mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-07-29 18:46:59 +00:00
c883367bd2
A mid-stream steer persists an interrupted-turn checkpoint so the model knows its reply was cut off. That scaffolding — "[This response was interrupted by a user correction.]" and the "Visible response before the interruption:" header — was written straight into message content, so every reload painted the raw machinery as an assistant bubble (and merged it into the preceding tool-call bubble). Steered transcripts became unreadable. Reuse the existing display/replay split instead of inventing new surface: - Carry the scaffolded form in the server-only api_content sidecar (the exact bytes replayed to the provider), keep content the user's/agent's real words. - When nothing reached the screen there is no clean form, so mark the row display_kind=hidden — replayed to the model, dropped by every transcript surface, exactly like compaction-reference rows. - Honor display_kind=hidden in the gateway's _history_to_messages projection (it only sniffed the [System: convention), so the checkpoint can't leak through the live/resume path to the TUI/CLI either. The model still receives the full interrupted context on the wire; the transcript shows the partial reply and the user's correction.
610 lines
25 KiB
Python
610 lines
25 KiB
Python
"""Tests for AIAgent.steer() — mid-run user message injection.
|
|
|
|
/steer lets the user add a note to the agent's next tool result without
|
|
interrupting the current tool call. The agent sees the note inline with
|
|
tool output on its next iteration, preserving message-role alternation
|
|
and prompt-cache integrity.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import threading
|
|
|
|
import pytest
|
|
|
|
from agent.prompt_builder import STEER_MARKER_OPEN, format_steer_marker
|
|
from run_agent import AIAgent
|
|
|
|
|
|
def _bare_agent() -> AIAgent:
|
|
"""Build an AIAgent without running __init__, then install the steer
|
|
state manually — matches the existing object.__new__ stub pattern
|
|
used elsewhere in the test suite.
|
|
"""
|
|
agent = object.__new__(AIAgent)
|
|
agent._pending_steer = None
|
|
agent._pending_steer_lock = threading.Lock()
|
|
agent._pending_redirect = None
|
|
agent._pending_redirect_lock = threading.Lock()
|
|
agent._model_request_active = threading.Event()
|
|
agent._executing_tools = False
|
|
agent._execution_thread_id = None
|
|
agent._interrupt_thread_signal_pending = False
|
|
agent._interrupt_requested = False
|
|
agent._interrupt_message = None
|
|
agent._active_children = []
|
|
agent._active_children_lock = threading.Lock()
|
|
agent._tool_worker_threads = None
|
|
agent._tool_worker_threads_lock = None
|
|
agent._current_streamed_assistant_text = ""
|
|
agent._stream_needs_break = False
|
|
agent._strip_think_blocks = lambda content: content
|
|
agent.quiet_mode = True
|
|
agent.api_mode = "chat_completions"
|
|
return agent
|
|
|
|
|
|
class TestSteerAcceptance:
|
|
def test_accepts_non_empty_text(self):
|
|
agent = _bare_agent()
|
|
assert agent.steer("go ahead and check the logs") is True
|
|
assert agent._pending_steer == "go ahead and check the logs"
|
|
|
|
def test_rejects_empty_string(self):
|
|
agent = _bare_agent()
|
|
assert agent.steer("") is False
|
|
assert agent._pending_steer is None
|
|
|
|
def test_rejects_whitespace_only(self):
