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

fix(agent): exempt :cloud GLM models from stop->length truncation rewrite

Browse Source 
Ollama cloud models (model name contains ':cloud') run generation on
Ollama's hosted server; the local 11434 endpoint is only a transparent
proxy that forwards finish_reason faithfully.  _is_ollama_glm_backend()
was matching these models because the proxy listens on the same port
as local Ollama, causing _should_treat_stop_as_truncated() to rewrite
a correct finish_reason='stop' into 'length'.

The 4-attempt continuation loop then injects a synthetic user nudge
that reasoning-capable GLM models spend their output budget deliberating
over, producing unpunctuated tails that re-trigger _has_natural_response_ending()
rejection -- a self-reinforcing loop that always exhausts retries.

Fix: add an early return in _is_ollama_glm_backend when the model name
contains ':cloud'.  Local GLM inference (no ':cloud') is still caught
by the existing port/URL/provider checks.

Fixes #98406
This commit is contained in:
salch-cred
2026-08-30 12:27:43 +05:30
committed by Teknium
parent 5360886f54
commit bbed304536
1 changed files with 11 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
run_agent.py
View File

@@ -1886,17 +1886,23 @@ class AIAgent:
report finish_reason correctly and were the source of #13971's
false-positive truncation continuations.
Also excludes Ollama Cloud (ollama.com) — the hosted service
correctly reports finish_reason and is not affected by the local
Ollama stop-reason bug (GH-72316).
Also excludes Ollama Cloud — the hosted service correctly reports
finish_reason and is not affected by the local Ollama stop-reason
bug (GH-72316). Two signatures identify it: the ``ollama.com`` host
(provider ``ollama-cloud``) and the ``:cloud`` model suffix (cloud
generation proxied through a local 11434 endpoint, #98406). Applying
the stop→length rewrite to them manufactures false truncations and
causes the continuation nudge to consume the model's output budget
on the next retry, making further false-positives more likely.
"""
model_lower = (self.model or "").lower()
provider_lower = (self.provider or "").lower()
if "glm" not in model_lower and provider_lower != "zai":
return False
base = self._base_url_lower
# Exclude Ollama Cloud (ollama.com) — hosted service, not local Ollama
if "ollama.com" in base:
# Ollama Cloud (hosted service or :cloud proxy) forwards finish_reason
# faithfully — do not rewrite.
if "ollama.com" in base or ":cloud" in model_lower:
return False
if "ollama" in base or ":11434" in base:
return True