Timmy_Foundation/hermes-agent
15
0
Fork 0
Code Issues 258 Pull Requests 56 Actions Packages Projects Releases Wiki Activity

fix(delegate): give subagents independent iteration budgets (#3004)

Browse Source 
Each subagent now gets its own IterationBudget instead of sharing the
parent's.  The per-subagent cap is controlled by delegation.max_iterations
in config.yaml (default 50).  Total iterations across parent + subagents
can exceed the parent's max_iterations, but the user retains control via
the config setting.

Previously, subagents shared the parent's budget, so three parallel
subagents configured for max_iterations=50 racing against a parent that
already used 60 of 90 would each only get ~10 iterations.

Inspired by PR #2928 (Bartok9) which identified the issue (#2873).
This commit is contained in:
Teknium
2026-03-25 11:29:49 -07:00
committed by GitHub
parent 65dace1b1a
commit 68ab37e891
3 changed files with 17 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
hermes_cli/config.py
View File

@@ -317,6 +317,8 @@ DEFAULT_CONFIG = {
"provider": "", # e.g. "openrouter" (empty = inherit parent provider + credentials)
"base_url": "", # direct OpenAI-compatible endpoint for subagents
"api_key": "", # API key for delegation.base_url (falls back to OPENAI_API_KEY)
"max_iterations": 50, # per-subagent iteration cap (each subagent gets its own budget,
# independent of the parent's max_iterations)
},
# Ephemeral prefill messages file — JSON list of {role, content} dicts

16
run_agent.py
View File

@@ -162,11 +162,15 @@ def _install_safe_stdio() -> None:
class IterationBudget:
"""Thread-safe shared iteration counter for parent and child agents.
"""Thread-safe iteration counter for an agent.
Tracks total LLM-call iterations consumed across a parent agent and all
its subagents. A single ``IterationBudget`` is created by the parent
and passed to every child so they share the same cap.
Each agent (parent or subagent) gets its own ``IterationBudget``.
The parent's budget is capped at ``max_iterations`` (default 90).
Each subagent gets an independent budget capped at
``delegation.max_iterations`` (default 50) — this means total
iterations across parent + subagents can exceed the parent's cap.
Users control the per-subagent limit via ``delegation.max_iterations``
in config.yaml.
``execute_code`` (programmatic tool calling) iterations are refunded via
:meth:`refund` so they don't eat into the budget.
@@ -5719,7 +5723,7 @@ class AIAgent:
api_call_count += 1
if not self.iteration_budget.consume():
if not self.quiet_mode:
self._safe_print(f"\n⚠️ Session iteration budget exhausted ({self.iteration_budget.max_total} total across agent + subagents)")
self._safe_print(f"\n⚠️ Iteration budget exhausted ({self.iteration_budget.used}/{self.iteration_budget.max_total} iterations used)")
break
# Fire step_callback for gateway hooks (agent:step event)
@@ -7178,7 +7182,7 @@ class AIAgent:
or self.iteration_budget.remaining <= 0
):
if self.iteration_budget.remaining <= 0 and not self.quiet_mode:
print(f"\n⚠️ Session iteration budget exhausted ({self.iteration_budget.used}/{self.iteration_budget.max_total} used, including subagents)")
print(f"\n⚠️ Iteration budget exhausted ({self.iteration_budget.used}/{self.iteration_budget.max_total} iterations used)")
final_response = self._handle_max_iterations(messages, api_call_count)
# Determine if conversation completed successfully

9
tools/delegate_tool.py
View File

@@ -191,9 +191,10 @@ def _build_child_agent(
# Build progress callback to relay tool calls to parent display
child_progress_cb = _build_child_progress_callback(task_index, parent_agent)
# Share the parent's iteration budget so subagent tool calls
# count toward the session-wide limit.
shared_budget = getattr(parent_agent, "iteration_budget", None)
# Each subagent gets its own iteration budget capped at max_iterations
# (configurable via delegation.max_iterations, default 50). This means
# total iterations across parent + subagents can exceed the parent's
# max_iterations. The user controls the per-subagent cap in config.yaml.
# Resolve effective credentials: config override > parent inherit
effective_model = model or parent_agent.model
@@ -230,7 +231,7 @@ def _build_child_agent(
providers_order=parent_agent.providers_order,
provider_sort=parent_agent.provider_sort,
tool_progress_callback=child_progress_cb,
iteration_budget=shared_budget,
iteration_budget=None, # fresh budget per subagent
)
# Set delegation depth so children can't spawn grandchildren
child._delegate_depth = getattr(parent_agent, '_delegate_depth', 0) + 1