mirrors/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-04-25 00:51:20 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'main' of github.com:NousResearch/hermes-agent into feat/ink-refactor

Browse source
This commit is contained in:
Brooklyn Nicholson 2026-04-15 17:43:41 -05:00
commit 72aebfbb24
21 changed files with 376 additions and 49 deletions
Download patch file Download diff file Expand all files Collapse all files

91
run_agent.py
View file

@ -457,6 +457,15 @@ def _sanitize_messages_non_ascii(messages: list) -> bool:
if sanitized != fn_args:
fn["arguments"] = sanitized
found = True
# Sanitize any additional top-level string fields (e.g. reasoning_content)
for key, value in msg.items():
if key in {"content", "name", "tool_calls", "role"}:
continue
if isinstance(value, str):
sanitized = _strip_non_ascii(value)
if sanitized != value:
msg[key] = sanitized
found = True
return found
@ -705,12 +714,13 @@ class AIAgent:
except Exception:
pass
# GPT-5.x models require the Responses API path — they are rejected
# on /v1/chat/completions by both OpenAI and OpenRouter. Also
# auto-upgrade for direct OpenAI URLs (api.openai.com) since all
# newer tool-calling models prefer Responses there.
# ACP runtimes are excluded: CopilotACPClient handles its own
# routing and does not implement the Responses API surface.
# GPT-5.x models usually require the Responses API path, but some
# providers have exceptions (for example Copilot's gpt-5-mini still
# uses chat completions). Also auto-upgrade for direct OpenAI URLs
# (api.openai.com) since all newer tool-calling models prefer
# Responses there. ACP runtimes are excluded: CopilotACPClient
# handles its own routing and does not implement the Responses API
# surface.
if (
self.api_mode == "chat_completions"
and self.provider != "copilot-acp"
@ -718,7 +728,10 @@ class AIAgent:
and not str(self.base_url or "").lower().startswith("acp+tcp://")
and (
self._is_direct_openai_url()
or self._model_requires_responses_api(self.model)
or self._provider_model_requires_responses_api(
self.model,
provider=self.provider,
)
)
):
self.api_mode = "codex_responses"
@ -1951,6 +1964,24 @@ class AIAgent:
m = m.rsplit("/", 1)[-1]
return m.startswith("gpt-5")
@staticmethod
def _provider_model_requires_responses_api(
model: str,
*,
provider: Optional[str] = None,
) -> bool:
"""Return True when this provider/model pair should use Responses API."""
normalized_provider = (provider or "").strip().lower()
if normalized_provider == "copilot":
try:
from hermes_cli.models import _should_use_copilot_responses_api
return _should_use_copilot_responses_api(model)
except Exception:
# Fall back to the generic GPT-5 rule if Copilot-specific
# logic is unavailable for any reason.
pass
return AIAgent._model_requires_responses_api(model)
def _max_tokens_param(self, value: int) -> dict:
"""Return the correct max tokens kwarg for the current provider.
@ -5721,9 +5752,13 @@ class AIAgent:
fb_api_mode = "anthropic_messages"
elif self._is_direct_openai_url(fb_base_url):
fb_api_mode = "codex_responses"
elif self._model_requires_responses_api(fb_model):
# GPT-5.x models need Responses API on every provider
# (OpenRouter, Copilot, direct OpenAI, etc.)
elif self._provider_model_requires_responses_api(
fb_model,
provider=fb_provider,
):
# GPT-5.x models usually need Responses API, but keep
# provider-specific exceptions like Copilot gpt-5-mini on
# chat completions.
fb_api_mode = "codex_responses"
old_model = self.model
@ -9108,7 +9143,19 @@ class AIAgent:
# ASCII codec: the system encoding can't handle
# non-ASCII characters at all. Sanitize all
# non-ASCII content from messages/tool schemas and retry.
# Sanitize both the canonical `messages` list and
# `api_messages` (the API-copy built before the retry
# loop, which may contain extra fields like
# reasoning_content that are not in `messages`).
_messages_sanitized = _sanitize_messages_non_ascii(messages)
if isinstance(api_messages, list):
_sanitize_messages_non_ascii(api_messages)
# Also sanitize the last api_kwargs if already built,
# so a leftover non-ASCII value in a transformed field
# (e.g. extra_body, reasoning_content) doesn't survive
# into the next attempt via _build_api_kwargs cache paths.
if isinstance(api_kwargs, dict):
_sanitize_structure_non_ascii(api_kwargs)
_prefill_sanitized = False
if isinstance(getattr(self, "prefill_messages", None), list):
_prefill_sanitized = _sanitize_messages_non_ascii(self.prefill_messages)
@ -9166,22 +9213,34 @@ class AIAgent:
force=True,
)
if (
# Always retry on ASCII codec detection —
# _force_ascii_payload guarantees the full
# api_kwargs payload is sanitized on the
# next iteration (line ~8475). Even when
# per-component checks above find nothing
# (e.g. non-ASCII only in api_messages'
# reasoning_content), the flag catches it.
# Bounded by _unicode_sanitization_passes < 2.
self._unicode_sanitization_passes += 1
_any_sanitized = (
_messages_sanitized
or _prefill_sanitized
or _tools_sanitized
or _system_sanitized
or _headers_sanitized
or _credential_sanitized
):
self._unicode_sanitization_passes += 1
)
if _any_sanitized:
self._vprint(
f"{self.log_prefix}⚠️ System encoding is ASCII — stripped non-ASCII characters from request payload. Retrying...",
force=True,
)
continue
# Nothing to sanitize in any payload component.
# Fall through to normal error path.
else:
self._vprint(
f"{self.log_prefix}⚠️ System encoding is ASCII — enabling full-payload sanitization for retry...",
force=True,
)
continue
status_code = getattr(api_error, "status_code", None)
error_context = self._extract_api_error_context(api_error)