ling/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-04-28 06:51:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

wip: tool result fixes -- persistence

Browse Source
This commit is contained in:
alt-glitch
2026-04-07 22:21:27 -07:00
committed by Teknium
parent 22d1bda185
commit 65e24c942e
11 changed files with 869 additions and 235 deletions
Download Patch File Download Diff File Expand all files Collapse all files

89
run_agent.py
View File

@@ -66,7 +66,8 @@ from model_tools import (
handle_function_call,
check_toolset_requirements,
)
from tools.terminal_tool import cleanup_vm
from tools.terminal_tool import cleanup_vm, get_active_env
from tools.tool_result_storage import maybe_persist_tool_result, enforce_turn_budget
from tools.interrupt import set_interrupt as _set_interrupt
from tools.browser_tool import cleanup_browser
@@ -411,63 +412,6 @@ def _strip_budget_warnings_from_history(messages: list) -> None:
# Large tool result handler — save oversized output to temp file
# =========================================================================
# Threshold at which tool results are saved to a file instead of kept inline.
# 100K chars ≈ 25K tokens — generous for any reasonable output but prevents
# catastrophic context explosions.
_LARGE_RESULT_CHARS = 100_000
# How many characters of the original result to include as an inline preview
# so the model has immediate context about what the tool returned.
_LARGE_RESULT_PREVIEW_CHARS = 1_500
def _save_oversized_tool_result(function_name: str, function_result: str) -> str:
"""Replace oversized tool results with a file reference + preview.
When a tool returns more than ``_LARGE_RESULT_CHARS`` characters, the full
content is written to a temporary file under ``HERMES_HOME/cache/tool_responses/``
and the result sent to the model is replaced with:
• a brief head preview (first ``_LARGE_RESULT_PREVIEW_CHARS`` chars)
• the file path so the model can use ``read_file`` / ``search_files``
Falls back to destructive truncation if the file write fails.
"""
original_len = len(function_result)
if original_len <= _LARGE_RESULT_CHARS:
return function_result
# Build the target directory
try:
response_dir = os.path.join(get_hermes_home(), "cache", "tool_responses")
os.makedirs(response_dir, exist_ok=True)
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
# Sanitize tool name for use in filename
safe_name = re.sub(r"[^\w\-]", "_", function_name)[:40]
filename = f"{safe_name}_{timestamp}.txt"
filepath = os.path.join(response_dir, filename)
with open(filepath, "w", encoding="utf-8") as f:
f.write(function_result)
preview = function_result[:_LARGE_RESULT_PREVIEW_CHARS]
return (
f"{preview}\n\n"
f"[Large tool response: {original_len:,} characters total — "
f"only the first {_LARGE_RESULT_PREVIEW_CHARS:,} shown above. "
f"Full output saved to: {filepath}\n"
f"Use read_file or search_files on that path to access the rest.]"
)
except Exception as exc:
# Fall back to destructive truncation if file write fails
logger.warning("Failed to save large tool result to file: %s", exc)
return (
function_result[:_LARGE_RESULT_CHARS]
+ f"\n\n[Truncated: tool response was {original_len:,} chars, "
f"exceeding the {_LARGE_RESULT_CHARS:,} char limit. "
f"File save failed: {exc}]"
)
class AIAgent:
"""
@@ -6262,15 +6206,17 @@ class AIAgent:
except Exception as cb_err:
logging.debug(f"Tool complete callback error: {cb_err}")
# Save oversized results to file instead of destructive truncation
function_result = _save_oversized_tool_result(name, function_result)
function_result = maybe_persist_tool_result(
content=function_result,
tool_name=name,
tool_use_id=tc.id,
env=get_active_env(effective_task_id),
)
# Discover subdirectory context files from tool arguments
subdir_hints = self._subdirectory_hints.check_tool_call(name, args)
if subdir_hints:
function_result += subdir_hints
# Append tool result message in order
tool_msg = {
"role": "tool",
"content": function_result,
@@ -6278,6 +6224,12 @@ class AIAgent:
}
messages.append(tool_msg)
# ── Per-turn aggregate budget enforcement ─────────────────────────
num_tools = len(parsed_calls)
if num_tools > 0:
turn_tool_msgs = messages[-num_tools:]
enforce_turn_budget(turn_tool_msgs, env=get_active_env(effective_task_id))
# ── Budget pressure injection ────────────────────────────────────
budget_warning = self._get_budget_warning(api_call_count)
if budget_warning and messages and messages[-1].get("role") == "tool":
@@ -6562,8 +6514,12 @@ class AIAgent:
except Exception as cb_err:
logging.debug(f"Tool complete callback error: {cb_err}")
# Save oversized results to file instead of destructive truncation
function_result = _save_oversized_tool_result(function_name, function_result)
function_result = maybe_persist_tool_result(
content=function_result,
tool_name=function_name,
tool_use_id=tool_call.id,
env=get_active_env(effective_task_id),
)
# Discover subdirectory context files from tool arguments
subdir_hints = self._subdirectory_hints.check_tool_call(function_name, function_args)
@@ -6601,6 +6557,11 @@ class AIAgent:
if self.tool_delay > 0 and i < len(assistant_message.tool_calls):
time.sleep(self.tool_delay)
# ── Per-turn aggregate budget enforcement ─────────────────────────
num_tools_seq = len(assistant_message.tool_calls)
if num_tools_seq > 0:
enforce_turn_budget(messages[-num_tools_seq:], env=get_active_env(effective_task_id))
# ── Budget pressure injection ─────────────────────────────────
# After all tool calls in this turn are processed, check if we're
# approaching max_iterations. If so, inject a warning into the LAST