|
|
agent = _bare_agent()
|
|
assert agent.steer(" \n\t ") is False
|
|
assert agent._pending_steer is None
|
|
|
|
def test_rejects_none(self):
|
|
agent = _bare_agent()
|
|
assert agent.steer(None) is False # type: ignore[arg-type]
|
|
assert agent._pending_steer is None
|
|
|
|
def test_strips_surrounding_whitespace(self):
|
|
agent = _bare_agent()
|
|
assert agent.steer(" hello world \n") is True
|
|
assert agent._pending_steer == "hello world"
|
|
|
|
def test_concatenates_multiple_steers_with_newlines(self):
|
|
agent = _bare_agent()
|
|
agent.steer("first note")
|
|
agent.steer("second note")
|
|
agent.steer("third note")
|
|
assert agent._pending_steer == "first note\nsecond note\nthird note"
|
|
|
|
|
|
class TestSteerDrain:
|
|
def test_drain_returns_and_clears(self):
|
|
agent = _bare_agent()
|
|
agent.steer("hello")
|
|
assert agent._drain_pending_steer() == "hello"
|
|
assert agent._pending_steer is None
|
|
|
|
def test_drain_on_empty_returns_none(self):
|
|
agent = _bare_agent()
|
|
assert agent._drain_pending_steer() is None
|
|
|
|
|
|
class TestActiveTurnRedirect:
|
|
def test_rejects_when_no_turn_is_active(self):
|
|
agent = _bare_agent()
|
|
assert agent.redirect("change course") is False
|
|
assert agent._pending_redirect is None
|
|
|
|
def test_cancels_only_an_active_model_request(self):
|
|
agent = _bare_agent()
|
|
agent._model_request_active.set()
|
|
|
|
assert agent.redirect("use Postgres") is True
|
|
assert agent._pending_redirect == "use Postgres"
|
|
assert agent._interrupt_requested is True
|
|
assert agent._interrupt_message is None
|
|
|
|
def test_multiple_redirects_preserve_message_boundaries(self):
|
|
agent = _bare_agent()
|
|
agent._model_request_active.set()
|
|
|
|
assert agent.redirect("first correction") is True
|
|
assert agent.redirect("second correction") is True
|
|
assert agent._pending_redirect == (
|
|
"first correction\n\n"
|
|
"[Additional user correction]\n"
|
|
"second correction"
|
|
)
|
|
|
|
def test_hard_interrupt_wins_over_new_redirect(self):
|
|
agent = _bare_agent()
|
|
agent._model_request_active.set()
|
|
agent._interrupt_requested = True
|
|
|
|
assert agent.redirect("too late") is False
|
|
assert agent._pending_redirect is None
|
|
|
|
def test_reasoning_deltas_are_display_only(self):
|
|
"""Streamed reasoning must never accumulate into replayable transcript
|
|
state — an assistant checkpoint that inlines chain-of-thought trips
|
|
Anthropic's output classifier and permanently bricks the session
|
|
(deterministic empty-response storms on every replay)."""
|
|
agent = _bare_agent()
|
|
seen = []
|
|
agent.reasoning_callback = seen.append
|
|
|
|
agent._fire_reasoning_delta("visible provider thinking")
|
|
|
|
# Displayed to the surface, but never checkpointed anywhere.
|
|
assert seen == ["visible provider thinking"]
|
|
assert not getattr(agent, "_current_streamed_reasoning_text", "")
|
|
|
|
def test_response_completion_before_redirect_lock_rejects_correction(self):
|
|
agent = _bare_agent()
|
|
agent._model_request_active.set()
|
|
started = threading.Event()
|
|
outcome = {}
|
|
|
|
def redirect():
|
|
started.set()
|
|
outcome["accepted"] = agent.redirect("late correction")
|
|
|
|
with agent._pending_redirect_lock:
|
|
worker = threading.Thread(target=redirect)
|
|
worker.start()
|
|
assert started.wait(timeout=1)
|
|
# Mirrors conversation_loop clearing the request-active marker
|
|
# under this same lock before redirect can commit its slot.
|
|
agent._model_request_active.clear()
|
|
worker.join(timeout=1)
|
|
|
|
assert outcome["accepted"] is False
|
|
assert agent._pending_redirect is None
|
|
|
|
def test_hard_stop_wins_concurrent_redirect(self):
|
|
agent = _bare_agent()
|
|
agent._model_request_active.set()
|
|
start = threading.Barrier(3)
|
|
outcome = {}
|
|
|
|
def redirect():
|
|
start.wait()
|
|
outcome["redirect"] = agent.redirect("change course")
|
|
|
|
def hard_stop():
|
|
start.wait()
|
|
agent.interrupt("stop requested")
|
|
|
|
redirect_thread = threading.Thread(target=redirect)
|
|
stop_thread = threading.Thread(target=hard_stop)
|
|
redirect_thread.start()
|
|
stop_thread.start()
|
|
start.wait()
|
|
redirect_thread.join(timeout=1)
|
|
stop_thread.join(timeout=1)
|
|
|
|
assert redirect_thread.is_alive() is False
|
|
assert stop_thread.is_alive() is False
|
|
assert agent._interrupt_requested is True
|
|
assert agent._interrupt_message == "stop requested"
|
|
assert agent._pending_redirect is None
|
|
|
|
def test_codex_app_server_hard_stop_reaches_native_session(self):
|
|
agent = _bare_agent()
|
|
calls = []
|
|
agent.api_mode = "codex_app_server"
|
|
agent._codex_session = type(
|
|
"_CodexSession",
|
|
(),
|
|
{"request_interrupt": lambda self: calls.append("interrupt")},
|
|
)()
|
|
|
|
agent.interrupt()
|
|
|
|
assert calls == ["interrupt"]
|
|
|
|
def test_codex_app_server_redirect_rejects_after_hard_stop(self):
|
|
agent = _bare_agent()
|
|
calls = []
|
|
agent.api_mode = "codex_app_server"
|
|
agent._interrupt_requested = True
|
|
agent._codex_session = type(
|
|
"_CodexSession",
|
|
(),
|
|
{"request_steer": lambda self, text: calls.append(text) or True},
|
|
)()
|
|
|
|
assert agent.redirect("too late") is False
|
|
assert calls == []
|
|
|
|
def test_redirect_during_tool_execution_uses_safe_steer_boundary(self):
|
|
agent = _bare_agent()
|
|
agent._executing_tools = True
|
|
|
|
assert agent.redirect("also check migrations") is True
|
|
assert agent._pending_redirect is None
|
|
assert agent._pending_steer == "also check migrations"
|
|
assert agent._interrupt_requested is False
|
|
|
|
|
|
class TestActiveTurnRedirectCheckpoint:
|
|
def test_assistant_tail_puts_correction_last(self):
|
|
from agent.conversation_loop import _apply_active_turn_redirect
|
|
|
|
agent = _bare_agent()
|
|
agent._current_streamed_assistant_text = "Visible draft."
|
|
messages = [
|
|
{"role": "user", "content": "start"},
|
|
{"role": "assistant", "content": "committed assistant item"},
|
|
]
|
|
|
|
_apply_active_turn_redirect(agent, messages, "Use Postgres instead.")
|
|
|
|
assert [m["role"] for m in messages] == ["user", "assistant", "user"]
|
|
assert messages[-1]["role"] == "user"
|
|
assert messages[-1]["content"] == "Use Postgres instead."
|
|
assert sum(1 for m in messages if m["role"] == "assistant") == 1
|
|
# Scaffolding is provider-replay text, carried in the sidecar so the
|
|
# model still sees the interrupted context — never in the transcript.
|
|
replayed = messages[-1]["api_content"]
|
|
assert "Visible draft." in replayed
|
|
assert "Context from the interrupted assistant response" in replayed
|
|
assert replayed.endswith("Use Postgres instead.")
|
|
|
|
def test_scaffolding_never_lands_in_transcript_content(self):
|
|
"""The checkpoint machinery is for the MODEL, not the transcript.
|
|
|
|
Persisting ``[This response was interrupted by a user correction.]``
|
|
into ``content`` painted raw scaffolding as an assistant bubble on
|
|
every reload. It must ride in ``api_content`` (replayed to the
|
|
provider) while ``content`` stays clean, or be marked
|
|
``display_kind="hidden"`` when there is no clean form at all.
