Summary

_execute_tool_calls_sequential in run_agent.py doesn't route memory-provider tools through MemoryManager, so honcho_profile, honcho_search, honcho_context, honcho_conclude (and any future memory-plugin tools) return {"error": "Unknown tool: <name>"} whenever the assistant makes a single tool call or the sequential dispatch path is otherwise selected. The concurrent path (_invoke_tool) handles this correctly, so multi-tool turns sometimes succeed — which makes the bug intermittent and easy to mistake for a config problem.

Reproduction

1. Configure Honcho per docs (hermes memory setup, or drop a valid ~/.hermes/honcho.json) with memory.provider: honcho and recall_mode: hybrid so honcho tool schemas are injected (confirmed via session JSON — schemas are present in the LLM payload).
2. Start any gateway (Telegram, Slack) or a non-interactive CLI chat.
3. In a turn where the assistant produces exactly one tool call, have the LLM call honcho_profile (e.g. prompt: "Call honcho_profile with no arguments").

$ hermes chat -Q --max-turns 2 -q "Call only honcho_profile and report the JSON result verbatim"
  ┊ ⚡ honcho_pr   0.0s [error]
# tool response: {"error": "Unknown tool: honcho_profile"}

Error origin: tools/registry.py:153 — _execute_tool_calls_sequential falls through to handle_function_call() which consults the main tool registry; honcho tools are registered with MemoryManager, not the main registry, so dispatch fails.

Expected

Same behavior as the concurrent path: MemoryManager.has_tool() check, then MemoryManager.handle_tool_call() dispatch. The tool returns a result like {"result": "No profile facts available yet."} instead of {"error": "Unknown tool"}.

Root cause

Two tool-dispatch paths in run_agent.py:

Schemas get injected correctly by MemoryManager.get_all_tool_schemas() at ~L1094 and valid_tool_names contains the tool names — so the LLM advertises/calls them. Only the sequential dispatch fails.

Suggested patch

Add the memory_manager branch to the sequential path, mirroring _invoke_tool:

# in _execute_tool_calls_sequential, between the delegate_task branch and
# the `elif self.quiet_mode:` fallthrough:
elif self._memory_manager and self._memory_manager.has_tool(function_name):
    try:
        function_result = self._memory_manager.handle_tool_call(function_name, function_args)
    except Exception as tool_error:
        function_result = json.dumps({"error": f"Memory tool '{function_name}' failed: {tool_error}"})
        logger.error("memory_manager.handle_tool_call raised for %s: %s", function_name, tool_error, exc_info=True)
    tool_duration = time.time() - tool_start_time
    if self.quiet_mode:
        self._vprint(f"  {_get_cute_tool_message_impl(function_name, function_args, tool_duration, result=function_result)}")

Applied locally against HEAD 29c98e8f; verified fix with the same one-shot CLI repro — honcho_profile now returns {"result": "No profile facts available yet."} as expected.

Affected

• All memory-plugin tools on single-tool-call turns (Telegram, Slack, CLI, any non-concurrent path).
• Makes memory tools appear "unavailable in this gateway", which misleads the model into claiming memory doesn't work — users file this as a config bug.

Environment

• hermes-agent HEAD: 29c98e8f
• Python 3.11
• Ubuntu 22.04 / systemd gateway
• Memory provider: honcho (hybrid mode)