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 2026-04-16 10:47:41 -05:00
commit 9c71f3a6ea
55 changed files with 1413 additions and 3197 deletions
Download patch file Download diff file Expand all files Collapse all files

107
run_agent.py
View file

@ -540,13 +540,6 @@ class AIAgent:
for AI models that support function calling.
"""
# ── Class-level context pressure dedup (survives across instances) ──
# The gateway creates a new AIAgent per message, so instance-level flags
# reset every time. This dict tracks {session_id: (warn_level, timestamp)}
# to suppress duplicate warnings within a cooldown window.
_context_pressure_last_warned: dict = {}
_CONTEXT_PRESSURE_COOLDOWN = 300 # seconds between re-warning same session
@property
def base_url(self) -> str:
return self._base_url
@ -826,12 +819,6 @@ class AIAgent:
self._budget_exhausted_injected = False
self._budget_grace_call = False
# Context pressure warnings: notify the USER (not the LLM) as context
# fills up. Purely informational — displayed in CLI output and sent via
# status_callback for gateway platforms. Does NOT inject into messages.
# Tiered: fires at 85% and again at 95% of compaction threshold.
self._context_pressure_warned_at = 0.0 # highest tier already shown
# Activity tracking — updated on each API call, tool execution, and
# stream chunk. Used by the gateway timeout handler to report what the
# agent was doing when it was killed, and by the "still working"
@ -4353,6 +4340,15 @@ class AIAgent:
def _create_openai_client(self, client_kwargs: dict, *, reason: str, shared: bool) -> Any:
from agent.auxiliary_client import _validate_base_url, _validate_proxy_env_urls
# Treat client_kwargs as read-only. Callers pass self._client_kwargs (or shallow
# copies of it) in; any in-place mutation leaks back into the stored dict and is
# reused on subsequent requests. #10933 hit this by injecting an httpx.Client
# transport that was torn down after the first request, so the next request
# wrapped a closed transport and raised "Cannot send a request, as the client
# has been closed" on every retry. The revert resolved that specific path; this
# copy locks the contract so future transport/keepalive work can't reintroduce
# the same class of bug.
client_kwargs = dict(client_kwargs)
_validate_proxy_env_urls()
_validate_base_url(client_kwargs.get("base_url"))
if self.provider == "copilot-acp" or str(client_kwargs.get("base_url", "")).startswith("acp://copilot"):
@ -7221,20 +7217,6 @@ class AIAgent:
self.context_compressor.last_prompt_tokens = _compressed_est
self.context_compressor.last_completion_tokens = 0
# Only reset the pressure warning if compression actually brought
# us below the warning level (85% of threshold). When compression
# can't reduce enough (e.g. threshold is very low, or system prompt
# alone exceeds the warning level), keep the tier set to prevent
# spamming the user with repeated warnings every loop iteration.
if self.context_compressor.threshold_tokens > 0:
_post_progress = _compressed_est / self.context_compressor.threshold_tokens
if _post_progress < 0.85:
self._context_pressure_warned_at = 0.0
# Clear class-level dedup for this session so a fresh
# warning cycle can start if context grows again.
_sid = self.session_id or "default"
AIAgent._context_pressure_last_warned.pop(_sid, None)
# Clear the file-read dedup cache. After compression the original
# read content is summarised away — if the model re-reads the same
# file it needs the full content, not a "file unchanged" stub.
@ -8034,45 +8016,6 @@ class AIAgent:
def _emit_context_pressure(self, compaction_progress: float, compressor) -> None:
"""Notify the user that context is approaching the compaction threshold.
Args:
compaction_progress: How close to compaction (0.0–1.0, where 1.0 = fires).
compressor: The ContextCompressor instance (for threshold/context info).
Purely user-facing — does NOT modify the message stream.
For CLI: prints a formatted line with a progress bar.
For gateway: fires status_callback so the platform can send a chat message.
"""
from agent.display import format_context_pressure, format_context_pressure_gateway
threshold_pct = compressor.threshold_tokens / compressor.context_length if compressor.context_length else 0.5
# CLI output — always shown (these are user-facing status notifications,
# not verbose debug output, so they bypass quiet_mode).
# Gateway users also get the callback below.
if self.platform in (None, "cli"):
line = format_context_pressure(
compaction_progress=compaction_progress,
threshold_tokens=compressor.threshold_tokens,
threshold_percent=threshold_pct,
compression_enabled=self.compression_enabled,
)
self._safe_print(line)
# Gateway / external consumers
if self.status_callback:
try:
msg = format_context_pressure_gateway(
compaction_progress=compaction_progress,
threshold_percent=threshold_pct,
compression_enabled=self.compression_enabled,
)
self.status_callback("context_pressure", msg)
except Exception:
logger.debug("status_callback error in context pressure", exc_info=True)
def _handle_max_iterations(self, messages: list, api_call_count: int) -> str:
"""Request a summary when max iterations are reached. Returns the final response text."""
print(f"⚠️ Reached maximum iterations ({self.max_iterations}). Requesting summary...")
@ -10800,38 +10743,6 @@ class AIAgent:
else:
_real_tokens = estimate_messages_tokens_rough(messages)
# ── Context pressure warnings (user-facing only) ──────────
# Notify the user (NOT the LLM) as context approaches the
# compaction threshold. Thresholds are relative to where
# compaction fires, not the raw context window.
# Does not inject into messages — just prints to CLI output
# and fires status_callback for gateway platforms.
# Tiered: 85% (orange) and 95% (red/critical).
if _compressor.threshold_tokens > 0:
_compaction_progress = _real_tokens / _compressor.threshold_tokens
# Determine the warning tier for this progress level
_warn_tier = 0.0
if _compaction_progress >= 0.95:
_warn_tier = 0.95
elif _compaction_progress >= 0.85:
_warn_tier = 0.85
if _warn_tier > self._context_pressure_warned_at:
# Class-level dedup: check if this session was already
# warned at this tier within the cooldown window.
_sid = self.session_id or "default"
_last = AIAgent._context_pressure_last_warned.get(_sid)
_now = time.time()
if _last is None or _last[0] < _warn_tier or (_now - _last[1]) >= self._CONTEXT_PRESSURE_COOLDOWN:
self._context_pressure_warned_at = _warn_tier
AIAgent._context_pressure_last_warned[_sid] = (_warn_tier, _now)
self._emit_context_pressure(_compaction_progress, _compressor)
# Evict stale entries (older than 2x cooldown)
_cutoff = _now - self._CONTEXT_PRESSURE_COOLDOWN * 2
AIAgent._context_pressure_last_warned = {
k: v for k, v in AIAgent._context_pressure_last_warned.items()
if v[1] > _cutoff
}
if self.compression_enabled and _compressor.should_compress(_real_tokens):
self._safe_print(" ⟳ compacting context…")
messages, active_system_prompt = self._compress_context(