|
|
"""
|
|
from agent.conversation_loop import _apply_active_turn_redirect
|
|
|
|
scaffolding = (
|
|
"[This response was interrupted by a user correction.]",
|
|
"Visible response before the interruption:",
|
|
"[Context from the interrupted assistant response]",
|
|
)
|
|
|
|
for tail_role in ("tool", "assistant"):
|
|
for streamed in ("Partial reply on screen.", ""):
|
|
agent = _bare_agent()
|
|
agent._current_streamed_assistant_text = streamed
|
|
messages = [{"role": "user", "content": "start"}]
|
|
if tail_role == "assistant":
|
|
messages.append({"role": "assistant", "content": "committed"})
|
|
else:
|
|
messages.append(
|
|
{"role": "assistant", "tool_calls": [{"id": "a"}]}
|
|
)
|
|
messages.append(
|
|
{"role": "tool", "content": "out", "tool_call_id": "a"}
|
|
)
|
|
|
|
_apply_active_turn_redirect(agent, messages, "New direction.")
|
|
|
|
for msg in messages:
|
|
if msg.get("display_kind") == "hidden":
|
|
continue # dropped by every transcript surface
|
|
content = str(msg.get("content", ""))
|
|
for marker in scaffolding:
|
|
assert marker not in content, (
|
|
f"scaffolding leaked into visible content "
|
|
f"(tail={tail_role}, streamed={bool(streamed)}): {content!r}"
|
|
)
|
|
|
|
# The user's correction is always shown verbatim.
|
|
assert messages[-1]["content"] == "New direction."
|
|
# ...and the model still receives the interrupted context.
|
|
replayed = "".join(
|
|
str(m.get("api_content") or m.get("content", "")) for m in messages
|
|
)
|
|
assert "[This response was interrupted by a user correction.]" in replayed
|
|
if streamed:
|
|
assert streamed in replayed
|
|
|
|
def test_checkpoint_never_replays_chain_of_thought(self):
|
|
"""Raw CoT serialized into checkpoint content reads to Anthropic's
|
|
output classifier as reasoning-injection; because the checkpoint is
|
|
persisted and replayed on every later call, one redirect during a
|
|
thinking phase permanently bricked sessions with deterministic
|
|
empty-response storms (July 2026). Reasoning must never appear in
|
|
replayable content — in either the assistant-checkpoint or the
|
|
merged-user-correction shape."""
|
|
from agent.conversation_loop import _apply_active_turn_redirect
|
|
|
|
for tail_role in ("user", "assistant"):
|
|
agent = _bare_agent()
|
|
# Simulate a surface having displayed reasoning this turn.
|
|
agent._current_streamed_reasoning_text = "SECRET chain of thought."
|
|
agent._current_streamed_assistant_text = "Visible draft."
|
|
messages = [{"role": "user", "content": "start"}]
|
|
if tail_role == "assistant":
|
|
messages.append({"role": "assistant", "content": "committed"})
|
|
|
|
_apply_active_turn_redirect(agent, messages, "Change course.")
|
|
|
|
# Check BOTH the transcript content and the replayed sidecar —
|
|
# the sidecar is what actually reaches the provider.
|
|
serialized = "".join(
|
|
str(m.get("content", "")) + str(m.get("api_content") or "")
|
|
for m in messages
|
|
)
|
|
assert "SECRET chain of thought." not in serialized
|
|
assert "Reasoning shown before the interruption" not in serialized
|
|
assert "Visible draft." in serialized
|
|
|
|
def test_checkpoint_omits_reasoning_label_when_nothing_visible(self):
|
|
from agent.conversation_loop import _apply_active_turn_redirect
|
|
|
|
agent = _bare_agent()
|
|
agent._current_streamed_reasoning_text = "thinking only, no text yet"
|
|
messages = [{"role": "user", "content": "start"}]
|
|
|
|
_apply_active_turn_redirect(agent, messages, "New direction.")
|
|
|
|
checkpoint_row = messages[-2]
|
|
# Nothing was on screen, so the row exists only for the model: hidden
|
|
# from every transcript surface, scaffolding replayed via the sidecar.
|
|
assert checkpoint_row["display_kind"] == "hidden"
|
|
assert (
|
|
checkpoint_row["api_content"]
|
|
== "[This response was interrupted by a user correction.]"
|
|
)
|
|
assert messages[-1]["content"] == "New direction."
