fix(review): isolate the background-review fork from the canonical session

The forked skill/memory review agent shares the parent's session_id for
prompt-cache warmth. Without isolation it wrote its harness turn ('Review the
conversation above and update the skill library…') plus its curator-mode reply
straight into the user's REAL session in state.db; the next live turn re-read
that injected user message as a standing instruction and the agent 'became' the
curator, refusing the actual task.

Root fix: a _persist_disabled flag on the fork that hard-stops every DB write
and lazy-open path (_flush_messages_to_session_db, _ensure_db_session,
_get_session_db_for_recall) — the review writes only to the skill/memory stores
via its tools. Defense-in-depth: _strip_background_review_harness drops any
stray harness message (and the assistant reply that followed) at load time in
get_messages_as_conversation, so an already-polluted session resumes clean.

Salvaged from #50296.

Co-authored-by: arminanton <29869547+arminanton@users.noreply.github.com>

agent/agent_init.py

@@ -1167,6 +1167,11 @@ def init_agent(
     # continuation row that must remain open after the helper is torn down;
     # those callers explicitly set this flag to False.
     agent._end_session_on_close = True
+    # When True, this agent NEVER persists to the canonical session store
+    # (state.db) or the JSON snapshot, regardless of session_id. Set on the
+    # background skill/memory review fork so its harness turn can't leak into
+    # the user's real session and hijack the next live turn. Default False.
+    agent._persist_disabled = False
     agent._session_init_model_config = {
         "max_iterations": agent.max_iterations,
         "reasoning_config": reasoning_config,

agent/background_review.py

@@ -674,6 +674,20 @@ def _run_review_in_thread(
             review_agent._user_profile_enabled = agent._user_profile_enabled
             review_agent._memory_nudge_interval = 0
             review_agent._skill_nudge_interval = 0
+            # PERSISTENCE ISOLATION (the curator-takeover root cause): the fork
+            # shares the parent's session_id (set below, for prompt-cache
+            # warmth), so without this it would write its harness turn ("Review
+            # the conversation above and update the skill library…") + its own
+            # response straight into the user's REAL session in state.db. On the
+            # user's next live turn the agent re-reads that injected user message
+            # as a standing instruction and "becomes" the curator, refusing the
+            # actual task. _persist_disabled hard-stops every DB write/lazy-open
+            # path (_flush_messages_to_session_db, _ensure_db_session,
+            # _get_session_db_for_recall); the review writes only to the skill
+            # and memory stores via its tools, which is all it needs.
+            review_agent._persist_disabled = True
+            review_agent._session_db = None
+            review_agent._session_json_enabled = False
             # Suppress all status/warning emits from the fork so the
             # user only sees the final successful-action summary.
             # Without this, mid-review "Iteration budget exhausted",

hermes_state.py

@@ -204,6 +204,61 @@ def get_last_init_error() -> Optional[str]:
     return _last_init_error
 
 
+# Distinctive opening shared by both background-review harness prompts
+# (_SKILL_REVIEW_PROMPT and _MEMORY_REVIEW_PROMPT in agent/background_review.py).
+# Matched case-sensitively against the leading content of a user/system message.
+_REVIEW_HARNESS_PREFIXES = (
+    "Review the conversation above and update the skill library",
+    "Review the conversation above and consider saving to memory",
+)
+
+
+def _is_background_review_harness_message(msg: Dict[str, Any]) -> bool:
+    """True when ``msg`` is a persisted background-review harness prompt.
+
+    These are user/system turns the forked skill/memory review agent wrote into
+    a real session in older builds (before the ``_persist_disabled`` isolation
+    fix). They instruct the agent to act as the curator under a hard tool
+    restriction, so replaying them as live history hijacks the session.
+    """
+    if not isinstance(msg, dict):
+        return False
+    if msg.get("role") not in {"user", "system"}:
+        return False
+    content = msg.get("content")
+    if not isinstance(content, str):
+        return False
+    head = content.lstrip()
+    return any(head.startswith(p) for p in _REVIEW_HARNESS_PREFIXES)
+
+
+def _strip_background_review_harness(
+    messages: List[Dict[str, Any]],
+) -> List[Dict[str, Any]]:
+    """Drop background-review harness messages and the curator-mode assistant
+    reply that immediately followed each one.
+
+    Walk the list once; when a harness user/system message is found, skip it and
+    also skip the next message if it is the assistant turn that answered it.
+    Everything else passes through untouched and in order.
+    """
+    if not messages:
+        return messages
+    out: List[Dict[str, Any]] = []
+    skip_next_assistant = False
+    for msg in messages:
+        if _is_background_review_harness_message(msg):
+            skip_next_assistant = True
+            continue
+        if skip_next_assistant:
+            skip_next_assistant = False
+            if isinstance(msg, dict) and msg.get("role") == "assistant":
+                # The curator-mode reply to the harness prompt — drop it.
+                continue
+        out.append(msg)
+    return out
+
+
 def format_session_db_unavailable(prefix: str = "Session database not available") -> str:
     """Format a user-facing 'session DB unavailable' message with cause.
 
@@ -3721,6 +3776,17 @@ class SessionDB:
             if include_ancestors and self._is_duplicate_replayed_user_message(messages, msg):
                 continue
             messages.append(msg)
+        # DEFENSE-IN-DEPTH against background-review session pollution: a forked
+        # skill/memory review that (in older builds, before the _persist_disabled
+        # fix) shared the parent's session_id wrote its harness turn into this
+        # real session. The harness is a user/system message instructing the
+        # agent to "Review the conversation above and update the skill library /
+        # save to memory" under a hard tool restriction; re-loading it as live
+        # history makes the agent adopt the curator role and refuse the user's
+        # actual task. Strip any such harness message AND the curator-mode
+        # assistant reply immediately following it, so a polluted session
+        # resumes clean even if stray rows exist.
+        messages = _strip_background_review_harness(messages)
         return messages
 
     def _session_lineage_root_to_tip(self, session_id: str) -> List[str]:

run_agent.py

@@ -580,6 +580,12 @@ class AIAgent:
         opening the default state DB instead of making the advertised
         ``session_search`` tool unusable.
         """
+        # Persistence-isolated forks (background review) must not lazily open the
+        # canonical state DB: doing so would re-arm _flush_messages_to_session_db
+        # to write the fork's harness turn into the user's real session. Recall
+        # degrades to None for them (they don't use session_search anyway).
+        if getattr(self, "_persist_disabled", False):
+            return None
         if self._session_db is not None:
             return self._session_db
         try:
@@ -593,6 +599,8 @@ class AIAgent:
 
     def _ensure_db_session(self) -> None:
         """Create session DB row on first use. Disables _session_db on failure."""
+        if getattr(self, "_persist_disabled", False):
+            return
         if self._session_db_created or not self._session_db:
             return
         source = _session_source_for_agent(self.platform)
@@ -1723,6 +1731,15 @@ class AIAgent:
         edits a persisted message's content/role in place expecting a re-write
         (in-place compaction resets the seed and re-diffs by identity).
         """
+        # Persistence-isolated agents (e.g. the background skill/memory review
+        # fork) must NEVER write into the canonical session store. The fork
+        # shares the parent's session_id for prompt-cache warmth, so any write
+        # here would land its harness turn ("Review the conversation above and
+        # update the skill library…") inside the user's real session history,
+        # where the next live turn re-reads it as an instruction and the agent
+        # "becomes" the curator. Hard-stop before any DB touch.
+        if getattr(self, "_persist_disabled", False):
+            return
         if not self._session_db:
             return
         self._apply_persist_user_message_override(messages)

tests/test_background_review_session_isolation.py

@@ -0,0 +1,104 @@
+"""Tests for background-review session-store isolation (hermes_state).
+
+The background skill/memory review fork shares the parent's ``session_id`` for
+prompt-cache warmth. Without the ``_persist_disabled`` isolation it wrote its
+harness turn ("Review the conversation above and update the skill library…")
+plus its curator-mode reply into the user's REAL session, and the next live
+turn re-read that injected user message as a standing instruction — the agent
+"became" the curator and refused the actual task.
+
+``_strip_background_review_harness`` is the load-on-read defense-in-depth that
+removes any such stray harness message (and the assistant reply that followed
+it) so a polluted session resumes clean.
+"""
+
+from hermes_state import (
+    _is_background_review_harness_message,
+    _strip_background_review_harness,
+)
+
+
+class TestIsBackgroundReviewHarnessMessage:
+    def test_matches_skill_review_prompt(self):
+        msg = {"role": "user", "content": "Review the conversation above and update the skill library now."}
+        assert _is_background_review_harness_message(msg) is True
+
+    def test_matches_memory_review_prompt(self):
+        msg = {"role": "system", "content": "Review the conversation above and consider saving to memory."}
+        assert _is_background_review_harness_message(msg) is True
+
+    def test_matches_after_leading_whitespace(self):
+        msg = {"role": "user", "content": "\n\n   Review the conversation above and update the skill library."}
+        assert _is_background_review_harness_message(msg) is True
+
+    def test_ignores_normal_user_message(self):
+        msg = {"role": "user", "content": "Please review my PR and update the changelog."}
+        assert _is_background_review_harness_message(msg) is False
+
+    def test_ignores_assistant_role(self):
+        # An assistant message that quotes the harness text is not itself a harness prompt.
+        msg = {"role": "assistant", "content": "Review the conversation above and update the skill library"}
+        assert _is_background_review_harness_message(msg) is False
+
+    def test_ignores_non_string_content(self):
+        msg = {"role": "user", "content": [{"type": "text", "text": "Review the conversation above and update the skill library"}]}
+        assert _is_background_review_harness_message(msg) is False
+
+    def test_ignores_non_dict(self):
+        assert _is_background_review_harness_message("not a dict") is False  # type: ignore[arg-type]
+
+
+class TestStripBackgroundReviewHarness:
+    def test_strips_harness_and_following_assistant_reply(self):
+        messages = [
+            {"role": "user", "content": "What's the weather?"},
+            {"role": "assistant", "content": "It's sunny."},
+            {"role": "user", "content": "Review the conversation above and update the skill library."},
+            {"role": "assistant", "content": "Nothing to save."},
+            {"role": "user", "content": "Thanks, now book a flight."},
+        ]
+        out = _strip_background_review_harness(messages)
+        contents = [m["content"] for m in out]
+        assert contents == ["What's the weather?", "It's sunny.", "Thanks, now book a flight."]
+
+    def test_strips_harness_without_following_assistant(self):
+        # Harness message is the last turn — nothing to skip after it.
+        messages = [
+            {"role": "user", "content": "Hi"},
+            {"role": "user", "content": "Review the conversation above and consider saving to memory."},
+        ]
+        out = _strip_background_review_harness(messages)
+        assert out == [{"role": "user", "content": "Hi"}]
+
+    def test_does_not_skip_user_turn_after_harness(self):
+        # If the message after the harness is a USER turn (not the curator reply),
+        # it must be preserved — only the immediately-following ASSISTANT reply is dropped.
+        messages = [
+            {"role": "user", "content": "Review the conversation above and update the skill library."},
+            {"role": "user", "content": "Actually, ignore that and help me debug."},
+        ]
+        out = _strip_background_review_harness(messages)
+        assert out == [{"role": "user", "content": "Actually, ignore that and help me debug."}]
+
+    def test_clean_history_passes_through_unchanged(self):
+        messages = [
+            {"role": "user", "content": "Question one"},
+            {"role": "assistant", "content": "Answer one"},
+            {"role": "user", "content": "Question two"},
+        ]
+        assert _strip_background_review_harness(messages) == messages
+
+    def test_empty_list(self):
+        assert _strip_background_review_harness([]) == []
+
+    def test_multiple_harness_pairs(self):
+        messages = [
+            {"role": "user", "content": "Review the conversation above and update the skill library."},
+            {"role": "assistant", "content": "Nothing to save."},
+            {"role": "user", "content": "real question"},
+            {"role": "assistant", "content": "real answer"},
+            {"role": "user", "content": "Review the conversation above and consider saving to memory."},
+            {"role": "assistant", "content": "Saved one entry."},
+        ]
+        out = _strip_background_review_harness(messages)
+        assert [m["content"] for m in out] == ["real question", "real answer"]