fix(server): preserve grounded agent stream content (#736)

* fix(server): preserve grounded agent stream content

* test(server): cover active grounded stream path

* fix(server): retain agent stream bridge

docs/user-guide/evaluations.md

@@ -31,6 +31,16 @@ uv sync --extra dev --extra eval-wandb     # Weights & Biases run tracking
 uv sync --extra dev --extra eval-sheets    # Google Sheets results export
 ```
 
+TauBench additionally requires Python 3.12 or newer and the upstream `tau2`
+package. Install the pinned revision explicitly before running that benchmark:
+
+```bash
+uv pip install "tau2 @ git+https://github.com/sierra-research/tau2-bench.git@fc0055dc4e0a316c3f83133267fbd6faaa770992"
+```
+
+OpenJarvis does not install third-party packages automatically when an
+evaluation is imported or run.
+
 !!! note "Python version requirement"
     Python 3.10 requires the `tomli` package for TOML config parsing. `openjarvis` declares it as a conditional dependency, so it is installed automatically.
 

src/openjarvis/agents/_stubs.py

@@ -155,6 +155,9 @@ class BaseAgent(ABC):
         conversation messages, and finally the user input.
         """
         messages: list[Message] = []
+        context_messages = (
+            list(context.conversation.messages) if context is not None else []
+        )
         # Check if the context already supplies a system message
         _context_has_system = (
             context
@@ -176,9 +179,28 @@ class BaseAgent(ABC):
             except Exception:
                 effective_system_prompt = None
         if effective_system_prompt:
+            context_system_text = "\n\n".join(
+                message.text
+                for message in context_messages
+                if message.role == Role.SYSTEM
+                and message.metadata.get("memory_context")
+                and message.text
+            )
+            if context_system_text:
+                effective_system_prompt = (
+                    f"{effective_system_prompt}\n\n{context_system_text}"
+                )
+                context_messages = [
+                    message
+                    for message in context_messages
+                    if not (
+                        message.role == Role.SYSTEM
+                        and message.metadata.get("memory_context")
+                    )
+                ]
             messages.append(Message(role=Role.SYSTEM, content=effective_system_prompt))
-        if context and context.conversation.messages:
-            messages.extend(context.conversation.messages)
+        if context_messages:
+            messages.extend(context_messages)
         messages.append(Message(role=Role.USER, content=input))
         return messages
 

src/openjarvis/cli/ask.py

@@ -248,6 +248,17 @@ def _get_memory_backend(config):
         return None
 
 
+def _get_memory_facts(config):
+    """Load facts captured by the automatic memory service."""
+    try:
+        from openjarvis.memory import load_configured_facts
+
+        return load_configured_facts(config)
+    except Exception as exc:
+        logger.debug("Automatic memory facts unavailable (optional): %s", exc)
+        return []
+
+
 _MEMORY_TOOLS = frozenset(
     {"retrieval", "memory_store", "memory_search", "memory_index", "memory_retrieve"}
 )
@@ -416,7 +427,8 @@ def _run_agent(
             from openjarvis.tools.storage.context import ContextConfig, inject_context
 
             backend = _get_memory_backend(config)
-            if backend is not None:
+            facts = _get_memory_facts(config)
+            if backend is not None or facts:
                 ctx_cfg = ContextConfig(
                     top_k=config.memory.context_top_k,
                     min_score=config.memory.context_min_score,
@@ -427,6 +439,7 @@ def _run_agent(
                     [],
                     backend,
                     config=ctx_cfg,
+                    facts=facts,
                 )
                 for msg in context_messages:
                     ctx.conversation.add(msg)
@@ -963,7 +976,8 @@ def ask(
             )
 
             backend = _get_memory_backend(config)
-            if backend is not None:
+            facts = _get_memory_facts(config)
+            if backend is not None or facts:
                 ctx_cfg = ContextConfig(
                     top_k=config.memory.context_top_k,
                     min_score=config.memory.context_min_score,
@@ -974,6 +988,7 @@ def ask(
                     messages,
                     backend,
                     config=ctx_cfg,
+                    facts=facts,
                 )
         except Exception as exc:
             logger.debug("Failed to inject memory context: %s", exc)

src/openjarvis/cli/chat_cmd.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import logging
 import sys
 from typing import List, Optional
 
@@ -15,6 +16,8 @@ from openjarvis.core.events import EventBus
 from openjarvis.core.types import Message, Role
 from openjarvis.memory import publish_completed_exchange
 
+logger = logging.getLogger(__name__)
+
 
 def _read_input(prompt: str = "You> ") -> Optional[str]:
     """Read user input with graceful EOF handling."""
@@ -194,6 +197,15 @@ def chat(
         console.print(f"[yellow]Memory service unavailable: {exc}[/yellow]")
         memory_service = None
 
+    # The document backend and automatic fact store are separate persistence
+    # mechanisms. Context injection combines both at read time so facts from
+    # previous sessions are immediately available without a manual index step.
+    memory_backend = None
+    if config.agent.context_from_memory:
+        from openjarvis.cli.ask import _get_memory_backend
+
+        memory_backend = _get_memory_backend(config)
+
     # Conversation state
     if not system_prompt:
         from openjarvis.prompt.builder import SystemPromptBuilder
@@ -262,15 +274,55 @@ def chat(
         # Add user message
         history.append(Message(role=Role.USER, content=user_input))
 
-        # Generate response
+        generation_history = history
+        agent_context_message = None
+        if config.agent.context_from_memory:
+            try:
+                from openjarvis.memory import load_configured_facts
+                from openjarvis.tools.storage.context import (
+                    ContextConfig,
+                    inject_context,
+                )
+
+                if memory_service is not None and hasattr(memory_service, "list_facts"):
+                    facts = memory_service.list_facts()
+                else:
+                    facts = load_configured_facts(config)
+                ctx_cfg = ContextConfig(
+                    top_k=config.memory.context_top_k,
+                    min_score=config.memory.context_min_score,
+                    max_context_tokens=config.memory.context_max_tokens,
+                )
+                context_messages = inject_context(
+                    user_input,
+                    [] if agent is not None else history,
+                    memory_backend,
+                    config=ctx_cfg,
+                    facts=facts,
+                )
+                if agent is not None:
+                    if context_messages:
+                        agent_context_message = context_messages[0]
+                else:
+                    generation_history = context_messages
+            except Exception:
+                logger.debug("Failed to inject memory context", exc_info=True)
+
+        # Generate response even when optional memory context is unavailable.
         try:
             if agent is not None:
-                response = agent.run(user_input)
+                agent_context = None
+                if agent_context_message is not None:
+                    from openjarvis.agents._stubs import AgentContext
+
+                    agent_context = AgentContext()
+                    agent_context.conversation.add(agent_context_message)
+                response = agent.run(user_input, context=agent_context)
                 content = (
                     response.content if hasattr(response, "content") else str(response)
                 )
             else:
-                result = engine.generate(history, model=model)
+                result = engine.generate(generation_history, model=model)
                 content = (
                     result.get("content", "")
                     if isinstance(result, dict)

src/openjarvis/evals/datasets/taubench.py

@@ -8,13 +8,12 @@ Reference: https://github.com/sierra-research/tau2-bench
 
 from __future__ import annotations
 
+import json
 import logging
 import os
-import subprocess
-import sys
+from importlib import metadata
 from typing import Iterable, List, Optional
 
-from openjarvis.core.paths import get_cache_dir
 from openjarvis.evals.core.dataset import DatasetProvider
 from openjarvis.evals.core.splits import apply_split
 from openjarvis.evals.core.types import EvalRecord
@@ -22,48 +21,50 @@ from openjarvis.evals.core.types import EvalRecord
 LOGGER = logging.getLogger(__name__)
 
 TAU2_REPO = "https://github.com/sierra-research/tau2-bench.git"
-CACHE_DIR = get_cache_dir() / "tau2-bench"
+# v1.0.1. Keep the full commit SHA here (rather than a movable tag) so every
+# TauBench setup uses the same third-party code.
+TAU2_REVISION = "fc0055dc4e0a316c3f83133267fbd6faaa770992"
+TAU2_INSTALL_SPEC = f"tau2 @ git+{TAU2_REPO}@{TAU2_REVISION}"
 
 DOMAINS = ("airline", "retail", "telecom")
 
 
 def _ensure_tau2() -> None:
-    """Ensure tau2 package is importable; install from cache if needed."""
+    """Ensure the explicitly installed, pinned tau2 package is importable."""
+    try:
+        distribution = metadata.distribution("tau2")
+    except metadata.PackageNotFoundError as exc:
+        raise ImportError(
+            "TauBench requires tau2, which OpenJarvis does not install at "
+            "runtime. Install the pinned dependency explicitly (Python >=3.12): "
+            f'uv pip install "{TAU2_INSTALL_SPEC}"'
+        ) from exc
+
+    try:
+        direct_url_text = distribution.read_text("direct_url.json")
+        direct_url = json.loads(direct_url_text or "")
+        vcs_info = direct_url.get("vcs_info", {})
+        installed_repo = direct_url.get("url")
+        installed_revision = vcs_info.get("commit_id")
+    except (json.JSONDecodeError, AttributeError):
+        installed_repo = None
+        installed_revision = None
+
+    if installed_repo != TAU2_REPO or installed_revision != TAU2_REVISION:
+        raise ImportError(
+            "The installed tau2 package does not match OpenJarvis's pinned "
+            "source revision. Reinstall it explicitly (Python >=3.12): "
+            f'uv pip install --force-reinstall "{TAU2_INSTALL_SPEC}"'
+        )
+
     try:
         import tau2  # noqa: F401
-    except ImportError:
-        # Clone and install from source
-        if not CACHE_DIR.exists():
-            LOGGER.info("Cloning tau2-bench from %s ...", TAU2_REPO)
-            CACHE_DIR.parent.mkdir(parents=True, exist_ok=True)
-            subprocess.run(
-                ["git", "clone", "--depth", "1", TAU2_REPO, str(CACHE_DIR)],
-                check=True,
-                capture_output=True,
-            )
-        LOGGER.info("Installing tau2-bench ...")
-        # Try `python -m pip` first; fall back to `uv pip` for uv-managed venvs
-        # which don't ship pip by default.
-        try:
-            subprocess.run(
-                [sys.executable, "-m", "pip", "install", "-e", str(CACHE_DIR)],
-                check=True,
-                capture_output=True,
-            )
-        except (subprocess.CalledProcessError, FileNotFoundError):
-            subprocess.run(
-                [
-                    "uv",
-                    "pip",
-                    "install",
-                    "--python",
-                    sys.executable,
-                    "-e",
-                    str(CACHE_DIR),
-                ],
-                check=True,
-                capture_output=True,
-            )
+    except ImportError as exc:
+        raise ImportError(
+            "The pinned tau2 package is installed but cannot be imported. "
+            "Reinstall it explicitly (Python >=3.12): "
+            f'uv pip install --force-reinstall "{TAU2_INSTALL_SPEC}"'
+        ) from exc
 
 
 class TauBenchDataset(DatasetProvider):

src/openjarvis/memory/__init__.py

@@ -19,6 +19,7 @@ from openjarvis.memory.store import (
     FactStore,
     LocalFactStore,
     create_fact_store,
+    load_configured_facts,
 )
 
 __all__ = [
@@ -29,5 +30,6 @@ __all__ = [
     "MemoryService",
     "build_memory_service",
     "create_fact_store",
+    "load_configured_facts",
     "publish_completed_exchange",
 ]

src/openjarvis/memory/store.py

@@ -16,7 +16,7 @@ import time
 from abc import ABC, abstractmethod
 from dataclasses import asdict, dataclass
 from pathlib import Path
-from typing import Iterable, List
+from typing import Any, Iterable, List
 
 from openjarvis.core.paths import get_config_dir
 from openjarvis.core.registry import FactStoreRegistry
@@ -205,4 +205,30 @@ def create_fact_store(
     return FactStoreRegistry.create(key, path, max_facts=max_facts)
 
 
-__all__ = ["Fact", "FactStore", "LocalFactStore", "create_fact_store"]
+def load_configured_facts(config: Any) -> List[Fact]:
+    """Load automatic-memory facts from *config* when the service is enabled.
+
+    Context injection is also used by short-lived commands such as
+    ``jarvis ask``, where no :class:`MemoryService` instance exists.  This
+    helper gives those callers the same configured fact-store view without
+    coupling them to the service lifecycle.
+    """
+    memory = getattr(config, "memory", None)
+    if memory is None or not getattr(memory, "enabled", False):
+        return []
+
+    store = create_fact_store(
+        getattr(memory, "backend", "local"),
+        path=getattr(memory, "facts_path", None),
+        max_facts=getattr(memory, "max_facts", 1000),
+    )
+    return store.list()
+
+
+__all__ = [
+    "Fact",
+    "FactStore",
+    "LocalFactStore",
+    "create_fact_store",
+    "load_configured_facts",
+]

src/openjarvis/sdk.py

@@ -522,14 +522,15 @@ class Jarvis:
         # Context injection
         if context and self._config.agent.context_from_memory:
             try:
-                from openjarvis.cli.ask import _get_memory_backend
+                from openjarvis.cli.ask import _get_memory_backend, _get_memory_facts
                 from openjarvis.tools.storage.context import (
                     ContextConfig,
                     inject_context,
                 )
 
                 backend = _get_memory_backend(self._config)
-                if backend is not None:
+                facts = _get_memory_facts(self._config)
+                if backend is not None or facts:
                     ctx_cfg = ContextConfig(
                         top_k=self._config.memory.context_top_k,
                         min_score=self._config.memory.context_min_score,
@@ -540,6 +541,7 @@ class Jarvis:
                         [],
                         backend,
                         config=ctx_cfg,
+                        facts=facts,
                     )
                     for msg in context_messages:
                         ctx.conversation.add(msg)
@@ -570,17 +572,24 @@ class Jarvis:
     ) -> List[Message]:
         """Inject memory context into messages."""
         try:
-            from openjarvis.cli.ask import _get_memory_backend
+            from openjarvis.cli.ask import _get_memory_backend, _get_memory_facts
             from openjarvis.tools.storage.context import ContextConfig, inject_context
 
             backend = _get_memory_backend(self._config)
-            if backend is not None:
+            facts = _get_memory_facts(self._config)
+            if backend is not None or facts:
                 ctx_cfg = ContextConfig(
                     top_k=self._config.memory.context_top_k,
                     min_score=self._config.memory.context_min_score,
                     max_context_tokens=self._config.memory.context_max_tokens,
                 )
-                return inject_context(query, messages, backend, config=ctx_cfg)
+                return inject_context(
+                    query,
+                    messages,
+                    backend,
+                    config=ctx_cfg,
+                    facts=facts,
+                )
         except Exception as exc:
             logger.warning("Failed to inject memory context: %s", exc)
         return messages

src/openjarvis/server/routes.py

@@ -11,7 +11,7 @@ from fastapi import APIRouter, HTTPException, Request
 from fastapi.responses import StreamingResponse
 
 from openjarvis.core.paths import get_config_dir
-from openjarvis.core.types import Message, Role
+from openjarvis.core.types import Message, Role, ToolCall
 from openjarvis.server.model_capabilities import is_embed_only_model
 from openjarvis.server.models import (
     ChatCompletionChunk,
@@ -40,6 +40,15 @@ def _to_messages(chat_messages) -> list[Message]:
                 role=role,
                 content=m.content or "",
                 name=m.name,
+                tool_calls=[
+                    ToolCall(
+                        id=tool_call.get("id", ""),
+                        name=tool_call.get("function", {}).get("name", ""),
+                        arguments=tool_call.get("function", {}).get("arguments", "{}"),
+                    )
+                    for tool_call in (m.tool_calls or [])
+                ]
+                or None,
                 tool_call_id=m.tool_call_id,
             )
         )
@@ -114,13 +123,15 @@ async def chat_completions(request_body: ChatCompletionRequest, request: Request
     memory_backend = getattr(request.app.state, "memory_backend", None)
     if (
         config is not None
-        and memory_backend is not None
         and config.agent.context_from_memory
         and request_body.messages
     ):
         try:
             from openjarvis.tools.storage.context import ContextConfig, inject_context
 
+            memory_service = getattr(request.app.state, "memory_service", None)
+            facts = memory_service.list_facts() if memory_service is not None else []
+
             # Extract query from the last user message
             query_text = ""
             for m in reversed(request_body.messages):
@@ -130,6 +141,7 @@ async def chat_completions(request_body: ChatCompletionRequest, request: Request
 
             if query_text:
                 messages = _to_messages(request_body.messages)
+                messages = _ensure_identity_prompt(messages, config)
                 ctx_cfg = ContextConfig(
                     top_k=config.memory.context_top_k,
                     min_score=config.memory.context_min_score,
@@ -140,22 +152,35 @@ async def chat_completions(request_body: ChatCompletionRequest, request: Request
                     messages,
                     memory_backend,
                     config=ctx_cfg,
+                    facts=facts,
                 )
-                # Rebuild request messages from enriched Message objects
-                if len(enriched) > len(messages):
-                    from openjarvis.server.models import ChatMessage
+                # Rebuild after identity/context merging so downstream engine
+                # adapters always receive exactly one system message.
+                from openjarvis.server.models import ChatMessage
 
-                    new_msgs = []
-                    for msg in enriched:
-                        new_msgs.append(
-                            ChatMessage(
-                                role=msg.role.value,
-                                content=msg.content,
-                                name=msg.name,
-                                tool_call_id=getattr(msg, "tool_call_id", None),
-                            )
+                new_msgs = []
+                for msg in enriched:
+                    new_msgs.append(
+                        ChatMessage(
+                            role=msg.role.value,
+                            content=msg.content,
+                            name=msg.name,
+                            tool_calls=[
+                                {
+                                    "id": tool_call.id,
+                                    "type": "function",
+                                    "function": {
+                                        "name": tool_call.name,
+                                        "arguments": tool_call.arguments,
+                                    },
+                                }
+                                for tool_call in (msg.tool_calls or [])
+                            ]
+                            or None,
+                            tool_call_id=getattr(msg, "tool_call_id", None),
                         )
-                    request_body.messages = new_msgs
+                    )
+                request_body.messages = new_msgs
         except Exception:
             logging.getLogger("openjarvis.server").debug(
                 "Memory context injection failed",

src/openjarvis/server/stream_bridge.py

@@ -246,62 +246,15 @@ class AgentStreamBridge:
                     {"results": tool_results_data},
                 )
 
-            # Stream content using real LLM token streaming via
-            # engine.stream_full() when the engine is available.
+            # ``agent.run()`` already produced the authoritative, grounded
+            # response.  Do not call the engine again here: a second inference
+            # would not have the agent's system prompt, tool transcript, or
+            # other internal context and could therefore contradict the
+            # result reported by the agent events.  Replay the final content
+            # in chunks so the OpenAI-compatible streaming response stays
+            # consistent with the completed agent run.
             content = agent_result.content or ""
-            engine = getattr(self._agent, "_engine", None)
-            used_real_streaming = False
-
-            if engine is not None and hasattr(engine, "stream_full") and content:
-                # Re-stream using the engine for real token delivery.
-                # Build the same messages the agent used for its final turn.
-                try:
-                    from openjarvis.core.types import Message as MsgType
-                    from openjarvis.core.types import Role as RoleType
-
-                    replay_messages = []
-                    for m in self._request.messages:
-                        role = (
-                            RoleType(m.role)
-                            if m.role in {r.value for r in RoleType}
-                            else RoleType.USER
-                        )
-                        replay_messages.append(
-                            MsgType(
-                                role=role,
-                                content=m.content or "",
-                                name=m.name,
-                                tool_call_id=m.tool_call_id,
-                            )
-                        )
-
-                    async for sc in engine.stream_full(
-                        replay_messages,
-                        model=self._model,
-                    ):
-                        if sc.content:
-                            chunk = ChatCompletionChunk(
-                                id=self._chunk_id,
-                                model=self._model,
-                                choices=[
-                                    StreamChoice(
-                                        delta=DeltaMessage(content=sc.content),
-                                    )
-                                ],
-                            )
-                            yield f"data: {chunk.model_dump_json()}\n\n"
-                    used_real_streaming = True
-                except Exception as stream_exc:
-                    import logging as _logging
-
-                    _logger = _logging.getLogger("openjarvis.server")
-                    _logger.warning(
-                        "Real streaming failed, falling back to word replay: %s",
-                        stream_exc,
-                    )
-
-            # Fallback: word-by-word replay if real streaming was not used
-            if not used_real_streaming and content:
+            if content:
                 words = content.split(" ")
                 for i, word in enumerate(words):
                     token = word if i == 0 else " " + word

src/openjarvis/system/orchestrator.py

@@ -40,8 +40,9 @@ class QueryOrchestrator:
 
         messages = [Message(role=Role.USER, content=query)]
 
-        if context and s.memory_backend and s.config.agent.context_from_memory:
+        if context and s.config.agent.context_from_memory:
             try:
+                from openjarvis.memory import load_configured_facts
                 from openjarvis.tools.storage.context import (
                     ContextConfig,
                     inject_context,
@@ -52,11 +53,13 @@ class QueryOrchestrator:
                     min_score=s.config.memory.context_min_score,
                     max_context_tokens=s.config.memory.context_max_tokens,
                 )
+                facts = load_configured_facts(s.config)
                 messages = inject_context(
                     query,
                     messages,
                     s.memory_backend,
                     config=ctx_cfg,
+                    facts=facts,
                 )
             except Exception as exc:
                 logger.warning("Failed to inject memory context: %s", exc)

src/openjarvis/tools/storage/context.py

@@ -2,13 +2,16 @@
 
 from __future__ import annotations
 
-from dataclasses import dataclass
-from typing import List, Optional
+from dataclasses import dataclass, replace
+from typing import TYPE_CHECKING, List, Optional, Sequence
 
 from openjarvis.core.events import EventType, get_event_bus
 from openjarvis.core.types import Message, Role
 from openjarvis.tools.storage._stubs import MemoryBackend, RetrievalResult
 
+if TYPE_CHECKING:
+    from openjarvis.memory.store import Fact
+
 
 @dataclass(slots=True)
 class ContextConfig:
@@ -46,28 +49,75 @@ def format_context(results: List[RetrievalResult]) -> str:
 
 def build_context_message(
     results: List[RetrievalResult],
+    facts: Sequence[Fact] = (),
 ) -> Message:
     """Create a system message with formatted context."""
-    context_text = format_context(results)
-    content = (
-        "The following context was retrieved from the knowledge"
-        " base. Use it to inform your response, citing sources"
-        " where applicable:\n\n" + context_text
+    sections = []
+    if facts:
+        fact_text = "\n".join(f"- {fact.text}" for fact in facts)
+        sections.append(
+            "The following durable facts were remembered from prior "
+            "conversations. Use them when relevant to the user's request:\n\n"
+            + fact_text
+        )
+    if results:
+        sections.append(
+            "The following context was retrieved from the knowledge"
+            " base. Use it to inform your response, citing sources"
+            " where applicable:\n\n" + format_context(results)
+        )
+    content = "\n\n".join(sections)
+    return Message(
+        role=Role.SYSTEM,
+        content=content,
+        metadata={"memory_context": True},
     )
-    return Message(role=Role.SYSTEM, content=content)
+
+
+def _merge_context_message(
+    messages: List[Message],
+    context_message: Message,
+) -> List[Message]:
+    """Return a copy with context folded into the existing system prompt."""
+    system_messages = [message for message in messages if message.role == Role.SYSTEM]
+    if not system_messages:
+        return [context_message, *messages]
+
+    content = "\n\n".join(
+        part
+        for part in (
+            *(message.text for message in system_messages),
+            context_message.text,
+        )
+        if part
+    )
+    combined = replace(system_messages[0], content=content)
+    merged: List[Message] = []
+    inserted = False
+    for message in messages:
+        if message.role == Role.SYSTEM:
+            if not inserted:
+                merged.append(combined)
+                inserted = True
+            continue
+        merged.append(message)
+    return merged
 
 
 def inject_context(
     query: str,
     messages: List[Message],
-    backend: MemoryBackend,
+    backend: Optional[MemoryBackend],
     *,
     config: Optional[ContextConfig] = None,
+    facts: Sequence[Fact] = (),
 ) -> List[Message]:
     """Retrieve relevant context and prepend it to *messages*.
 
     Returns a **new** list — the original list is not mutated.
-    If no results pass the score threshold, returns the original
+    Automatic-memory facts are included independently of the retrieval
+    backend, so persisted facts remain recallable even when the document
+    store is empty. If no facts or results are available, returns the original
     messages unchanged.
 
     Parameters
@@ -77,33 +127,55 @@ def inject_context(
     messages:
         The existing message list.
     backend:
-        The memory backend to search.
+        The memory backend to search, or ``None`` when only facts are available.
     config:
         Context injection settings (uses defaults if ``None``).
+    facts:
+        Durable facts captured by the automatic memory service.
     """
     cfg = config or ContextConfig()
     if not cfg.enabled:
         return messages
 
-    results = backend.retrieve(query, top_k=cfg.top_k)
+    results = backend.retrieve(query, top_k=cfg.top_k) if backend is not None else []
 
     # Filter by minimum score
     results = [r for r in results if r.score >= cfg.min_score]
 
-    if not results:
-        return messages
-
-    # Truncate to max_context_tokens
-    truncated: List[RetrievalResult] = []
+    # When both sources have data, cap facts at half the total budget so they
+    # cannot starve query-specific document retrieval. Unused fact budget is
+    # still available to documents. Newest facts win within the fact budget.
+    fact_budget = cfg.max_context_tokens
+    if results:
+        fact_budget //= 2
+    selected_facts: List[Fact] = []
     total_tokens = 0
+    for fact in reversed(facts):
+        tokens = _count_tokens(fact.text)
+        if total_tokens + tokens > fact_budget:
+            continue
+        selected_facts.append(fact)
+        total_tokens += tokens
+
+    # Fill the remaining context budget with retrieved documents.
+    truncated: List[RetrievalResult] = []
     for r in results:
         tokens = _count_tokens(r.content)
+        if total_tokens + tokens > cfg.max_context_tokens:
+            # A large top result should not disappear solely because facts
+            # consumed their reserved share. Prefer that result when it fits
+            # the total budget on its own.
+            if not truncated and selected_facts and tokens <= cfg.max_context_tokens:
+                selected_facts = []
+                total_tokens = 0
+            else:
+                break
         if total_tokens + tokens > cfg.max_context_tokens:
             break
         truncated.append(r)
         total_tokens += tokens
 
-    if not truncated:
+    if not selected_facts and not truncated:
         return messages
 
     # Publish event
@@ -114,13 +186,14 @@ def inject_context(
             "context_injection": True,
             "query": query,
             "num_results": len(truncated),
+            "num_facts": len(selected_facts),
             "total_tokens": total_tokens,
         },
     )
 
     # Build context message and prepend
-    ctx_msg = build_context_message(truncated)
-    return [ctx_msg] + list(messages)
+    ctx_msg = build_context_message(truncated, selected_facts)
+    return _merge_context_message(messages, ctx_msg)
 
 
 __all__ = [

tests/agents/test_base_agent.py

@@ -205,6 +205,40 @@ class TestBuildMessages:
         assert messages[1].content == "prev"
         assert messages[2].content == "new"
 
+    def test_prompt_builder_merges_context_system_message(self):
+        engine = MagicMock()
+        prompt_builder = MagicMock()
+        prompt_builder.build.return_value = "You are OpenJarvis."
+        agent = _ConcreteAgent(engine, "m", prompt_builder=prompt_builder)
+        conv = Conversation()
+        conv.add(
+            Message(
+                role=Role.SYSTEM,
+                content="Remember: user likes jazz.",
+                metadata={"memory_context": True},
+            )
+        )
+        ctx = AgentContext(conversation=conv)
+
+        messages = agent._build_messages("new", ctx)
+
+        system_messages = [m for m in messages if m.role == Role.SYSTEM]
+        assert len(system_messages) == 1
+        assert "You are OpenJarvis." in system_messages[0].content
+        assert "user likes jazz" in system_messages[0].content
+
+    def test_prompt_builder_preserves_caller_system_context(self):
+        engine = MagicMock()
+        prompt_builder = MagicMock()
+        prompt_builder.build.return_value = "Agent instructions."
+        agent = _ConcreteAgent(engine, "m", prompt_builder=prompt_builder)
+        conv = Conversation()
+        conv.add(Message(role=Role.SYSTEM, content="You are helpful."))
+
+        messages = agent._build_messages("new", AgentContext(conversation=conv))
+
+        assert any(message.content == "You are helpful." for message in messages)
+
 
 class TestGenerate:
     def test_delegates_to_engine(self):

tests/cli/test_chat_cmd.py

@@ -18,6 +18,7 @@ from openjarvis.core.config import JarvisConfig
 from openjarvis.core.events import Event, EventBus, EventType
 from openjarvis.core.registry import AgentRegistry, ToolRegistry
 from openjarvis.core.types import ToolCall, ToolResult
+from openjarvis.memory.store import LocalFactStore
 from openjarvis.tools._stubs import BaseTool, ToolSpec
 
 
@@ -97,6 +98,79 @@ class TestReadInput:
 
 
 class TestChatAgents:
+    def test_direct_chat_injects_auto_memory_facts(self, tmp_path) -> None:
+        facts_path = tmp_path / "facts.jsonl"
+        LocalFactStore(facts_path).add(
+            "The user's favorite color is blue",
+            source="auto",
+        )
+
+        engine = MagicMock()
+        engine.engine_id = "mock"
+        engine.generate.return_value = {"content": "Blue."}
+        config = JarvisConfig()
+        config.intelligence.default_model = "test-model"
+        config.memory.enabled = True
+        config.memory.facts_path = str(facts_path)
+        config.agent.context_from_memory = True
+
+        with (
+            patch("openjarvis.cli.chat_cmd.load_config", return_value=config),
+            patch("openjarvis.engine.get_engine", return_value=("mock", engine)),
+            patch("openjarvis.intelligence.register_builtin_models"),
+            patch("openjarvis.memory.build_memory_service", return_value=None),
+            patch("openjarvis.cli.ask._get_memory_backend", return_value=None),
+        ):
+            result = CliRunner().invoke(
+                chat,
+                ["--model", "test-model"],
+                input="What is my favorite color?\n/quit\n",
+            )
+
+        assert result.exit_code == 0
+        messages = engine.generate.call_args.args[0]
+        assert messages[0].role.value == "system"
+        assert "favorite color is blue" in messages[0].content
+
+    def test_chat_generation_survives_fact_store_failure(self) -> None:
+        class _FailingMemoryService:
+            def start(self) -> None:
+                pass
+
+            def stop(self, timeout: float = 2.0) -> None:
+                pass
+
+            def list_facts(self):
+                raise OSError("fact store unavailable")
+
+        engine = MagicMock()
+        engine.engine_id = "mock"
+        engine.generate.return_value = {"content": "Still working."}
+        config = JarvisConfig()
+        config.intelligence.default_model = "test-model"
+        config.memory.enabled = True
+        config.agent.context_from_memory = True
+
+        with (
+            patch("openjarvis.cli.chat_cmd.load_config", return_value=config),
+            patch("openjarvis.engine.get_engine", return_value=("mock", engine)),
+            patch("openjarvis.intelligence.register_builtin_models"),
+            patch(
+                "openjarvis.memory.build_memory_service",
+                return_value=_FailingMemoryService(),
+            ),
+            patch("openjarvis.cli.ask._get_memory_backend", return_value=None),
+        ):
+            result = CliRunner().invoke(
+                chat,
+                ["--model", "test-model"],
+                input="hello\n/quit\n",
+            )
+
+        assert result.exit_code == 0
+        assert "Still working." in result.output
+        engine.generate.assert_called_once()
+
     def test_simple_agent_does_not_receive_tool_only_kwargs(self) -> None:
         engine = MagicMock()
         engine.engine_id = "mock"

tests/evals/datasets/test_taubench.py

@@ -0,0 +1,98 @@
+"""Tests for the TauBench optional dependency boundary."""
