diff --git a/agent/auxiliary_client.py b/agent/auxiliary_client.py
index 9156eaa26fed..2c2ac9138b52 100644
--- a/agent/auxiliary_client.py
+++ b/agent/auxiliary_client.py
@@ -97,17 +97,20 @@ def _normalize_aux_provider(provider: Optional[str]) -> str:
 
 _FIXED_TEMPERATURE_MODELS: Dict[str, float] = {
     "kimi-for-coding": 0.6,
+    # Kimi Coding Plan now rejects kimi-k2.5 unless temperature is exactly 1.0.
+    # Keep this explicit so it overrides the broader non-thinking-family default.
+    "kimi-k2.5": 1.0,
+    "moonshotai/kimi-k2.5": 1.0,
 }
 
-# Moonshot's kimi-for-coding endpoint (api.kimi.com/coding) documents:
-# "k2.5 model will use a fixed value 1.0, non-thinking mode will use a fixed
-# value 0.6.  Any other value will result in an error."  The same lock applies
-# to the other k2.* models served on that endpoint.  Enumerated explicitly so
-# non-coding siblings like `kimi-k2-instruct` (variable temperature, served on
-# the standard chat API and third parties) are NOT clamped.
+# Moonshot's kimi-for-coding endpoint (api.kimi.com/coding) documents fixed
+# temperatures for the K2 family. In current live behavior, `kimi-k2.5`
+# requires exactly 1.0, while the turbo/preview non-thinking siblings still use
+# 0.6. Enumerated explicitly so non-coding siblings like `kimi-k2-instruct`
+# (variable temperature, served on the standard chat API and third parties) are
+# NOT clamped.
 # Source: https://platform.kimi.ai/docs/guide/kimi-k2-5-quickstart
 _KIMI_INSTANT_MODELS: frozenset = frozenset({
-    "kimi-k2.5",
     "kimi-k2-turbo-preview",
     "kimi-k2-0905-preview",
 })
@@ -120,10 +123,11 @@ def _normalize_aux_provider(provider: Optional[str]) -> str:
 def _fixed_temperature_for_model(model: Optional[str]) -> Optional[float]:
     """Return a required temperature override for models with strict contracts.
 
-    Moonshot's kimi-for-coding endpoint rejects any non-approved temperature on
-    the k2.5 family.  Non-thinking variants require exactly 0.6; thinking
-    variants require 1.0.  An optional ``vendor/`` prefix (e.g.
-    ``moonshotai/kimi-k2.5``) is tolerated for aggregator routings.
+    Moonshot's K2 coding-family endpoints reject non-approved temperatures.
+    `kimi-k2.5` currently requires exactly 1.0, turbo/preview non-thinking
+    variants require 0.6, and thinking variants require 1.0. An optional
+    ``vendor/`` prefix (e.g. ``moonshotai/kimi-k2.5``) is tolerated for
+    aggregator routings.
 
     Returns ``None`` for every other model, including ``kimi-k2-instruct*``
     which is the separate non-coding K2 family with variable temperature.
diff --git a/agent/prompt_builder.py b/agent/prompt_builder.py
index 2a210434949b..05628ede7253 100644
--- a/agent/prompt_builder.py
+++ b/agent/prompt_builder.py
@@ -620,6 +620,17 @@ def build_skills_system_prompt(
         or ""
     )
     disabled = get_disabled_skill_names()
+
+    # Load curated skill filter early so it participates in caching
+    _sp_skills: "list[str]" = []
+    try:
+        from hermes_cli.config import load_config
+        _cfg = load_config()
+        _raw = _cfg.get("skills", {}).get("system_prompt_skills")
+        if isinstance(_raw, list):
+            _sp_skills = [str(s).strip() for s in _raw if str(s).strip()]
+    except Exception:
+        pass
     cache_key = (
         str(skills_dir.resolve()),
         tuple(str(d) for d in external_dirs),
@@ -627,6 +638,7 @@ def build_skills_system_prompt(
         tuple(sorted(str(ts) for ts in (available_toolsets or set()))),
         _platform_hint,
         tuple(sorted(disabled)),
+        tuple(sorted(_sp_skills)),
     )
     with _SKILLS_PROMPT_CACHE_LOCK:
         cached = _SKILLS_PROMPT_CACHE.get(cache_key)
@@ -763,6 +775,16 @@ def build_skills_system_prompt(
     if not skills_by_category:
         result = ""
     else:
+        # Filter to curated system-prompt skills if configured
+        if _sp_skills:
+            _allowed = set(_sp_skills)
+            filtered: dict[str, list[tuple[str, str]]] = {}
+            for cat, items in skills_by_category.items():
+                kept = [(n, d) for n, d in items if n in _allowed]
+                if kept:
+                    filtered[cat] = kept
+            skills_by_category = filtered
+
         index_lines = []
         for category in sorted(skills_by_category.keys()):
             cat_desc = category_descriptions.get(category, "")
@@ -781,6 +803,12 @@ def build_skills_system_prompt(
                 else:
                     index_lines.append(f"    - {name}")
 
+        _extra_skills_note = ""
+        if _sp_skills:
+            _extra_skills_note = (
+                "\nAdditional skills are available — call skills_list() to browse the full catalog."
+            )
+
         result = (
             "## Skills (mandatory)\n"
             "Before replying, scan the skills below. If a skill matches or is even partially relevant "
@@ -801,6 +829,7 @@ def build_skills_system_prompt(
             "<available_skills>\n"
             + "\n".join(index_lines) + "\n"
             "</available_skills>\n"
+            + _extra_skills_note + "\n"
             "\n"
             "Only proceed without loading a skill if genuinely none are relevant to the task."
         )
diff --git a/gateway/platforms/whatsapp.py b/gateway/platforms/whatsapp.py
index d1de5b856870..39dcabbc3496 100644
--- a/gateway/platforms/whatsapp.py
+++ b/gateway/platforms/whatsapp.py
@@ -766,6 +766,17 @@ async def send_video(
         """Send a video natively via bridge — plays inline in WhatsApp."""
         return await self._send_media_to_bridge(chat_id, video_path, "video", caption)
 
+    async def send_voice(
+        self,
+        chat_id: str,
+        audio_path: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        **kwargs,
+    ) -> SendResult:
+        """Send an audio file as a voice note via bridge."""
+        return await self._send_media_to_bridge(chat_id, audio_path, "audio", caption)
+
     async def send_document(
         self,
         chat_id: str,
diff --git a/gateway/run.py b/gateway/run.py
index 60c57495b447..2b2e44339269 100644
--- a/gateway/run.py
+++ b/gateway/run.py
@@ -6043,8 +6043,17 @@ def _should_send_voice_reply(
         # When streaming already delivered the text (already_sent=True),
         # the base adapter will receive None and can't run auto-TTS,
         # so the runner must take over.
+        # For platforms without voice-channel auto-TTS (e.g. WhatsApp),
+        # the runner must always handle voice replies.
         if is_voice_input and not already_sent:
-            return False
+            adapter = self.adapters.get(event.source.platform)
+            has_auto_tts = (
+                adapter
+                and hasattr(adapter, "play_in_voice_channel")
+                and hasattr(adapter, "is_in_voice_channel")
+            )
+            if has_auto_tts:
+                return False
 
         return True
 
diff --git a/hermes_cli/config.py b/hermes_cli/config.py
index 9040eac0ba45..68f8bf797423 100644
--- a/hermes_cli/config.py
+++ b/hermes_cli/config.py
@@ -693,6 +693,11 @@ def _ensure_hermes_home_managed(home: Path):
     # always goes to ~/.hermes/skills/.
     "skills": {
         "external_dirs": [],   # e.g. ["~/.agents/skills", "/shared/team-skills"]
+        # Curated list of skill names to show in the system prompt index.
+        # When set, only these skills appear in the auto-injected skills list,
+        # reducing context bloat. Other skills remain available via skills_list.
+        # Example: ["github-pr-workflow", "plan", "systematic-debugging"]
+        "system_prompt_skills": [],
     },
 
     # Honcho AI-native memory -- reads ~/.honcho/config.json as single source of truth.
diff --git a/model_tools.py b/model_tools.py
index 0e8bc877e2cd..68ba98953af1 100644
--- a/model_tools.py
+++ b/model_tools.py
@@ -412,6 +412,31 @@ def _coerce_value(value: str, expected_type):
         return _coerce_number(value, integer_only=(expected_type == "integer"))
     if expected_type == "boolean":
         return _coerce_boolean(value)
+    if expected_type == "array":
+        return _coerce_json(value, list)
+    if expected_type == "object":
+        return _coerce_json(value, dict)
+    return value
+
+
+def _coerce_json(value: str, expected_python_type: type):
+    """Parse *value* as JSON when the schema expects an array or object.
+
+    Handles model output drift where a complex oneOf/discriminated-union schema
+    causes the LLM to emit the array/object as a JSON string instead of a native
+    structure.  Returns the original string if parsing fails or yields the wrong
+    Python type.
+    """
+    try:
+        parsed = json.loads(value)
+    except (ValueError, TypeError):
+        return value
+    if isinstance(parsed, expected_python_type):
+        logger.debug(
+            "coerce_tool_args: coerced string to %s via json.loads",
+            expected_python_type.__name__,
+        )
+        return parsed
     return value
 
 
diff --git a/run_agent.py b/run_agent.py
index 85eaad1b3753..4ddd0e9937e6 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -919,6 +919,13 @@ def __init__(
         self._current_tool: str | None = None
         self._api_call_count: int = 0
 
