mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-04-25 00:51:20 +00:00
fix(streaming): silent retry when stream dies mid tool-call (#14151)
When the streaming connection dropped AFTER user-visible text was delivered but a tool call was in flight, we stubbed the turn with a '⚠ Stream stalled mid tool-call; Ask me to retry' warning — costing an iteration and breaking the flow. Users report this happening increasingly often on long SSE streams through flaky provider routes. Fix: in the existing inner stream-retry loop, relax the deltas_were_sent short-circuit. If a tool call was in flight (partial_tool_names populated) AND the error is a transient connection error (timeout, RemoteProtocolError, SSE 'connection lost', etc.), silently retry instead of bailing out. Fire a brief 'Connection dropped mid tool-call; reconnecting…' marker so the user understands the preamble is about to be re-streamed. Researched how Claude Code (tombstone + non-streaming fallback), OpenCode (blind Effect.retry wrapping whole stream), and Clawdbot (4-way gate: stopReason==error + output==0 + !hadPotentialSideEffects) handle this. Chose the narrow Clawdbot-style gate: retry only when (a) a tool call was actually in flight (otherwise the existing stub-with-recovered-text is correct for pure-text stalls) and (b) the error is transient. Side-effect safety is automatic — no tool has been dispatched within this single API call yet. UX trade-off: user sees preamble text twice on retry (OpenCode-style). Strictly better than a lost action with a 'retry manually' message. If retries exhaust, falls through to the existing stub-with-warning path so the user isn't left with zero signal. Tests: 3 new tests in TestSilentRetryMidToolCall covering (1) silent retry recovers tool call; (2) exhausted retries fall back to stub; (3) text-only stalls don't trigger retry. 30/30 pass.
This commit is contained in:
Teknium
GitHub
parent
88564ad8bc
commit
ea67e49574
2 changed files with 339 additions and 10 deletions
127
run_agent.py
127
run_agent.py
|
|
@ -5826,16 +5826,6 @@ class AIAgent:
|
|||
result["response"] = _call_chat_completions()
|
||||
return # success
|
||||
except Exception as e:
|
||||
if deltas_were_sent["yes"]:
|
||||
# Streaming failed AFTER some tokens were already
|
||||
# delivered. Don't retry or fall back — partial
|
||||
# content already reached the user.
|
||||
logger.warning(
|
||||
"Streaming failed after partial delivery, not retrying: %s", e
|
||||
)
|
||||
result["error"] = e
|
||||
return
|
||||
|
||||
_is_timeout = isinstance(
|
||||
e, (_httpx.ReadTimeout, _httpx.ConnectTimeout, _httpx.PoolTimeout)
|
||||
)
|
||||
|
|
@ -5843,6 +5833,123 @@ class AIAgent:
|
|||
e, (_httpx.ConnectError, _httpx.RemoteProtocolError, ConnectionError)
|
||||
)
|
||||
|
||||
# If the stream died AFTER some tokens were delivered:
|
||||
# normally we don't retry (the user already saw text,
|
||||
# retrying would duplicate it). BUT: if a tool call
|
||||
# was in-flight when the stream died, silently aborting
|
||||
# discards the tool call entirely. In that case we
|
||||
# prefer to retry — the user sees a brief
|
||||
# "reconnecting" marker + duplicated preamble text,
|
||||
# which is strictly better than a failed action with
|
||||
# a "retry manually" message. Limit this to transient
|
||||
# connection errors (Clawdbot-style narrow gate): no
|
||||
# tool has executed yet within this API call, so
|
||||
# silent retry is safe wrt side-effects.
|
||||
if deltas_were_sent["yes"]:
|
||||
_partial_tool_in_flight = bool(
|
||||
result.get("partial_tool_names")
|
||||
)
|
||||
_is_sse_conn_err_preview = False
|
||||
if not _is_timeout and not _is_conn_err:
|
||||
from openai import APIError as _APIError
|
||||
if isinstance(e, _APIError) and not getattr(e, "status_code", None):
|
||||
_err_lower_preview = str(e).lower()
|
||||
_SSE_PREVIEW_PHRASES = (
|
||||
"connection lost",
|
||||
"connection reset",
|
||||
"connection closed",
|
||||
"connection terminated",
|
||||
"network error",
|
||||
"network connection",
|
||||
"terminated",
|
||||
"peer closed",
|
||||
"broken pipe",
|
||||
"upstream connect error",
|
||||
)
|
||||
_is_sse_conn_err_preview = any(
|
||||
phrase in _err_lower_preview
|
||||
for phrase in _SSE_PREVIEW_PHRASES
|
||||
)
|
||||
_is_transient = (
|
||||
_is_timeout or _is_conn_err or _is_sse_conn_err_preview
|
||||
)
|
||||
_can_silent_retry = (
|
||||
_partial_tool_in_flight
|
||||
and _is_transient
|
||||
and _stream_attempt < _max_stream_retries
|
||||
)
|
||||
if not _can_silent_retry:
|
||||
# Either no tool call was in-flight (so the
|
||||
# turn was a pure text response — current
|
||||
# stub-with-recovered-text behaviour is
|
||||
# correct), or retries are exhausted, or the
|
||||
# error isn't transient. Fall through to the
|
||||
# stub path.
|
||||
logger.warning(
|
||||
"Streaming failed after partial delivery, not retrying: %s", e
|
||||
)
|
||||
result["error"] = e
|
||||
return
|
||||
# Tool call was in-flight AND error is transient:
|
||||
# retry silently. Clear per-attempt state so the
|
||||
# next stream starts clean. Fire a "reconnecting"
|
||||
# marker so the user sees why the preamble is
|
||||
# about to be re-streamed.
|
||||
logger.info(
|
||||
"Streaming attempt %s/%s died mid tool-call "
|
||||
"(%s: %s) after user-visible text; retrying "
|
||||
"silently to avoid losing the action. "
|
||||
"Preamble will re-stream.",
|
||||
_stream_attempt + 1,
|
||||
_max_stream_retries + 1,
|
||||
type(e).__name__,
|
||||
e,
|
||||
)
|
||||
try:
|
||||
self._fire_stream_delta(
|
||||
"\n\n⚠ Connection dropped mid tool-call; "
|
||||
"reconnecting…\n\n"
|
||||
)
|
||||
except Exception:
|
||||
pass
|
||||
# Reset the streamed-text buffer so the retry's
|
||||
# fresh preamble doesn't get double-recorded in
|
||||
# _current_streamed_assistant_text (which would
|
||||
# pollute the interim-visible-text comparison).
|
||||
try:
|
||||
self._reset_stream_delivery_tracking()
|
||||
except Exception:
|
||||
pass
|
||||
# Reset in-memory accumulators so the next
|
||||
# attempt's chunks don't concat onto the dead
|
||||
# stream's partial JSON.
|
||||
result["partial_tool_names"] = []
|
||||
deltas_were_sent["yes"] = False
|
||||
first_delta_fired["done"] = False
|
||||
self._emit_status(
|
||||
f"⚠️ Connection dropped mid tool-call "
|
||||
f"({type(e).__name__}). Reconnecting… "
|
||||
f"(attempt {_stream_attempt + 2}/{_max_stream_retries + 1})"
|
||||
)
|
||||
self._touch_activity(
|
||||
f"stream retry {_stream_attempt + 2}/{_max_stream_retries + 1} "
|
||||
f"mid tool-call after {type(e).__name__}"
|
||||
)
|
||||
stale = request_client_holder.get("client")
|
||||
if stale is not None:
|
||||
self._close_request_openai_client(
|
||||
stale, reason="stream_mid_tool_retry_cleanup"
|
||||
)
|
||||
request_client_holder["client"] = None
|
||||
try:
|
||||
self._replace_primary_openai_client(
|
||||
reason="stream_mid_tool_retry_pool_cleanup"
|
||||
)
|
||||
except Exception:
|
||||
pass
|
||||
self._emit_status("🔄 Reconnected — resuming…")
|
||||
continue
|
||||
|
||||
# SSE error events from proxies (e.g. OpenRouter sends
|
||||
# {"error":{"message":"Network connection lost."}}) are
|
||||
# raised as APIError by the OpenAI SDK. These are
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue