Hermes: using-superpowers bootstrap is injected into delegate_task subagents (Hermes counterpart of #2160)

• I searched existing issues and this is not a duplicate. #2160 reports the same class for OpenCode; I found no Hermes report (searched "hermes subagent", "subagent bootstrap", "SUBAGENT-STOP"). #2223, the OpenCode fix, was closed unmerged and does not touch .hermes-plugin/.

Environment (required)

Is this a Superpowers issue or a platform issue?

• I confirmed this issue does not occur without Superpowers installed. The injected text is the plugin's own bootstrap and is the only pre_llm_call result returned.

The injection comes from .hermes-plugin/__init__.py. Hermes runs every pre_llm_call hook for child agents too; the only skip is _persist_disabled (agent/turn_context.py:801), which delegate_task children do not set.

What happened?

pre_llm_call returns the bootstrap whenever is_first_turn is true (.hermes-plugin/__init__.py:121). A delegate_task child starts with no history, so on its first turn it receives the full <EXTREMELY_IMPORTANT> bootstrap, byte-identical to the one the parent gets. The bootstrap then relies on <SUBAGENT-STOP> for the child to ignore it. #2160 shows that is not reliable: an OpenCode worker invoked brainstorming and restarted the design cycle instead of doing its bounded task.

Hermes already tells the hook that the caller is a child: delegate_task builds children with platform="subagent" (tools/delegate_tool.py:277), the agent keeps it (AIAgent(platform="subagent").platform == "subagent"), and _collect_pre_llm_call_context forwards it as platform=.

Steps to reproduce

No model call is needed. From the Hermes source checkout, with its own interpreter:

import types
from agent.turn_context import _collect_pre_llm_call_context
from agent.delegation_context import delegated_child_context

def stub(platform):
    return types.SimpleNamespace(session_id="probe", model="m", platform=platform,
                                 _parent_session_id="", _user_id="", _persist_disabled=False)

kw = dict(effective_task_id="t", turn_id="1", original_user_message="hi",
          messages=[{"role": "user", "content": "hi"}])
parent = _collect_pre_llm_call_context(stub("desktop"), conversation_history=None, **kw)
with delegated_child_context("probe-child"):
    child = _collect_pre_llm_call_context(stub("subagent"), conversation_history=None, **kw)
    later = _collect_pre_llm_call_context(stub("subagent"), conversation_history=[{"role": "user", "content": "x"}], **kw)
print(len(parent), len(child), len(later), parent == child)

Expected behavior

A delegate_task child gets no bootstrap: the parent has already chosen the process, and the child's brief carries what it needs.

Actual behavior

5729 5729 0 True

The parent and the subagent first turn both get 5,729 characters (roughly 1,400 tokens) of identical bootstrap. Later child turns get nothing.

Not verified on Hermes: whether a child actually ignores <SUBAGENT-STOP>. The behavioural failure is the one #2160 reports for OpenCode; I did not run a live delegation for this report.

Suggested fix

        if is_first_turn and platform != "subagent":
            return {"context": bootstrap}
        return None

Checked on a scratch copy of the adapter with this one-line change: a subagent first turn now gets nothing, while desktop and cron first turns still get the full bootstrap and later turns still get nothing.

platform is the explicit signal. parent_session_id does not mean "subagent" by itself: Hermes also records a parent for the continuation session it creates on context compression (agent/conversation_compression.py, publish_compression_child(parent_session_id=...)).

Debug log or conversation transcript

Injected context is not persisted in Hermes' session records, so a transcript would not show it. The probe above is the reproduction.