+        # Tool call loop breaker — detects when the model repeatedly issues
+        # the *exact same* tool call (same name + same args) and forces a
+        # state change after N consecutive duplicates.  This prevents the
+        # agent from getting stuck in infinite tool-call loops.
+        self._tool_call_history: list[tuple[str, str]] = []  # (tool_name, args_hash)
+        self._tool_call_breach_threshold: int = 3  # consecutive identical calls before breaking
+
         # Rate limit tracking — updated from x-ratelimit-* response headers
         # after each API call.  Accessed by /usage slash command.
         self._rate_limit_state: Optional["RateLimitState"] = None
@@ -1312,6 +1319,12 @@ def __init__(
         self._memory_flush_min_turns = 6
         self._turns_since_memory = 0
         self._iters_since_skill = 0
+        # Circuit breaker: detect and break infinite tool-call loops.
+        # Tracks the last (tool_name, args_hash) pair.  When the same
+        # tool+args is called consecutively N times, we abort to force
+        # the LLM into a different strategy.
+        self._consecutive_tool_calls: list[tuple[str, str]] = []
+        self._circuit_breaker_threshold: int = 3
         if not skip_memory:
             try:
                 mem_config = _agent_cfg.get("memory", {})
@@ -7856,6 +7869,35 @@ def _invoke_tool(self, function_name: str, function_args: dict, effective_task_i
         if block_message is not None:
             return json.dumps({"error": block_message}, ensure_ascii=False)
 
+        # ── Circuit breaker: detect infinite tool-call loops ─────────────
+        # Build a stable hash of (tool_name, args) to detect identical
+        # consecutive calls.  When the same tool+args fires N times in a
+        # row, the LLM is likely stuck in a retry loop — abort and force
+        # it to try a different approach.
+        import hashlib as _hashlib
+        args_key = json.dumps(function_args, sort_keys=True, ensure_ascii=False)
+        call_sig = f"{function_name}:{_hashlib.md5(args_key.encode('utf-8')).hexdigest()[:16]}"
+        
+        # Always record the call (append new entry each time)
+        self._consecutive_tool_calls.append((function_name, call_sig))
+        
+        # Check if the last N entries are all the same
+        n = len(self._consecutive_tool_calls)
+        if n >= self._circuit_breaker_threshold:
+            # Check if the last N entries all have the same signature
+            recent_sigs = [s for _, s in self._consecutive_tool_calls[-self._circuit_breaker_threshold:]]
+            if len(set(recent_sigs)) == 1:
+                # All same — circuit breaker triggered!
+                streak = len(self._consecutive_tool_calls)
+                msg = (
+                    f"[熔断器触发] 工具 '{function_name}' 连续 {streak} 次调用完全相同的参数，"
+                    f"已强制终止。请换用不同的工具或策略。"
+                )
+                logger.warning("Circuit breaker triggered: %s (streak=%d)", function_name, streak)
+                # Reset streak so future calls can succeed
+                self._consecutive_tool_calls.clear()
+                return json.dumps({"error": msg}, ensure_ascii=False)
+
         if function_name == "todo":
             from tools.todo_tool import todo_tool as _todo_tool
             return _todo_tool(
@@ -8359,12 +8401,44 @@ def _execute_tool_calls_sequential(self, assistant_message, messages: list, effe
                 except Exception:
                     pass  # never block tool execution
 
