai_compatible_provider() else "openai")
 
     async def _codex_health(self) -> dict[str, Any]:
         command = self._codex_command()
@@ -671,6 +693,19 @@ class OllamaAgent:
             "Content-Type": "application/json",
         }
 
+    def _openai_request_options(self, stream: bool) -> dict[str, Any]:
+        if self.provider == "deepseek":
+            options: dict[str, Any] = {}
+            if self.reasoning_effort in {"none", "minimal"}:
+                options["thinking"] = {"type": "disabled"}
+            else:
+                options["thinking"] = {"type": "enabled"}
+                options["reasoning_effort"] = "max" if self.reasoning_effort in {"xhigh", "max"} else "high"
+            if stream:
+                options["stream_options"] = {"include_usage": True}
+            return options
+        return {"reasoning_effort": self.reasoning_effort}
+
     def _openai_messages(
         self,
         query: str,
@@ -706,6 +741,8 @@ class OllamaAgent:
                 entry["content"] = content_parts
             if role == "assistant" and message.get("tool_calls"):
                 entry["tool_calls"] = message["tool_calls"]
+            if role == "assistant" and message.get("reasoning_content"):
+                entry["reasoning_content"] = message["reasoning_content"]
             if role == "tool":
                 entry["tool_call_id"] = message.get("tool_call_id") or message.get("tool_name") or "tool"
             normalized.append(entry)
@@ -729,10 +766,29 @@ class OllamaAgent:
     def _provider_label(self) -> str:
         if self.provider == "openai":
             return "OpenAI model"
+        if self.provider == "deepseek":
+            return "DeepSeek model"
         if self.provider == "codex":
             return "Codex model"
         return "local model"
 
+    def _is_openai_compatible_provider(self) -> bool:
+        return self.provider in {"openai", "deepseek"}
+
+    def _tool_rounds(self):
+        if self.provider == "deepseek":
+            while True:
+                yield None
+            return
+        for _ in range(10):
+            yield None
+
+    def _tool_round_limit_message(self) -> str:
+        return "I hit the tool-call limit while working on that. Try narrowing the request or approve any pending action first."
+
+    def _wake_tool_round_limit_message(self) -> str:
+        return "I hit the tool-call limit while running this scheduled wake job. Check the job prompt or pending approvals."
+
     @staticmethod
     def _merge_openai_tool_call(target: dict[int, dict[str, Any]], delta: dict[str, Any]) -> None:
         index = int(delta.get("index") or 0)
@@ -1175,20 +1231,21 @@ class OllamaAgent:
                 models.append(slug)
         return sorted(set(models))
 
-    def _runtime_context(
+    def _runtime_context_parts(
         self,
         query: str,
         previous_interaction: dict[str, Any] | None = None,
         thread_id: str | None = DEFAULT_THREAD_ID,
         attached_image_count: int = 0,
-    ) -> str:
+    ) -> tuple[list[str], list[str]]:
         local_zone = get_localzone()
-        parts = [
+        stable_parts: list[str] = []
+        volatile_parts = [
             f"Current local date/time: {iso_now()} UTC; {iso_now_in_zone(local_zone)} {local_zone}.",
         ]
         if attached_image_count:
             label = "image" if attached_image_count == 1 else "images"
-            parts.append(
+            volatile_parts.append(
                 f"Current user message includes {attached_image_count} pasted {label}. "
                 "You can inspect them visually. If the user wants one reused in a marketplace listing draft, "
                 "call draft_marketplace_listing or draft_marketplace_listing_with_cornerstone_image with "
@@ -1202,34 +1259,34 @@ class OllamaAgent:
             if uex.bearer_token:
                 auth_methods.append("bearer token")
             if auth_methods:
-                parts.append(
+                stable_parts.append(
                     "UEX API authentication is configured server-side with "
                     + " and ".join(auth_methods)
                     + "; use authenticated UEX tools directly and do not ask for tokens."
                 )
             else:
-                parts.append("UEX API authentication is not configured server-side.")
+                stable_parts.append("UEX API authentication is not configured server-side.")
         if self.user_name:
-            parts.append(f"Known user name/handle: {self.user_name}.")
+            stable_parts.append(f"Known user name/handle: {self.user_name}.")
 
         if self.memory is None:
-            return "\n".join(parts)
+            return stable_parts, volatile_parts
 
         profile = self.memory.get_profile()
         if profile:
             identity = self._profile_identity(profile)
             if identity:
-                parts.append(identity)
-            parts.append(f"Known user profile JSON: {json.dumps(self._profile_for_prompt(profile), ensure_ascii=True)}.")
+                stable_parts.append(identity)
+            stable_parts.append(f"Known user profile JSON: {json.dumps(self._profile_for_prompt(profile), ensure_ascii=True)}.")
 
         last = previous_interaction if previous_interaction is not None else self.memory.last_interaction(thread_id)
         if last:
-            parts.append(
+            volatile_parts.append(
                 f"Previous interaction before this message: {last['created_at']} "
                 f"({time_since(last['created_at'])}, role {last['role']})."
             )
         else:
-            parts.append("Previous interaction before this message: none recorded.")
+            volatile_parts.append("Previous interaction before this message: none recorded.")
 
         memories = self.memory.recall(query, limit=6)
         if memories:
@@ -1237,17 +1294,24 @@ class OllamaAgent:
                 f"- [{item['kind']}, importance {item['importance']}] {item['content']}"
                 for item in memories
             )
-            parts.append(f"Relevant long-term memories:\n{memory_text}")
+            volatile_parts.append(f"Relevant long-term memories:\n{memory_text}")
 
-        recent = self.memory.recent_conversation(limit=6, thread_id=thread_id)
-        if recent:
-            recent_text = "\n".join(
-                f"- {item['created_at']} {item['role']}: {item['content'][:500]}"
-                for item in recent
-            )
-            parts.append(f"Recent conversation excerpts from this chat:\n{recent_text}")
+        return stable_parts, volatile_parts
 
-        return "\n".join(parts)
+    def _runtime_context(
+        self,
+        query: str,
+        previous_interaction: dict[str, Any] | None = None,
+        thread_id: str | None = DEFAULT_THREAD_ID,
+        attached_image_count: int = 0,
+    ) -> str:
+        stable_parts, volatile_parts = self._runtime_context_parts(
+            query,
+            previous_interaction=previous_interaction,
+            thread_id=thread_id,
+            attached_image_count=attached_image_count,
+        )
+        return "\n".join(part for part in ("\n".join(stable_parts), "\n".join(volatile_parts)) if part)
 
     def _messages_for_thread(self, thread_id: str | None) -> list[dict[str, Any]]:
         resolved_thread_id = self._thread_id(thread_id)
@@ -1283,7 +1347,7 @@ class OllamaAgent:
             f"Message: {first_message[:800]}"
         )
         try:
-            if self.provider == "openai":
+            if self._is_openai_compatible_provider():
                 async with httpx.AsyncClient(timeout=20) as client:
                     response = await client.post(
                         f"{self.base_url}/chat/completions",
@@ -1295,6 +1359,7 @@ class OllamaAgent:
                                 {"role": "user", "content": prompt},
                             ],
                             "stream": False,
+                            **self._openai_request_options(stream=False),
                         },
                     )
                 response.raise_for_status()
@@ -1330,6 +1395,34 @@ class OllamaAgent:
         except Exception:
             return ""
 
+    async def generate_plan_draft(
+        self,
+        title: str = "",
+        objective: str = "",
+        kind: str = "buying",
+        constraints: dict[str, Any] | None = None,
+        items: list[dict[str, Any]] | None = None,
+    ) -> dict[str, Any]:
+        clean_title = str(title or "").strip()
+        clean_objective = str(objective or "").strip()
+        clean_kind = str(kind or "buying").strip().casefold() or "buying"
+        clean_constraints = dict(constraints or {})
+        clean_items = self._normalize_plan_items(items or [])
+        seed = {
+            "title": clean_title,
+            "objective": clean_objective,
+            "kind": clean_kind,
+            "constraints": clean_constraints,
+            "items": clean_items,
+        }
+        prompt = self._plan_draft_prompt(seed)
+        fallback = self._heuristic_plan_draft(seed)
+        try:
+            payload = await self._generate_plain_text(prompt, system_prompt="You draft structured continual plan JSON for TraderAI.")
+            return self._normalize_plan_draft(payload, seed, fallback)
+        except Exception:
+            return fallback
+
     @staticmethod
     def _thread_id(thread_id: str | None) -> str:
         return (thread_id or DEFAULT_THREAD_ID).strip() or DEFAULT_THREAD_ID
@@ -1346,6 +1439,254 @@ class OllamaAgent:
             text = " ".join(words[:8])
         return text[:64]
 
+    async def _generate_plain_text(self, prompt: str, system_prompt: str) -> str:
+        if self._is_openai_compatible_provider():
+            async with httpx.AsyncClient(timeout=30) as client:
+                response = await client.post(
+                    f"{self.base_url}/chat/completions",
+                    headers=self._openai_headers(),
+                    json={
+                        "model": self.model,
+                        "messages": [
+                            {"role": "system", "content": system_prompt},
+                            {"role": "user", "content": prompt},
+                        ],
+                        "stream": False,
+                        **self._openai_request_options(stream=False),
+                    },
+                )
+            response.raise_for_status()
+            choice = (response.json().get("choices") or [{}])[0]
+            message = choice.get("message") or {}
+            return str(message.get("content") or "")
+        if self.provider == "codex":
+            result = await self._codex_app_server_turn(
+                prompt,
+                [
+                    {"role": "system", "content": system_prompt},
+                    {"role": "user", "content": prompt},
+                ],
+                thread_id="plan-draft",
+            )
+            return str(result.get("message") or "")
+        async with httpx.AsyncClient(timeout=30) as client:
+            response = await client.post(
+                f"{self.base_url}/api/chat",
+                json={
+                    "model": self.model,
+                    "messages": [
+                        {"role": "system", "content": system_prompt},
+                        {"role": "user", "content": prompt},
+                    ],
+                    "options": self._ollama_options(),
+                    "stream": False,
+                },
+            )
+        response.raise_for_status()
+        message = response.json().get("message") or {}
+        return str(message.get("content") or "")
+
+    @classmethod
+    def _normalize_plan_draft(
+        cls,
+        raw_text: str,
+        seed: dict[str, Any],
+        fallback: dict[str, Any] | None = None,
+    ) -> dict[str, Any]:
+        base = dict(fallback or cls._heuristic_plan_draft(seed))
+        payload = cls._parse_json_object(raw_text)
+        if not isinstance(payload, dict):
+            return base
+
+        title = str(payload.get("title") or seed.get("title") or base.get("title") or "").strip()
+        objective = str(payload.get("objective") or seed.get("objective") or base.get("objective") or "").strip()
+        kind = str(payload.get("kind") or seed.get("kind") or base.get("kind") or "buying").strip().casefold() or "buying"
+        cadence = cls._normalize_plan_cadence(payload.get("cadence")) or base.get("cadence")
+        constraints = cls._normalize_plan_constraints(payload.get("constraints"), seed.get("constraints") or {}, base.get("constraints") or {})
+        items = cls._normalize_plan_items(payload.get("items") or seed.get("items") or base.get("items") or [])
+
+        if kind == "buying" and not items:
+            items = list(base.get("items") or [])
+        if not constraints.get("instructions"):
+            constraints["instructions"] = (base.get("constraints") or {}).get("instructions") or cls._default_plan_instructions(kind)
+        if not constraints.get("message_tone"):
+            constraints["message_tone"] = (base.get("constraints") or {}).get("message_tone") or "friendly and direct"
+
+        return {
+            "title": title or base.get("title") or "Continual plan",
+            "objective": objective or base.get("objective") or title or "Continue this plan",
+            "kind": kind,
+            "cadence": cadence,
+            "constraints": constraints,
+            "items": items,
+        }
+
+    @classmethod
+    def _heuristic_plan_draft(cls, seed: dict[str, Any]) -> dict[str, Any]:
+        title = str(seed.get("title") or "").strip()
+        objective = str(seed.get("objective") or "").strip()
+        kind = str(seed.get("kind") or "buying").strip().casefold() or "buying"
+        constraints = cls._normalize_plan_constraints(seed.get("constraints"), {}, {})
+        items = cls._normalize_plan_items(seed.get("items") or [])
