PR #24812: fix(agent): exit loop early when terminal command is blocked by user

• Repository: NousResearch/hermes-agent
• Author: waynehuu
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/24812

Description (problem / solution / changelog)

Closes #24806

Problem

When a user denies a terminal command via the approval dialog, terminal_tool returns a result with status: "blocked". The agent loop in _execute_tool_calls_sequential appended this as a normal tool message and continued running. The model was expected to stop based on the error text, but in practice the agent could keep running — and during gateway drain (e.g. triggered by a Feishu WS reconnect), this kept the agent alive for the full 60s timeout before being force-killed.

Fix

After each tool execution batch, scan the most recent tool messages for any terminal result with status == "blocked". If found, break the agent loop immediately with turn_exit_reason = "terminal_blocked", mirroring the existing guardrail_halt pattern.

Changes

• run_agent.py: add blocked-terminal check after the guardrail halt check in the main agent loop

Changed files

• run_agent.py (modified, +23/-0)

PR #24908: fix(run_agent): exit agent loop early when terminal command is denied

• Repository: NousResearch/hermes-agent
• Author: shanewas
• State: closed | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/24908

Description (problem / solution / changelog)

Summary

Fixes #24806: When a user denies a terminal command in the approval dialog (returns ), the agent was continuing to process remaining tool calls instead of exiting the loop.

Root Cause

in was not checking for on terminal results. It would record the blocked result and continue looping through remaining tool calls — creating a feedback loop where the model retries the same denied command repeatedly.

What was changed

**** — Three changes:

1. Added attribute (line ~10415): A -typed attribute that stores the halt signal, analogous to .

2. Added method (line ~10428): Creates and records a with , . Idempotent — no-op if already set.

3. **Modified ** (line ~11385): After processing a tool result, checks for on a terminal tool call and calls . This causes 's existing guardrail-halt check to exit the loop immediately.

Why this approach

• Reuses existing infrastructure for consistency
• mirrors so handles both decision types uniformly via its existing halt check at line ~14780
• No new exit paths — just wired into the existing guardrail halt mechanism
• Idempotent: second calls are no-ops, so model retries don't double-set

Tests

— 9 tests covering:

• sets the decision correctly
• Second call is idempotent (same object reused)
• Decision is cleared between turns (via )
• exits early on blocked terminal
• Non-blocked terminal result does NOT trigger early exit
• Non-terminal blocked result does NOT trigger early exit
• Subsequent tool calls are skipped after blocked terminal
• Blocked result is added to messages before break
• is called with the blocked message

9/9 pass, ruff clean.

Changed files

• run_agent.py (modified, +39/-0)
• tests/run_agent/test_terminal_blocked_exit.py (added, +209/-0)