+            # ── Circuit breaker: detect infinite tool-call loops ──────────
+            # Same logic as in _invoke_tool() — the sequential path has its
+            # own inline dispatch and must also be protected.
+            _cb_bypass = False
+            if _block_msg is None:
+                import hashlib as _hashlib
+                _cb_args_key = json.dumps(function_args, sort_keys=True, ensure_ascii=False)
+                _cb_call_sig = f"{function_name}:{_hashlib.md5(_cb_args_key.encode('utf-8')).hexdigest()[:16]}"
+                
+                # Always record the call
+                self._consecutive_tool_calls.append((function_name, _cb_call_sig))
+                
+                # Check if the last N entries are all the same
+                _cb_n = len(self._consecutive_tool_calls)
+                if _cb_n >= self._circuit_breaker_threshold:
+                    _cb_recent_sigs = [s for _, s in self._consecutive_tool_calls[-self._circuit_breaker_threshold:]]
+                    if len(set(_cb_recent_sigs)) == 1:
+                        _cb_streak = len(self._consecutive_tool_calls)
+                        function_result = json.dumps({
+                            "error": (
+                                f"[熔断器触发] 工具 '{function_name}' 连续 {_cb_streak} 次调用完全相同的参数，"
+                                f"已强制终止。请换用不同的工具或策略。"
+                            )
+                        }, ensure_ascii=False)
+                        logger.warning("Circuit breaker triggered (sequential): %s (streak=%d)", function_name, _cb_streak)
+                        self._consecutive_tool_calls.clear()
+                        tool_duration = 0.0
+                        _cb_bypass = True
+
             tool_start_time = time.time()
 
-            if _block_msg is not None:
-                # Tool blocked by plugin policy — return error without executing.
-                function_result = json.dumps({"error": _block_msg}, ensure_ascii=False)
-                tool_duration = 0.0
+            if _block_msg is not None or _cb_bypass:
+                # Tool blocked by plugin policy or circuit breaker — return
+                # error without executing.  (_cb_bypass: function_result was
+                # already set in the circuit breaker block above.)
+                if _block_msg is not None:
+                    function_result = json.dumps({"error": _block_msg}, ensure_ascii=False)
+                    tool_duration = 0.0
             elif function_name == "todo":
                 from tools.todo_tool import todo_tool as _todo_tool
                 function_result = _todo_tool(
diff --git a/tests/agent/test_auxiliary_client.py b/tests/agent/test_auxiliary_client.py
index aea8152a53e5..9e1f844a83b1 100644
--- a/tests/agent/test_auxiliary_client.py
+++ b/tests/agent/test_auxiliary_client.py
@@ -699,8 +699,8 @@ def test_500_not_connection(self):
 class TestKimiForCodingTemperature:
     """Moonshot kimi-for-coding models require fixed temperatures.
 
-    k2.5 / k2-turbo-preview / k2-0905-preview → 0.6 (non-thinking lock).
-    k2-thinking / k2-thinking-turbo → 1.0 (thinking lock).
+    kimi-k2.5 / k2-thinking / k2-thinking-turbo → 1.0.
+    k2-turbo-preview / k2-0905-preview → 0.6.
     kimi-k2-instruct* and every other model preserve the caller's temperature.
     """
 
@@ -780,19 +780,20 @@ async def test_auto_routed_kimi_for_coding_async_call_uses_fixed_temperature(sel
     @pytest.mark.parametrize(
         "model,expected",
         [
-            ("kimi-k2.5", 0.6),
+            ("kimi-k2.5", 1.0),
             ("kimi-k2-turbo-preview", 0.6),
             ("kimi-k2-0905-preview", 0.6),
             ("kimi-k2-thinking", 1.0),
             ("kimi-k2-thinking-turbo", 1.0),
-            ("moonshotai/kimi-k2.5", 0.6),
+            ("moonshotai/kimi-k2.5", 1.0),
             ("moonshotai/Kimi-K2-Thinking", 1.0),
         ],
     )
     def test_kimi_k2_family_temperature_override(self, model, expected):
         """Moonshot kimi-k2.* models only accept fixed temperatures.
 
-        Non-thinking models → 0.6, thinking-mode models → 1.0.
+        kimi-k2.5 and thinking-mode models → 1.0.
+        turbo/preview non-thinking variants → 0.6.
         """
         from agent.auxiliary_client import _build_call_kwargs
 
diff --git a/tools/tts_tool.py b/tools/tts_tool.py
index adc6524c46f7..bab216c83a5f 100644
--- a/tools/tts_tool.py
+++ b/tools/tts_tool.py
@@ -167,8 +167,8 @@ def _convert_to_opus(mp3_path: str) -> Optional[str]:
     ogg_path = mp3_path.rsplit(".", 1)[0] + ".ogg"
     try:
         result = subprocess.run(
-            ["ffmpeg", "-i", mp3_path, "-acodec", "libopus",
-             "-ac", "1", "-b:a", "64k", "-vbr", "off", ogg_path, "-y"],
+            ["ffmpeg", "-i", mp3_path, "-c:a", "libopus",
+             "-ar", "48000", "-ac", "1", "-b:a", "24k", "-application", "voip", ogg_path, "-y"],
             capture_output=True, timeout=30,
         )
         if result.returncode != 0: