Merge branch 'main' of github.com:NousResearch/hermes-agent into bb/gui

# Conflicts:
#	tui_gateway/server.py
This commit is contained in:
Brooklyn Nicholson
2026-05-30 13:19:27 -05:00
157 changed files with 10059 additions and 831 deletions
+43 -5
View File
@@ -474,6 +474,13 @@ class GatewayConfig:
# Delivery settings
always_log_local: bool = True # Always save cron outputs to local files
# Drop outbound "silence narration" messages (e.g. *(silent)*, 🔇, a bare
# ".") pre-send. These are model hallucinations emitted when a persona has
# nothing actionable to say; in bot-to-bot channels they mirror back and
# forth, burning tokens and crashing models. Substrate-level guard that
# survives SOUL.md/prompt drift across providers. Opt out with False for
# raw passthrough.
filter_silence_narration: bool = True
# STT settings
stt_enabled: bool = True # Whether to auto-transcribe inbound voice messages
@@ -582,6 +589,7 @@ class GatewayConfig:
"quick_commands": self.quick_commands,
"sessions_dir": str(self.sessions_dir),
"always_log_local": self.always_log_local,
"filter_silence_narration": self.filter_silence_narration,
"stt_enabled": self.stt_enabled,
"group_sessions_per_user": self.group_sessions_per_user,
"thread_sessions_per_user": self.thread_sessions_per_user,
@@ -650,6 +658,9 @@ class GatewayConfig:
quick_commands=quick_commands,
sessions_dir=sessions_dir,
always_log_local=_coerce_bool(data.get("always_log_local"), True),
filter_silence_narration=_coerce_bool(
data.get("filter_silence_narration"), True
),
stt_enabled=_coerce_bool(stt_enabled, True),
group_sessions_per_user=_coerce_bool(group_sessions_per_user, True),
thread_sessions_per_user=_coerce_bool(thread_sessions_per_user, False),
@@ -757,21 +768,32 @@ def load_gateway_config() -> GatewayConfig:
if "always_log_local" in yaml_cfg:
gw_data["always_log_local"] = yaml_cfg["always_log_local"]
if "filter_silence_narration" in yaml_cfg:
gw_data["filter_silence_narration"] = yaml_cfg[
"filter_silence_narration"
]
if "unauthorized_dm_behavior" in yaml_cfg:
gw_data["unauthorized_dm_behavior"] = _normalize_unauthorized_dm_behavior(
yaml_cfg.get("unauthorized_dm_behavior"),
"pair",
)
# Merge platforms section from config.yaml into gw_data so that
# nested keys like platforms.webhook.extra.routes are loaded.
yaml_platforms = yaml_cfg.get("platforms")
# Merge platform config into gw_data so runtime-only settings under
# ``gateway.platforms`` are loaded the same way as top-level
# ``platforms``. Merge nested first so top-level config keeps
# precedence, matching the existing gateway.streaming fallback.
gateway_cfg = yaml_cfg.get("gateway")
gateway_platforms = gateway_cfg.get("platforms") if isinstance(gateway_cfg, dict) else None
platforms_data = gw_data.setdefault("platforms", {})
if not isinstance(platforms_data, dict):
platforms_data = {}
gw_data["platforms"] = platforms_data
if isinstance(yaml_platforms, dict):
for plat_name, plat_block in yaml_platforms.items():
def _merge_platform_map(source_platforms: Any) -> None:
if not isinstance(source_platforms, dict):
return
for plat_name, plat_block in source_platforms.items():
if not isinstance(plat_block, dict):
continue
existing = platforms_data.get(plat_name, {})
@@ -785,6 +807,10 @@ def load_gateway_config() -> GatewayConfig:
if merged_extra:
merged["extra"] = merged_extra
platforms_data[plat_name] = merged
_merge_platform_map(gateway_platforms)
_merge_platform_map(yaml_cfg.get("platforms"))
if platforms_data:
gw_data["platforms"] = platforms_data
# Iterate built-in platforms plus any registered plugin platforms
# so plugin authors get the same shared-key bridging (#24836).
@@ -890,6 +916,18 @@ def load_gateway_config() -> GatewayConfig:
if entry.apply_yaml_config_fn is None:
continue
platform_cfg = yaml_cfg.get(entry.name)
# Fall back to the platform's block under ``platforms`` /
# ``gateway.platforms`` so adapter hooks still run when the
# user configured the platform only under those nested paths
# (e.g. ``platforms.discord.extra.allow_from``) and not via a
# top-level ``discord:`` block.
if not isinstance(platform_cfg, dict):
for _src in (gateway_platforms, yaml_cfg.get("platforms")):
if isinstance(_src, dict):
_candidate = _src.get(entry.name)
if isinstance(_candidate, dict):
platform_cfg = _candidate
break
if not isinstance(platform_cfg, dict):
continue
try:
+61
View File
@@ -9,6 +9,8 @@ Routes messages to the appropriate destination based on:
"""
import logging
import os
import re
from pathlib import Path
from datetime import datetime
from dataclasses import dataclass
@@ -21,6 +23,32 @@ logger = logging.getLogger(__name__)
MAX_PLATFORM_OUTPUT = 4000
TRUNCATED_VISIBLE = 3800
# Matches strings that are *only* a "silence" narration with optional markdown
# wrappers. Covers: *(silent)*, _silent_, `silent`, ~silent~, (silent), silent,
# 🔇, a bare ".", "…", and the whitespace/marker-padded variants seen in the
# wild. Anchored to start/end so substantive messages that merely *contain* the
# word "silent" are never matched.
_SILENCE_NARRATION = re.compile(
r'^[\s*_~`]*\(?\s*(silent|silence|no\s+response|no\s+reply)\s*\.?\)?[\s*_~`]*$'
r'|^[\s*_~`]*[\U0001F507\.\u2026]+[\s*_~`]*$',
re.IGNORECASE,
)
def _is_silence_narration(content: Optional[str]) -> bool:
"""Return True when ``content`` is *only* a silence-narration token.
Length-guarded (real messages are longer) and anchored to the whole string
so legitimate prose like "The deployment ran silently" or "Silence is
golden — here is the plan..." is never flagged.
"""
if not content:
return False
stripped = content.strip()
if not stripped or len(stripped) > 64: # length guard
return False
return bool(_SILENCE_NARRATION.match(stripped))
from .config import Platform, GatewayConfig
from .session import SessionSource
@@ -261,6 +289,18 @@ class DeliveryRouter:
path.write_text(content)
return path
def _filter_silence_narration_enabled(self) -> bool:
"""Whether the outbound silence-narration filter is active.
``HERMES_FILTER_SILENCE_NARRATION`` env var overrides config when set;
otherwise the ``gateway.filter_silence_narration`` config flag wins
(default True).
"""
env = os.getenv("HERMES_FILTER_SILENCE_NARRATION")
if env is not None:
return env.strip().lower() in ("1", "true", "yes", "on")
return bool(getattr(self.config, "filter_silence_narration", True))
async def _deliver_to_platform(
self,
target: DeliveryTarget,
@@ -286,6 +326,27 @@ class DeliveryRouter:
+ f"\n\n... [truncated, full output saved to {saved_path}]"
)
# Substrate-level anti-loop guard: drop hallucinated "silence narration"
# (*(silent)*, 🔇, a bare ".", etc.) before it ever reaches the adapter.
# In bot-to-bot channels these tokens mirror back and forth until a
# model crashes with "no content after all retries". Behavioral prompt
# rules drift across providers; this single chokepoint covers every
# platform adapter regardless of which persona's prompt failed.
# Local/file delivery (_deliver_local) is a separate path and is never
# filtered — saved silence has no loop risk.
if self._filter_silence_narration_enabled() and _is_silence_narration(content):
logger.warning(
"Dropped silence-narration outbound to %s (chat=%s): %r",
target.platform.value,
target.chat_id,
content[:40],
)
return {
"success": True,
"filtered": "silence_narration",
"delivered": False,
}
send_metadata = dict(metadata or {})
is_named_telegram_private_topic = False
named_telegram_private_topic_name: Optional[str] = None
+6 -3
View File
@@ -1191,10 +1191,12 @@ _MEDIA_EXT_ALTERNATION = "|".join(
# bare-path detector (extract_local_files) downstream rather than silently
# deleted. Shared by the non-streaming dispatch path and the streaming
# consumer so both behave identically.
# Path anchors: ``~/`` (Unix home-relative), ``/`` (Unix absolute),
# ``X:\\`` or ``X:/`` (Windows drive-letter absolute — #34632).
MEDIA_TAG_CLEANUP_RE = re.compile(
r'''[`"']?MEDIA:\s*'''
r'''(?P<path>`[^`\n]+`|"[^"\n]+"|'[^'\n]+'|'''
r'''(?:~/|/)\S+(?:[^\S\n]+\S+)*?\.(?:''' + _MEDIA_EXT_ALTERNATION + r'''))'''
r'''(?:~/|/|[A-Za-z]:[/\\])\S+(?:[^\S\n]+\S+)*?\.(?:''' + _MEDIA_EXT_ALTERNATION + r'''))'''
r'''(?=[\s`"',;:)\]}]|$)[`"']?''',
re.IGNORECASE,
)
@@ -2665,9 +2667,10 @@ class BasePlatformAdapter(ABC):
# (?<![/:\w.]) prevents matching inside URLs (e.g. https://…/img.png)
# and relative paths (./foo.png)
# (?:~/|/) anchors to absolute or home-relative paths
# (?:~/|/) anchors to absolute or home-relative Unix paths
# (?:[A-Za-z]:[/\\]) anchors to Windows drive-letter paths (#34632)
path_re = re.compile(
r'(?<![/:\w.])(?:~/|/)(?:[\w.\-]+/)*[\w.\-]+\.(?:' + ext_part + r')\b',
r'(?<![/:\w.])(?:~/|/|[A-Za-z]:[/\\])(?:[\w.\-]+[/\\])*[\w.\-]+\.(?:' + ext_part + r')\b',
re.IGNORECASE,
)
+90 -21
View File
@@ -2804,21 +2804,8 @@ class TelegramAdapter(BasePlatformAdapter):
return slug
try:
# Build provider buttons — 2 per row
buttons: list = []
for p in providers:
count = p.get("total_models", len(p.get("models", [])))
label = f"{p['name']} ({count})"
if p.get("is_current"):
label = f"{label}"
# Compact callback data: mp:<slug> (max 64 bytes)
buttons.append(
InlineKeyboardButton(label, callback_data=f"mp:{p['slug']}")
)
rows = [buttons[i : i + 2] for i in range(0, len(buttons), 2)]
rows.append([InlineKeyboardButton("✗ Cancel", callback_data="mx")])
keyboard = InlineKeyboardMarkup(rows)
# Build provider buttons — folds provider groups (display only).
keyboard = self._build_provider_keyboard(providers)
provider_label = get_label(current_provider)
text = self.format_message(
@@ -2865,6 +2852,56 @@ class TelegramAdapter(BasePlatformAdapter):
_MODEL_PAGE_SIZE = 8
def _build_provider_keyboard(self, providers: list):
"""Build the top-level provider keyboard, folding provider groups.
Provider families (Kimi/Moonshot, MiniMax, xAI Grok, ...) collapse to
a single ``mpg:<gid>`` button; tapping it drills into a member
sub-keyboard. Single providers (and groups with only one authenticated
member) render as direct ``mp:<slug>`` buttons. Grouping mirrors the
CLI ``hermes model`` picker via the shared ``group_providers`` fold,
so all surfaces stay consistent.
"""
try:
from hermes_cli.models import group_providers
except Exception:
group_providers = None
by_slug = {p.get("slug"): p for p in providers}
def _provider_button(p):
count = p.get("total_models", len(p.get("models", [])))
label = f"{p['name']} ({count})"
if p.get("is_current"):
label = f"{label}"
return InlineKeyboardButton(label, callback_data=f"mp:{p['slug']}")
buttons: list = []
if group_providers is not None:
for row in group_providers([p.get("slug") for p in providers]):
if row["kind"] == "group":
members = [by_slug[m] for m in row["members"] if m in by_slug]
count = sum(
m.get("total_models", len(m.get("models", []))) for m in members
)
label = f"{row['label']} ▸ ({count})"
if any(m.get("is_current") for m in members):
label = f"{label}"
buttons.append(
InlineKeyboardButton(label, callback_data=f"mpg:{row['group_id']}")
)
else:
p = by_slug.get(row["slug"])
if p is not None:
buttons.append(_provider_button(p))
else:
for p in providers:
buttons.append(_provider_button(p))
rows = [buttons[i : i + 2] for i in range(0, len(buttons), 2)]
rows.append([InlineKeyboardButton("✗ Cancel", callback_data="mx")])
return InlineKeyboardMarkup(rows)
def _build_model_keyboard(self, models: list, page: int) -> tuple:
"""Build paginated model buttons. Returns (keyboard, page_info_text)."""
page_size = self._MODEL_PAGE_SIZE
@@ -3043,10 +3080,23 @@ class TelegramAdapter(BasePlatformAdapter):
# Clean up state
self._model_picker_state.pop(chat_id, None)
elif data == "mb":
# --- Back to provider list ---
elif data.startswith("mpg:"):
# --- Provider group selected: show member providers ---
group_id = data[4:]
try:
from hermes_cli.models import PROVIDER_GROUPS
_label, member_slugs = PROVIDER_GROUPS.get(group_id, ("", []))
except Exception:
_label, member_slugs = "", []
by_slug = {p["slug"]: p for p in state["providers"]}
members = [by_slug[m] for m in member_slugs if m in by_slug]
if not members:
await query.answer(text="Group not found.")
return
buttons = []
for p in state["providers"]:
for p in members:
count = p.get("total_models", len(p.get("models", [])))
label = f"{p['name']} ({count})"
if p.get("is_current"):
@@ -3054,11 +3104,30 @@ class TelegramAdapter(BasePlatformAdapter):
buttons.append(
InlineKeyboardButton(label, callback_data=f"mp:{p['slug']}")
)
rows = [buttons[i : i + 2] for i in range(0, len(buttons), 2)]
rows.append([InlineKeyboardButton("✗ Cancel", callback_data="mx")])
rows.append([
InlineKeyboardButton("◀ Back", callback_data="mb"),
InlineKeyboardButton("✗ Cancel", callback_data="mx"),
])
keyboard = InlineKeyboardMarkup(rows)
await query.edit_message_text(
text=self.format_message(
(
f"⚙ *Model Configuration*\n\n"
f"Provider family: *{_label or group_id}*\n\n"
f"Select a provider:"
)
),
parse_mode=ParseMode.MARKDOWN_V2,
reply_markup=keyboard,
)
await query.answer()
elif data == "mb":
# --- Back to provider list (folds groups) ---
keyboard = self._build_provider_keyboard(state["providers"])
try:
provider_label = get_label(state["current_provider"])
except Exception:
@@ -3107,7 +3176,7 @@ class TelegramAdapter(BasePlatformAdapter):
query_user_name = getattr(query.from_user, "first_name", None)
# --- Model picker callbacks ---
if data.startswith(("mp:", "mm:", "mb", "mx", "mg:")):
if data.startswith(("mp:", "mpg:", "mm:", "mb", "mx", "mg:")):
chat_id = str(query.message.chat_id) if query.message else None
if chat_id:
await self._handle_model_picker_callback(query, data, chat_id)
+115 -6
View File
@@ -1180,12 +1180,48 @@ class WeixinAdapter(BasePlatformAdapter):
default=False,
)
# Text debounce batching (mirrors Telegram adapter pattern).
# iLink delivers messages individually, so rapid multi-message
# bursts (forwarded batches, paste-splits) each trigger a
# separate agent invocation. Default 3s delay / 5s split delay
# are tuned for iLink's typical delivery cadence. Tunable via
# config.yaml under
# ``gateway.platforms.weixin.extra.text_batch_delay_seconds`` /
# ``text_batch_split_delay_seconds``.
self._text_batch_delay_seconds = self._coerce_float_extra(
"text_batch_delay_seconds", 3.0
)
self._text_batch_split_delay_seconds = self._coerce_float_extra(
"text_batch_split_delay_seconds", 5.0
)
self._pending_text_batches: Dict[str, MessageEvent] = {}
self._pending_text_batch_tasks: Dict[str, asyncio.Task] = {}
if self._account_id and not self._token:
persisted = load_weixin_account(hermes_home, self._account_id)
if persisted:
self._token = str(persisted.get("token") or "").strip()
self._base_url = str(persisted.get("base_url") or self._base_url).strip().rstrip("/")
def _coerce_float_extra(self, key: str, default: float) -> float:
"""Read a float from ``config.extra``, guarding against bad/non-finite values.
The result is fed directly to ``asyncio.sleep()``, so NaN/Inf and
unparseable values fall back to ``default``.
"""
import math
value = self.config.extra.get(key) if getattr(self.config, "extra", None) else None
if value is None:
return float(default)
try:
parsed = float(value)
except (TypeError, ValueError):
return float(default)
if not math.isfinite(parsed) or parsed < 0:
return float(default)
return parsed
@staticmethod
def _coerce_list(value: Any) -> List[str]:
if value is None:
@@ -1247,6 +1283,11 @@ class WeixinAdapter(BasePlatformAdapter):
async def disconnect(self) -> None:
_LIVE_ADAPTERS.pop(self._token, None)
self._running = False
for task in self._pending_text_batch_tasks.values():
if not task.done():
task.cancel()
self._pending_text_batches.clear()
self._pending_text_batch_tasks.clear()
if self._poll_task and not self._poll_task.done():
self._poll_task.cancel()
try:
@@ -1395,12 +1436,10 @@ class WeixinAdapter(BasePlatformAdapter):
timestamp=datetime.now(),
)
logger.info("[%s] inbound from=%s type=%s media=%d", self.name, _safe_id(sender_id), source.chat_type, len(media_paths))
await self.handle_message(event)
@property
def enforces_own_access_policy(self) -> bool:
"""Weixin gates DM/group access at intake via dm_policy/group_policy."""
return True
if event.message_type == MessageType.TEXT:
self._enqueue_text_event(event)
else:
await self.handle_message(event)
def _is_dm_allowed(self, sender_id: str) -> bool:
if self._dm_policy == "disabled":
@@ -1409,6 +1448,76 @@ class WeixinAdapter(BasePlatformAdapter):
return sender_id in self._allow_from
return True
@property
def enforces_own_access_policy(self) -> bool:
"""Weixin gates DM/group access at intake via dm_policy/group_policy."""
return True
# ------------------------------------------------------------------
# Text debounce batching
# ------------------------------------------------------------------
_SPLIT_THRESHOLD = 1800 # iLink chunks at ~2048 chars
def _text_batch_key(self, event: MessageEvent) -> str:
"""Session-scoped key for text message batching."""
from gateway.session import build_session_key
return build_session_key(
event.source,
group_sessions_per_user=self.config.extra.get("group_sessions_per_user", True),
thread_sessions_per_user=self.config.extra.get("thread_sessions_per_user", False),
)
def _enqueue_text_event(self, event: MessageEvent) -> None:
"""Buffer a text event and reset the flush timer.
When users forward multiple messages or send rapid-fire texts
via WeChat, each arrives as a separate iLink message. This
concatenates them and waits for a short quiet period before
dispatching the combined message.
"""
key = self._text_batch_key(event)
existing = self._pending_text_batches.get(key)
chunk_len = len(event.text or "")
if existing is None:
event._last_chunk_len = chunk_len # type: ignore[attr-defined]
self._pending_text_batches[key] = event
else:
if event.text:
existing.text = f"{existing.text}\n{event.text}" if existing.text else event.text
existing._last_chunk_len = chunk_len # type: ignore[attr-defined]
if event.media_urls:
existing.media_urls.extend(event.media_urls)
existing.media_types.extend(event.media_types)
prior_task = self._pending_text_batch_tasks.get(key)
if prior_task and not prior_task.done():
prior_task.cancel()
self._pending_text_batch_tasks[key] = asyncio.create_task(
self._flush_text_batch(key)
)
async def _flush_text_batch(self, key: str) -> None:
"""Wait for quiet period then dispatch aggregated text."""
current_task = asyncio.current_task()
try:
pending = self._pending_text_batches.get(key)
last_len = getattr(pending, "_last_chunk_len", 0) if pending else 0
if last_len >= self._SPLIT_THRESHOLD:
delay = self._text_batch_split_delay_seconds
else:
delay = self._text_batch_delay_seconds
await asyncio.sleep(delay)
if self._pending_text_batch_tasks.get(key) is not current_task:
return
event = self._pending_text_batches.pop(key, None)
if not event:
return
await self.handle_message(event)
finally:
if self._pending_text_batch_tasks.get(key) is current_task:
self._pending_text_batch_tasks.pop(key, None)
async def _collect_media(self, item: Dict[str, Any], media_paths: List[str], media_types: List[str]) -> None:
item_type = item.get("type")
if item_type == ITEM_IMAGE:
+102 -2
View File
@@ -278,6 +278,43 @@ class WhatsAppAdapter(BasePlatformAdapter):
# notification before the normal "✓ whatsapp disconnected" fires.
self._shutting_down: bool = False
# Text debounce batching (mirrors Telegram adapter pattern).
# WhatsApp often delivers multiple messages in rapid succession
# (e.g. forwarded batches, paste-splits) — without debounce each
# message triggers a separate agent invocation, wasting tokens and
# flooding the user with reply fragments. Default 5s delay /
# 10s split delay are conservative for WhatsApp's delivery cadence.
# Tunable via config.yaml under
# ``gateway.platforms.whatsapp.extra.text_batch_delay_seconds`` /
# ``text_batch_split_delay_seconds``.
self._text_batch_delay_seconds = self._coerce_float_extra(
"text_batch_delay_seconds", 5.0
)
self._text_batch_split_delay_seconds = self._coerce_float_extra(
"text_batch_split_delay_seconds", 10.0
)
self._pending_text_batches: Dict[str, MessageEvent] = {}
self._pending_text_batch_tasks: Dict[str, asyncio.Task] = {}
def _coerce_float_extra(self, key: str, default: float) -> float:
"""Read a float from ``config.extra``, guarding against bad/non-finite values.
The result is fed directly to ``asyncio.sleep()``, so NaN/Inf and
unparseable values fall back to ``default``.
"""
import math
value = self.config.extra.get(key) if getattr(self.config, "extra", None) else None
if value is None:
return float(default)
try:
parsed = float(value)
except (TypeError, ValueError):
return float(default)
if not math.isfinite(parsed) or parsed < 0:
return float(default)
return parsed
def _effective_reply_prefix(self) -> str:
"""Return the prefix the Node bridge will add in self-chat mode."""
whatsapp_mode = os.getenv("WHATSAPP_MODE", "self-chat")
@@ -1139,7 +1176,10 @@ class WhatsAppAdapter(BasePlatformAdapter):
for msg_data in messages:
event = await self._build_message_event(msg_data)
if event:
await self.handle_message(event)
if event.message_type == MessageType.TEXT:
self._enqueue_text_event(event)
else:
await self.handle_message(event)
except asyncio.CancelledError:
break
except Exception as e:
@@ -1151,7 +1191,67 @@ class WhatsAppAdapter(BasePlatformAdapter):
await asyncio.sleep(5)
await asyncio.sleep(1) # Poll interval
# ── Text debounce batching ──────────────────────────────────────
_SPLIT_THRESHOLD = 6000 # WhatsApp supports ~65K chars; generous threshold
def _text_batch_key(self, event: MessageEvent) -> str:
"""Session-scoped key for text message batching."""
from gateway.session import build_session_key
return build_session_key(
event.source,
group_sessions_per_user=self.config.extra.get("group_sessions_per_user", True),
thread_sessions_per_user=self.config.extra.get("thread_sessions_per_user", False),
)
def _enqueue_text_event(self, event: MessageEvent) -> None:
"""Buffer a text event and reset the flush timer.
When WhatsApp delivers rapid-fire messages (e.g. forwarded
batches), this concatenates them and waits for a short quiet
period before dispatching the combined message.
"""
key = self._text_batch_key(event)
existing = self._pending_text_batches.get(key)
chunk_len = len(event.text or "")
if existing is None:
event._last_chunk_len = chunk_len # type: ignore[attr-defined]
self._pending_text_batches[key] = event
else:
if event.text:
existing.text = f"{existing.text}\n{event.text}" if existing.text else event.text
existing._last_chunk_len = chunk_len # type: ignore[attr-defined]
if event.media_urls:
existing.media_urls.extend(event.media_urls)
existing.media_types.extend(event.media_types)
prior_task = self._pending_text_batch_tasks.get(key)
if prior_task and not prior_task.done():
prior_task.cancel()
self._pending_text_batch_tasks[key] = asyncio.create_task(
self._flush_text_batch(key)
)
async def _flush_text_batch(self, key: str) -> None:
"""Wait for quiet period then dispatch aggregated text."""
current_task = asyncio.current_task()
try:
pending = self._pending_text_batches.get(key)
last_len = getattr(pending, "_last_chunk_len", 0) if pending else 0
if last_len >= self._SPLIT_THRESHOLD:
delay = self._text_batch_split_delay_seconds
else:
delay = self._text_batch_delay_seconds
await asyncio.sleep(delay)
event = self._pending_text_batches.pop(key, None)
if not event:
return
await self.handle_message(event)
finally:
if self._pending_text_batch_tasks.get(key) is current_task:
self._pending_text_batch_tasks.pop(key, None)
async def _build_message_event(self, data: Dict[str, Any]) -> Optional[MessageEvent]:
"""Build a MessageEvent from bridge message data, downloading images to cache."""
try:
+142 -43
View File
@@ -1730,6 +1730,14 @@ class GatewayRunner:
self._running_agents: Dict[str, Any] = {}
self._running_agents_ts: Dict[str, float] = {} # start timestamp per session
self._pending_messages: Dict[str, str] = {} # Queued messages during interrupt
# Last successfully-resolved (non-empty) model, keyed by session. Used
# as a fallback when a fresh config read transiently returns an empty
# model (e.g. an mtime-keyed config-cache miss during a post-interrupt
# recovery turn). Without this, the agent is built with model="" and
# every API call fails HTTP 400 "No models provided" — the session goes
# silent until the user manually re-sends. See #35314. ``"*"`` holds a
# process-wide last-known-good for sessions seen for the first time.
self._last_resolved_model: Dict[str, str] = {}
# Overflow buffer for explicit /queue commands. The adapter-level
# _pending_messages dict is a single slot per session (designed for
# "next-turn" follow-ups where repeated sends collapse into one
@@ -2488,6 +2496,32 @@ class GatewayRunner:
except Exception:
pass
# Final safety net (#35314): if resolution still produced an empty
# model — e.g. a transient config-cache miss during a post-interrupt
# recovery turn returned an empty user_config — reuse the last model we
# successfully resolved for this session (or, failing that, the most
# recent one resolved process-wide). Building an agent with model=""
# makes every API call fail HTTP 400 "No models provided" and the
# session goes silent until the user manually re-sends. ``getattr``
# guards against bare test runners built via ``object.__new__``.
_last_good = getattr(self, "_last_resolved_model", None)
if _last_good is not None:
if not model:
_recovered = _last_good.get(resolved_session_key or "") or _last_good.get("*")
if _recovered:
logger.warning(
"Empty model resolved for session=%s — recovering "
"last-known-good model %s (config read likely returned "
"empty; see #35314)",
resolved_session_key or "", _recovered,
)
model = _recovered
elif model:
# Cache the good resolution for future recovery turns.
if resolved_session_key:
_last_good[resolved_session_key] = model
_last_good["*"] = model
return model, runtime_kwargs
def _resolve_turn_agent_config(self, user_message: str, model: str, runtime_kwargs: dict) -> dict:
@@ -2784,10 +2818,12 @@ class GatewayRunner:
"""Mark a queued platform as paused — keep it in ``_failed_platforms``
but stop the reconnect watcher from hammering it.
Used by the circuit breaker after ``_PAUSE_AFTER_FAILURES`` consecutive
retryable failures, and by ``/platform pause <name>`` for manual
intervention. Paused platforms are surfaced in ``/platform list``
and resumed with ``/platform resume <name>``.
Used by ``/platform pause <name>`` for manual operator intervention.
Paused platforms are surfaced in ``/platform list`` and resumed with
``/platform resume <name>``. Note: the reconnect watcher does NOT
auto-pause retryable (network/DNS) failures keep retrying at the
backoff cap indefinitely so a transient outage self-heals without
manual intervention.
"""
info = getattr(self, "_failed_platforms", {}).get(platform)
if info is None:
@@ -5865,15 +5901,17 @@ class GatewayRunner:
"""Background task that periodically retries connecting failed platforms.
Uses exponential backoff: 30s 60s 120s 240s 300s (cap).
Retryable failures keep retrying at the backoff cap indefinitely
but if a platform fails ``_PAUSE_AFTER_FAILURES`` times in a row
without ever succeeding, it is *paused*: kept in the retry queue
but no longer hammered. The user surfaces it with ``/platform list``
and resumes it with ``/platform resume <name>``. Non-retryable
failures (bad auth, etc.) still drop out of the queue immediately.
Retryable failures (network/DNS blips) keep retrying at the backoff
cap indefinitely they self-heal once connectivity returns, so a
transient outage never requires manual intervention. Non-retryable
failures (bad auth, etc.) drop out of the queue immediately. The
circuit breaker (``_pause_failed_platform`` / ``/platform pause``)
remains available for manual operator control via ``/platform list``
and ``/platform resume <name>``, but is no longer triggered
automatically auto-pausing a recovered platform was the cause of
bots silently staying dead after a transient DNS failure.
"""
_BACKOFF_CAP = 300 # 5 minutes max between retries
_PAUSE_AFTER_FAILURES = 10 # circuit-breaker threshold
await asyncio.sleep(10) # initial delay — let startup finish
while self._running:
@@ -5968,14 +6006,14 @@ class GatewayRunner:
"Reconnect %s failed, next retry in %ds",
platform.value, backoff,
)
if attempt >= _PAUSE_AFTER_FAILURES:
self._pause_failed_platform(
platform,
reason=(
adapter.fatal_error_message
or "failed to reconnect"
),
)
# Retryable failures (network/DNS blips) keep retrying
# at the backoff cap indefinitely — they self-heal once
# connectivity returns. We do NOT auto-pause them: a
# transient outage must never require manual `/platform
# resume` to recover. Non-retryable failures (bad auth,
# etc.) already drop out of the queue via the
# `not fatal_error_retryable` branch above, so anything
# reaching here is by definition retryable.
except Exception as e:
self._update_platform_runtime_status(
platform.value,
@@ -5990,8 +6028,9 @@ class GatewayRunner:
"Reconnect %s error: %s, next retry in %ds",
platform.value, e, backoff,
)
if attempt >= _PAUSE_AFTER_FAILURES:
self._pause_failed_platform(platform, reason=str(e))
# A raised exception during reconnect (connect timeout, DNS
# resolution failure, etc.) is inherently transient — keep
# retrying at the backoff cap rather than auto-pausing.
# Check every 10 seconds for platforms that need reconnection
for _ in range(10):
@@ -10531,6 +10570,22 @@ class GatewayRunner:
except Exception as exc:
logger.warning("Picker model switch failed for cached agent: %s", exc)
# Persist the new model to the session DB so the
# dashboard shows the updated model (#34850).
_sess_db = getattr(_self, "_session_db", None)
if _sess_db is not None:
try:
_sess_entry = _self.session_store.get_or_create_session(
event.source
)
_sess_db.update_session_model(
_sess_entry.session_id, result.new_model
)
except Exception as exc:
logger.debug(
"Failed to persist model switch to DB: %s", exc
)
# Store model note + session override
if not hasattr(_self, "_pending_model_notes"):
_self._pending_model_notes = {}
@@ -10668,6 +10723,20 @@ class GatewayRunner:
except Exception as exc:
logger.warning("In-place model switch failed for cached agent: %s", exc)
# Persist the new model to the session DB so the dashboard
# shows the updated model (#34850).
_sess_db = getattr(self, "_session_db", None)
if _sess_db is not None:
try:
_sess_entry = self.session_store.get_or_create_session(source)
_sess_db.update_session_model(
_sess_entry.session_id, result.new_model
)
except Exception as exc:
logger.debug(
"Failed to persist model switch to DB: %s", exc
)
# Store a note to prepend to the next user message so the model
# knows about the switch (avoids system messages mid-history).
if not hasattr(self, "_pending_model_notes"):
@@ -15313,8 +15382,52 @@ class GatewayRunner:
("compression", "target_ratio"),
("compression", "protect_last_n"),
("agent", "disabled_toolsets"),
("memory", "provider"),
)
_HONCHO_CACHE_BUSTING_KEYS = (
"honcho.peer_name",
"honcho.ai_peer",
"honcho.pin_peer_name",
"honcho.runtime_peer_prefix",
"honcho.user_peer_aliases",
)
_HONCHO_CACHE_BUSTING_MEMO: dict[tuple[str, int | None], dict[str, Any]] = {}
@classmethod
def _empty_honcho_cache_busting_config(cls) -> dict[str, Any]:
return {key: None for key in cls._HONCHO_CACHE_BUSTING_KEYS}
@classmethod
def _extract_honcho_cache_busting_config(cls) -> dict[str, Any]:
"""Extract Honcho identity keys, memoized by honcho.json mtime."""
try:
from plugins.memory.honcho.client import HonchoClientConfig, resolve_config_path
path = resolve_config_path()
try:
mtime_ns = path.stat().st_mtime_ns
except OSError:
mtime_ns = None
memo_key = (str(path), mtime_ns)
cached = cls._HONCHO_CACHE_BUSTING_MEMO.get(memo_key)
if cached is not None:
return dict(cached)
hcfg = HonchoClientConfig.from_global_config(config_path=path)
aliases = hcfg.user_peer_aliases or {}
values = {
"honcho.peer_name": hcfg.peer_name,
"honcho.ai_peer": hcfg.ai_peer,
"honcho.pin_peer_name": bool(hcfg.pin_peer_name),
"honcho.runtime_peer_prefix": hcfg.runtime_peer_prefix or "",
"honcho.user_peer_aliases": sorted(aliases.items()) if isinstance(aliases, dict) else [],
}
cls._HONCHO_CACHE_BUSTING_MEMO = {memo_key: values}
return dict(values)
except Exception:
return cls._empty_honcho_cache_busting_config()
@classmethod
def _extract_cache_busting_config(cls, user_config: dict | None) -> dict:
"""Pull values that must bust the cached agent.
@@ -15345,26 +15458,12 @@ class GatewayRunner:
out["tools.registry_generation"] = None
# Honcho identity-mapping keys live in honcho.json, not user_config.
# HonchoSessionManager freezes the resolved peer_name / ai_peer /
# pin / aliases / prefix at construction; without busting here,
# mid-flight honcho.json edits go unread until the next unrelated
# cache eviction.
try:
from plugins.memory.honcho.client import HonchoClientConfig
hcfg = HonchoClientConfig.from_global_config()
out["honcho.peer_name"] = hcfg.peer_name
out["honcho.ai_peer"] = hcfg.ai_peer
out["honcho.pin_peer_name"] = bool(hcfg.pin_peer_name)
out["honcho.runtime_peer_prefix"] = hcfg.runtime_peer_prefix or ""
aliases = hcfg.user_peer_aliases or {}
out["honcho.user_peer_aliases"] = sorted(aliases.items()) if isinstance(aliases, dict) else []
except Exception:
out["honcho.peer_name"] = None
out["honcho.ai_peer"] = None
out["honcho.pin_peer_name"] = None
out["honcho.runtime_peer_prefix"] = None
out["honcho.user_peer_aliases"] = None
# Only read that file when Honcho is the active memory provider.
provider = cfg_get(cfg, "memory", "provider")
if isinstance(provider, str) and provider.lower() == "honcho":
out.update(cls._extract_honcho_cache_busting_config())
else:
out.update(cls._empty_honcho_cache_busting_config())
return out
@@ -17203,7 +17302,7 @@ class GatewayRunner:
_hc = _hm.get("content", "")
if "MEDIA:" in _hc:
_TOOL_MEDIA_RE = re.compile(
r'MEDIA:((?:/|~\/)\S+\.(?:png|jpe?g|gif|webp|'
r'MEDIA:((?:[A-Za-z]:[/\\]|/|~\/)\S+\.(?:png|jpe?g|gif|webp|'
r'mp4|mov|avi|mkv|webm|ogg|opus|mp3|wav|m4a|'
r'flac|epub|pdf|zip|rar|7z|docx?|xlsx?|pptx?|'
r'txt|csv|apk|ipa))',
@@ -17529,7 +17628,7 @@ class GatewayRunner:
content = msg.get("content", "")
if "MEDIA:" in content:
_TOOL_MEDIA_RE = re.compile(
r'MEDIA:((?:/|~\/)\S+\.(?:png|jpe?g|gif|webp|'
r'MEDIA:((?:[A-Za-z]:[/\\]|/|~\/)\S+\.(?:png|jpe?g|gif|webp|'
r'mp4|mov|avi|mkv|webm|ogg|opus|mp3|wav|m4a|'
r'flac|epub|pdf|zip|rar|7z|docx?|xlsx?|pptx?|'
r'txt|csv|apk|ipa))',