hermes-admin/hermes-agent
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

fix(compression): match the Codex stall marker on the raw error text

Browse Source 
The stall check relied on the lowercased error string, which only works
while CODEX_STREAM_STALL_MARKER happens to be all-lowercase. Match against
str(e) so the shared marker stays authoritative regardless of case, and
fold the two duplicate #124077 comments into one explaining the split
(stall -> retry-ladder timeout; transport timeouts stay terminal).
This commit is contained in:
kshitijk4poor
2026-09-27 13:09:34 +05:30
committed by kshitij
parent 8c9f2162c7
commit d8be403097
1 changed files with 4 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
agent/context_compressor.py
View File

@@ -745,9 +745,10 @@ def _classify_summary_failure(e: Exception) -> _SummaryFailureKind:
"""
status = _exc_status_code(e)
err = str(e).lower()
# The Codex aux stream guard's mid-stream stall is transient (#124077). Its no-progress and
# hard-ceiling timeouts deliberately stay terminal network failures.
stall = isinstance(e, TimeoutError) and CODEX_STREAM_STALL_MARKER in err
# #124077: only the Codex aux stream guard's mid-stream stall is a retry-ladder timeout; real
# transport timeouts and the guard's no-progress/hard-ceiling timeouts stay terminal
# network failures (#29559/#94448).
stall = isinstance(e, TimeoutError) and CODEX_STREAM_STALL_MARKER in str(e)
return _SummaryFailureKind(
# Permanent-looking error on a distinct summary model: fall back to main instead of cooldown.
model_not_found=status in {404, 503}
@@ -757,8 +758,6 @@ def _classify_summary_failure(e: Exception) -> _SummaryFailureKind:
# APIResponseValidationError "expecting value"; treat as transient.
json_decode=isinstance(e, json.JSONDecodeError) or "expecting value" in err,
# httpx premature-close errors are transient; treat like a timeout, not a 60s cooldown.
# Real transport timeouts stay here (terminal abort, #29559/#94448); only the stream-guard
# stall is a retry-ladder timeout, not a network failure (#124077).
streaming_closed=_is_connection_error(e) and not stall,
# HTTP 200 with empty body from a degraded provider, plus the sibling "no usable response"
# shapes from _validate_llm_response.