+
+from __future__ import annotations
+
+import builtins
+import sys
+from types import ModuleType
+from unittest.mock import Mock
+
+import pytest
+
+from openjarvis.evals.datasets import taubench
+
+
+def _mock_direct_url(monkeypatch, direct_url):
+    distribution = Mock()
+    distribution.read_text.return_value = direct_url
+    monkeypatch.setattr(
+        taubench.metadata, "distribution", Mock(return_value=distribution)
+    )
+
+
+def test_ensure_tau2_accepts_the_pinned_source_revision(monkeypatch):
+    monkeypatch.setitem(sys.modules, "tau2", ModuleType("tau2"))
+    _mock_direct_url(
+        monkeypatch,
+        (
+            '{"url": "https://github.com/sierra-research/tau2-bench.git", '
+            '"vcs_info": {"vcs": "git", '
+            f'"commit_id": "{taubench.TAU2_REVISION}"}}}}'
+        ),
+    )
+
+    taubench._ensure_tau2()
+
+
+def test_ensure_tau2_requires_explicit_pinned_install(monkeypatch):
+    monkeypatch.setitem(sys.modules, "tau2", None)
+    monkeypatch.setattr(
+        taubench.metadata,
+        "distribution",
+        Mock(side_effect=taubench.metadata.PackageNotFoundError),
+    )
+
+    with pytest.raises(ImportError) as exc_info:
+        taubench._ensure_tau2()
+
+    message = str(exc_info.value)
+    assert "does not install at runtime" in message
+    assert taubench.TAU2_REVISION in message
+    assert "uv pip install" in message
+
+
+@pytest.mark.parametrize(
+    "direct_url",
+    [
+        # Editable install left behind by the previous runtime installer.
+        '{"url": "file:///home/user/.openjarvis/cache/tau2-bench", '
+        '"dir_info": {"editable": true}}',
+        # A git install from an arbitrary upstream revision.
+        '{"url": "https://github.com/sierra-research/tau2-bench.git", '
+        '"vcs_info": {"vcs": "git", "commit_id": "deadbeef"}}',
+        # Registry installs do not carry PEP 610 direct-origin metadata.
+        None,
+    ],
+)
+def test_ensure_tau2_rejects_unpinned_install(monkeypatch, direct_url):
+    _mock_direct_url(monkeypatch, direct_url)
+    original_import = builtins.__import__
+
+    def guarded_import(name, *args, **kwargs):
+        if name == "tau2":
+            raise AssertionError("unverified tau2 package was imported")
+        return original_import(name, *args, **kwargs)
+
+    monkeypatch.setattr(builtins, "__import__", guarded_import)
+
+    with pytest.raises(ImportError) as exc_info:
+        taubench._ensure_tau2()
+
+    message = str(exc_info.value)
+    assert "does not match" in message
+    assert taubench.TAU2_REVISION in message
+    assert "--force-reinstall" in message
+
+
+def test_verify_requirements_reports_install_instruction(monkeypatch):
+    monkeypatch.setitem(sys.modules, "tau2", None)
+    monkeypatch.setattr(
+        taubench.metadata,
+        "distribution",
+        Mock(side_effect=taubench.metadata.PackageNotFoundError),
+    )
+
+    issues = taubench.TauBenchDataset().verify_requirements()
+
+    assert len(issues) == 1
+    assert taubench.TAU2_REVISION in issues[0]

tests/memory/test_context.py

@@ -7,6 +7,7 @@ from typing import Any, Dict, List, Optional
 
 from openjarvis.core.events import EventBus, EventType
 from openjarvis.core.types import Message, Role
+from openjarvis.memory.store import Fact
 from openjarvis.tools.storage._stubs import MemoryBackend, RetrievalResult
 from openjarvis.tools.storage.context import (
     ContextConfig,
@@ -167,6 +168,113 @@ def test_inject_context_no_results_returns_original():
     assert augmented is messages
 
 
+def test_inject_context_adds_auto_memory_facts_without_backend():
+    messages = [Message(role=Role.USER, content="What is my favorite color?")]
+    facts = [Fact(text="The user's favorite color is blue", source="auto")]
+
+    augmented = inject_context("favorite color", messages, None, facts=facts)
+
+    assert len(augmented) == 2
+    assert augmented[0].role == Role.SYSTEM
+    assert "remembered from prior conversations" in augmented[0].content
+    assert "favorite color is blue" in augmented[0].content
+
+
+def test_inject_context_prioritizes_newest_facts_within_token_budget():
+    messages = [Message(role=Role.USER, content="What do you remember?")]
+    facts = [
+        Fact(text="old fact uses four tokens"),
+        Fact(text="new fact uses four tokens"),
+    ]
+
+    augmented = inject_context(
+        "remember",
+        messages,
+        None,
+        config=ContextConfig(max_context_tokens=5),
+        facts=facts,
+    )
+
+    assert "new fact uses four tokens" in augmented[0].content
+    assert "old fact uses four tokens" not in augmented[0].content
+
+
+def test_inject_context_merges_with_existing_system_message():
+    messages = [
+        Message(role=Role.SYSTEM, content="You are OpenJarvis."),
+        Message(role=Role.USER, content="What is my favorite color?"),
+    ]
+    facts = [Fact(text="The user's favorite color is blue")]
+
+    augmented = inject_context("favorite color", messages, None, facts=facts)
+
+    system_messages = [m for m in augmented if m.role == Role.SYSTEM]
+    assert len(system_messages) == 1
+    assert "You are OpenJarvis." in system_messages[0].content
+    assert "favorite color is blue" in system_messages[0].content
+    assert messages[0].content == "You are OpenJarvis."
+
+
+def test_inject_context_collapses_multiple_system_messages():
+    messages = [
+        Message(role=Role.SYSTEM, content="Identity."),
+        Message(role=Role.SYSTEM, content="Persona."),
+        Message(role=Role.USER, content="What do you remember?"),
+    ]
+
+    augmented = inject_context(
+        "remember",
+        messages,
+        None,
+        facts=[Fact(text="User likes jazz")],
+    )
+
+    system_messages = [m for m in augmented if m.role == Role.SYSTEM]
+    assert len(system_messages) == 1
+    assert "Identity." in system_messages[0].content
+    assert "Persona." in system_messages[0].content
+    assert "User likes jazz" in system_messages[0].content
+
+
+def test_inject_context_reserves_budget_for_retrieved_documents():
+    backend = _FakeMemory(
+        [RetrievalResult(content="d1 d2 d3 d4 d5", score=1.0, source="doc")]
+    )
+    facts = [
+        Fact(text="old1 old2 old3 old4 old5"),
+        Fact(text="new1 new2 new3 new4 new5"),
+    ]
+
+    augmented = inject_context(
+        "query",
+        [Message(role=Role.USER, content="query")],
+        backend,
+        config=ContextConfig(max_context_tokens=10),
+        facts=facts,
+    )
+
+    assert "new1 new2 new3 new4 new5" in augmented[0].content
+    assert "d1 d2 d3 d4 d5" in augmented[0].content
+    assert "old1 old2 old3 old4 old5" not in augmented[0].content
+
+
+def test_inject_context_prefers_large_document_that_fits_total_budget():
+    backend = _FakeMemory(
+        [RetrievalResult(content="d1 d2 d3 d4 d5 d6 d7 d8", score=1.0)]
+    )
+
+    augmented = inject_context(
+        "query",
+        [Message(role=Role.USER, content="query")],
+        backend,
+        config=ContextConfig(max_context_tokens=10),
+        facts=[Fact(text="f1 f2 f3 f4 f5")],
+    )
+
+    assert "d1 d2 d3 d4 d5 d6 d7 d8" in augmented[0].content
+    assert "f1 f2 f3 f4 f5" not in augmented[0].content
+
+
 def test_inject_context_publishes_event():
     bus = EventBus(record_history=True)
     results = [

tests/memory/test_fact_store.py

@@ -7,7 +7,11 @@ import json
 import pytest
 
 from openjarvis.core.registry import FactStoreRegistry
-from openjarvis.memory.store import LocalFactStore, create_fact_store
+from openjarvis.memory.store import (
+    LocalFactStore,
+    create_fact_store,
+    load_configured_facts,
+)
 
 
 def test_add_and_list(tmp_path):
@@ -145,3 +149,28 @@ def test_create_fact_store_default_path_uses_openjarvis_home(tmp_path, monkeypat
 def test_create_fact_store_unknown_backend(tmp_path):
     with pytest.raises(ValueError):
         create_fact_store("cloud", path=tmp_path / "f.jsonl")
+
+
+def test_load_configured_facts_reads_enabled_store(tmp_path):
+    from types import SimpleNamespace
+
+    path = tmp_path / "facts.jsonl"
+    LocalFactStore(path).add("User likes jazz", source="auto")
+    config = SimpleNamespace(
+        memory=SimpleNamespace(
+            enabled=True,
+            backend="local",
+            facts_path=str(path),
+            max_facts=1000,
+        )
+    )
+
+    assert [fact.text for fact in load_configured_facts(config)] == ["User likes jazz"]
+
+
+def test_load_configured_facts_skips_disabled_memory():
+    from types import SimpleNamespace
+
+    config = SimpleNamespace(memory=SimpleNamespace(enabled=False))
+
+    assert load_configured_facts(config) == []

tests/server/test_routes.py

@@ -11,6 +11,7 @@ fastapi = pytest.importorskip("fastapi")
 from fastapi.testclient import TestClient  # noqa: E402
 
 from openjarvis.core.events import EventBus, EventType  # noqa: E402
+from openjarvis.core.types import Role  # noqa: E402
 from openjarvis.server.app import create_app  # noqa: E402
 
 # ---------------------------------------------------------------------------
@@ -846,6 +847,47 @@ class TestIdentityPromptInjection:
         assert len(system_msgs) == 1
         assert system_msgs[0].content == "Be terse."
 
+    def test_stream_uses_grounded_agent_result_without_replay(self):
+        """Regression for #734: web streaming emits the agent's final answer."""
+        from openjarvis.core.events import EventBus
+
+        captured: list = []
+        engine = _make_capturing_engine(captured)
+        agent = _make_agent(content="My name is Jarvis Prime.")
+        agent._tools = [object()]
+        agent._engine = engine
+        client = TestClient(
+            create_app(
+                engine,
+                "test-model",
+                agent=agent,
+                bus=EventBus(),
+                config=_identity_config(),
+            )
+        )
+
+        resp = client.post(
+            "/v1/chat/completions",
+            json={
+                "model": "test-model",
+                "messages": [{"role": "user", "content": "who are you?"}],
+                "stream": True,
+            },
+        )
+
+        assert resp.status_code == 200
+        streamed_content = ""
+        for line in resp.text.splitlines():
+            if not line.startswith("data: {"):
+                continue
+            payload = json.loads(line.removeprefix("data: "))
+            choices = payload.get("choices", [])
+            if choices and choices[0]["delta"].get("content"):
+                streamed_content += choices[0]["delta"]["content"]
+        assert streamed_content == "My name is Jarvis Prime."
+        assert captured == []
+        agent.run.assert_called_once()
+
     def test_direct_injects_identity_when_absent(self):
         captured: list = []
         engine = _make_capturing_engine(captured)
@@ -886,6 +928,101 @@ class TestIdentityPromptInjection:
         assert len(system_msgs) == 1
         assert system_msgs[0].content == "Be terse."
 
+    def test_direct_merges_identity_and_auto_memory_into_one_system_message(self):
+        from openjarvis.memory.store import Fact
+
+        class _MemoryService:
+            def list_facts(self):
+                return [Fact(text="The user's favorite color is blue")]
+
+        captured: list = []
+        engine = _make_capturing_engine(captured)
+        cfg = _identity_config()
+        cfg.agent.context_from_memory = True
+        client = TestClient(
+            create_app(
+                engine,
+                "test-model",
+                config=cfg,
+                memory_service=_MemoryService(),
+            )
+        )
+
+        resp = client.post(
+            "/v1/chat/completions",
+            json={
+                "model": "test-model",
+                "messages": [{"role": "user", "content": "What is my favorite color?"}],
+            },
+        )
+
+        assert resp.status_code == 200
+        messages = engine.generate.call_args.args[0]
+        system_messages = [m for m in messages if m.role == Role.SYSTEM]
+        assert len(system_messages) == 1
+        assert "OpenJarvis" in system_messages[0].content
+        assert "favorite color is blue" in system_messages[0].content
+
+    def test_memory_context_preserves_assistant_tool_calls(self):
+        from openjarvis.memory.store import Fact
+
+        class _MemoryService:
+            def list_facts(self):
+                return [Fact(text="User likes jazz")]
+
+        captured: list = []
+        engine = _make_capturing_engine(captured)
+        cfg = _identity_config()
+        cfg.agent.context_from_memory = True
+        client = TestClient(
+            create_app(
+                engine,
+                "test-model",
+                config=cfg,
+                memory_service=_MemoryService(),
+            )
+        )
+
+        resp = client.post(
+            "/v1/chat/completions",
+            json={
+                "model": "test-model",
+                "messages": [
+                    {"role": "user", "content": "Run the lookup"},
+                    {
+                        "role": "assistant",
+                        "content": "",
+                        "tool_calls": [
+                            {
+                                "id": "call_1",
+                                "type": "function",
+                                "function": {
+                                    "name": "lookup",
+                                    "arguments": '{"query":"jazz"}',
+                                },
+                            }
+                        ],
+                    },
+                    {
+                        "role": "tool",
+                        "content": "result",
+                        "tool_call_id": "call_1",
+                    },
+                    {"role": "user", "content": "What did it find?"},
+                ],
+            },
+        )
+
+        assert resp.status_code == 200
+        messages = engine.generate.call_args.args[0]
+        assistant = next(
+            message for message in messages if message.role == Role.ASSISTANT
+        )
+        assert assistant.tool_calls is not None
+        assert assistant.tool_calls[0].id == "call_1"
+        assert assistant.tool_calls[0].name == "lookup"
+        assert assistant.tool_calls[0].arguments == '{"query":"jazz"}'
+
     def test_direct_injects_soul_persona_when_present(self, tmp_path):
         """Regression: /v1/chat/completions previously injected only the bare
         ``default_system_prompt`` blurb via a hand-rolled lookup, bypassing

tests/server/test_stream_bridge.py

@@ -1,6 +1,68 @@
-import json
+"""Regression tests for streaming completed agent responses."""
 
-from openjarvis.server.stream_bridge import AgentStreamBridge
+from __future__ import annotations
+
+import asyncio
+import json
+from unittest.mock import MagicMock
+
+import pytest
+
+pytest.importorskip("fastapi")
+
+from openjarvis.agents._stubs import AgentResult  # noqa: E402
+from openjarvis.core.events import EventBus  # noqa: E402
+from openjarvis.core.types import ToolResult  # noqa: E402
+from openjarvis.server.models import ChatCompletionRequest  # noqa: E402
+from openjarvis.server.stream_bridge import AgentStreamBridge  # noqa: E402
+
+
+def _streamed_content(events: list[str]) -> str:
+    """Join assistant content from OpenAI-compatible data chunks."""
+    content = []
+    for event in events:
+        if not event.startswith("data: {"):
+            continue
+        payload = json.loads(event.removeprefix("data: ").strip())
+        choices = payload.get("choices")
+        if choices and choices[0]["delta"].get("content"):
+            content.append(choices[0]["delta"]["content"])
+    return "".join(content)
+
+
+def test_stream_replays_grounded_agent_result_without_second_inference():
+    grounded_content = "My name is Jarvis. The tool reports 72 degrees."
+    agent = MagicMock()
+    agent._model = "configured-model"
+    agent.run.return_value = AgentResult(
+        content=grounded_content,
+        tool_results=[
+            ToolResult(tool_name="weather", content="72 degrees", success=True)
+        ],
+        metadata={"prompt_tokens": 10, "completion_tokens": 12, "total_tokens": 22},
+    )
+
+    async def ungrounded_replay(*args, **kwargs):
+        raise AssertionError("stream_full must not run after agent.run")
+        yield  # pragma: no cover
+
+    agent._engine.stream_full = ungrounded_replay
+    request = ChatCompletionRequest(
+        model="requested-model",
+        messages=[{"role": "user", "content": "Who are you, and what's outside?"}],
+        stream=True,
+    )
+    bridge = AgentStreamBridge(agent, EventBus(), request.model, request)
+
+    async def collect_events() -> list[str]:
+        return [event async for event in bridge.stream()]
+
+    events = asyncio.run(collect_events())
+
+    assert _streamed_content(events) == grounded_content
+    assert any(event.startswith("event: tool_results\n") for event in events)
+    agent.run.assert_called_once()
+    assert agent._model == "configured-model"
 
 
 def test_tool_call_start_serializes_arguments_for_sse_without_mutating_event():