+
+        if not items and kind == "buying":
+            inferred_names = cls._infer_item_names(f"{title}\n{objective}")
+            items = [{"item_name": name, "desired_quantity": 1, "max_unit_price": None} for name in inferred_names[:8]]
+
+        if not constraints.get("message_tone"):
+            constraints["message_tone"] = "friendly and direct"
+        if not constraints.get("instructions"):
+            constraints["instructions"] = cls._default_plan_instructions(kind)
+
+        return {
+            "title": title or "Continual plan",
+            "objective": objective or title or "Continue this plan",
+            "kind": kind,
+            "cadence": cls._normalize_plan_cadence(seed.get("cadence")) or ("0 */6 * * *" if kind == "buying" else "0 */4 * * *"),
+            "constraints": constraints,
+            "items": items,
+        }
+
+    @staticmethod
+    def _default_plan_instructions(kind: str) -> str:
+        if kind == "custom":
+            return "Check for meaningful updates, summarize what changed, and suggest the next move."
+        return "Track the best active listings, avoid bad prices, and draft messages for approval when a strong candidate appears."
+
+    @staticmethod
+    def _plan_draft_prompt(seed: dict[str, Any]) -> str:
+        return (
+            "Draft a continual TraderAI plan as strict JSON.\n"
+            "Return one JSON object only with keys: title, objective, kind, cadence, constraints, items.\n"
+            "constraints may include message_tone, instructions, preferred_locations, excluded_sellers, max_unit_price.\n"
+            "items must be an array of objects with item_name, desired_quantity, max_unit_price.\n"
+            "If the request is vague, still fill cadence, message_tone, and instructions.\n"
+            "Only include checklist items when they can be reasonably inferred from the request or existing draft.\n"
+            "Do not wrap the JSON in markdown.\n\n"
+            f"Current draft seed: {json.dumps(seed, ensure_ascii=True)}"
+        )
+
+    @staticmethod
+    def _parse_json_object(raw_text: str) -> dict[str, Any] | None:
+        text = str(raw_text or "").strip()
+        if not text:
+            return None
+        try:
+            parsed = json.loads(text)
+            return parsed if isinstance(parsed, dict) else None
+        except ValueError:
+            pass
+        start = text.find("{")
+        end = text.rfind("}")
+        if start == -1 or end <= start:
+            return None
+        try:
+            parsed = json.loads(text[start : end + 1])
+            return parsed if isinstance(parsed, dict) else None
+        except ValueError:
+            return None
+
+    @classmethod
+    def _normalize_plan_constraints(cls, value: Any, seed: dict[str, Any], fallback: dict[str, Any]) -> dict[str, Any]:
+        merged: dict[str, Any] = {}
+        for source in (fallback, seed, value if isinstance(value, dict) else {}):
+            if not isinstance(source, dict):
+                continue
+            for key, item in source.items():
+                if item in (None, "", [], {}):
+                    continue
+                if key in {"preferred_locations", "excluded_sellers"}:
+                    if isinstance(item, list):
+                        merged[key] = [str(entry).strip() for entry in item if str(entry).strip()]
+                elif key == "max_unit_price":
+                    try:
+                        merged[key] = float(item)
+                    except (TypeError, ValueError):
+                        continue
+                else:
+                    merged[key] = str(item).strip()
+        return merged
+
+    @staticmethod
+    def _normalize_plan_cadence(value: Any) -> str | None:
+        text = str(value or "").strip()
+        if not text:
+            return None
+        parts = text.split()
+        return text if len(parts) == 5 else None
+
+    @classmethod
+    def _normalize_plan_items(cls, items: Any) -> list[dict[str, Any]]:
+        if not isinstance(items, list):
+            return []
+        normalized: list[dict[str, Any]] = []
+        for item in items:
+            if not isinstance(item, dict):
+                continue
+            name = str(item.get("item_name") or item.get("name") or "").strip()
+            if not name:
+                continue
+            normalized_item: dict[str, Any] = {"item_name": name}
+            try:
+                normalized_item["desired_quantity"] = max(1, int(item.get("desired_quantity") or item.get("quantity") or 1))
+            except (TypeError, ValueError):
+                normalized_item["desired_quantity"] = 1
+            try:
+                if item.get("max_unit_price") not in (None, ""):
+                    normalized_item["max_unit_price"] = float(item.get("max_unit_price"))
+                elif item.get("max_price") not in (None, ""):
+                    normalized_item["max_unit_price"] = float(item.get("max_price"))
+                else:
+                    normalized_item["max_unit_price"] = None
+            except (TypeError, ValueError):
+                normalized_item["max_unit_price"] = None
+            normalized.append(normalized_item)
+        return normalized
+
+    @staticmethod
+    def _infer_item_names(text: str) -> list[str]:
+        source = str(text or "")
+        quoted = [match.strip() for match in re.findall(r'"([^"\n]{2,80})"|\'([^\'\n]{2,80})\'', source)]
+        names = [next((part for part in group if part), "") for group in quoted]
+        if names:
+            return [name for name in names if name]
+
+        lines = []
+        for raw_line in source.splitlines():
+            line = raw_line.strip(" -*\t")
+            if not line:
+                continue
+            if any(token in line for token in [",", ";", "/"]):
+                parts = re.split(r"[,;/]+", line)
+                lines.extend(part.strip() for part in parts if part.strip())
+            else:
+                lines.append(line)
+
+        stopwords = {
+            "need", "needs", "want", "wants", "find", "draft", "deal", "deals", "parts", "part", "items",
+            "watch", "track", "check", "buy", "buying", "for", "the", "and", "with", "from", "best", "cheapest",
+        }
+        inferred = []
+        for line in lines:
+            clean = re.sub(r"\s+", " ", line).strip().strip(".")
+            if len(clean) < 3:
+                continue
+            lowered = clean.casefold()
+            if lowered in stopwords:
+                continue
+            if any(phrase in lowered for phrase in ["find and draft", "check for", "continue this plan"]):
+                continue
+            inferred.append(clean[:120])
+        deduped = []
+        seen = set()
+        for item in inferred:
+            key = item.casefold()
+            if key in seen:
+                continue
+            seen.add(key)
+            deduped.append(item)
+        return deduped
+
     def _pending_payloads(self) -> list[dict[str, Any]]:
         return [
             {
@@ -1405,6 +1746,8 @@ class OllamaAgent:
             "get_scwiki_page": "Reading Star Citizen Wiki page",
             "search_scwiki_vehicles": "Searching Star Citizen Wiki vehicles",
             "get_scwiki_vehicle": "Fetching Star Citizen Wiki vehicle",
+            "search_wikelo_ship_projects": "Searching Wikelo ship projects",
+            "get_wikelo_ship_project": "Fetching Wikelo ship requirements",
             "search_cornerstone_items": "Searching Cornerstone items",
             "get_cornerstone_item_locations": "Fetching Cornerstone item locations",
             "get_cornerstone_item_media": "Fetching Cornerstone item media",
@@ -1442,6 +1785,21 @@ class OllamaAgent:
             "writing_tokens_per_second": rate(output_tokens, output_duration),
         }
 
+    @staticmethod
+    def _cloud_usage_metrics(usage: dict[str, Any]) -> dict[str, Any]:
+        prompt_tokens = int(usage.get("prompt_tokens") or 0)
+        completion_tokens = int(usage.get("completion_tokens") or 0)
+        cache_hit_tokens = int(usage.get("prompt_cache_hit_tokens") or 0)
+        cache_miss_tokens = int(usage.get("prompt_cache_miss_tokens") or 0)
+        metrics = {
+            "reading_tokens": prompt_tokens,
+            "writing_tokens": completion_tokens,
+        }
+        if cache_hit_tokens or cache_miss_tokens:
+            metrics["cache_hit_tokens"] = cache_hit_tokens
+            metrics["cache_miss_tokens"] = cache_miss_tokens
+        return metrics
+
     @staticmethod
     def _profile_identity(profile: dict[str, Any]) -> str:
         user = profile.get("uex_user")