|
|
|
|
|
|
class TestSteerInjection:
|
|
def test_appends_to_last_tool_result(self):
|
|
agent = _bare_agent()
|
|
agent.steer("please also check auth.log")
|
|
messages = [
|
|
{"role": "user", "content": "what's in /var/log?"},
|
|
{"role": "assistant", "tool_calls": [{"id": "a"}, {"id": "b"}]},
|
|
{"role": "tool", "content": "ls output A", "tool_call_id": "a"},
|
|
{"role": "tool", "content": "ls output B", "tool_call_id": "b"},
|
|
]
|
|
agent._apply_pending_steer_to_tool_results(messages, num_tool_msgs=2)
|
|
# The LAST tool result is modified; earlier ones are untouched.
|
|
assert messages[2]["content"] == "ls output A"
|
|
assert "ls output B" in messages[3]["content"]
|
|
assert STEER_MARKER_OPEN in messages[3]["content"]
|
|
assert "please also check auth.log" in messages[3]["content"]
|
|
# And pending_steer is consumed.
|
|
assert agent._pending_steer is None
|
|
|
|
def test_no_op_when_no_steer_pending(self):
|
|
agent = _bare_agent()
|
|
messages = [
|
|
{"role": "assistant", "tool_calls": [{"id": "a"}]},
|
|
{"role": "tool", "content": "output", "tool_call_id": "a"},
|
|
]
|
|
agent._apply_pending_steer_to_tool_results(messages, num_tool_msgs=1)
|
|
assert messages[-1]["content"] == "output" # unchanged
|
|
|
|
def test_no_op_when_num_tool_msgs_zero(self):
|
|
agent = _bare_agent()
|
|
agent.steer("steer")
|
|
messages = [{"role": "user", "content": "hi"}]
|
|
agent._apply_pending_steer_to_tool_results(messages, num_tool_msgs=0)
|
|
# Steer should remain pending (nothing to drain into)
|
|
assert agent._pending_steer == "steer"
|
|
|
|
def test_marker_labels_text_as_out_of_band_user_message(self):
|
|
"""The injection marker must attribute the appended text to the user
|
|
via the explicit out-of-band marker (which the system prompt tells the
|
|
model to trust) — otherwise the model reads it as untrusted tool output
|
|
and refuses it as suspected prompt injection. Cache-safe: it only
|
|
rewrites existing tool content, never the message-role sequence.
|
|
"""
|
|
agent = _bare_agent()
|
|
agent.steer("stop after next step")
|
|
messages = [{"role": "tool", "content": "x", "tool_call_id": "1"}]
|
|
agent._apply_pending_steer_to_tool_results(messages, num_tool_msgs=1)
|
|
content = messages[-1]["content"]
|
|
assert STEER_MARKER_OPEN in content
|
|
assert "stop after next step" in content
|
|
|
|
def test_multimodal_content_list_preserved(self):
|
|
"""Anthropic-style list content should be preserved, with the steer
|
|
appended as a text block."""
|
|
agent = _bare_agent()
|
|
agent.steer("extra note")
|
|
original_blocks = [{"type": "text", "text": "existing output"}]
|
|
messages = [
|
|
{"role": "tool", "content": list(original_blocks), "tool_call_id": "1"}
|
|
]
|
|
agent._apply_pending_steer_to_tool_results(messages, num_tool_msgs=1)
|
|
new_content = messages[-1]["content"]
|
|
assert isinstance(new_content, list)
|
|
assert len(new_content) == 2
|
|
assert new_content[0] == {"type": "text", "text": "existing output"}
|
|
assert new_content[1]["type"] == "text"
|
|
assert "extra note" in new_content[1]["text"]
|
|
|
|
def test_restashed_when_no_tool_result_in_batch(self):
|
|
"""If the 'batch' contains no tool-role messages (e.g. all skipped
|
|
after an interrupt), the steer should be put back into the pending
|
|
slot so the caller's fallback path can deliver it."""
|
|
agent = _bare_agent()
|
|
agent.steer("ping")
|
|
messages = [
|
|
{"role": "user", "content": "x"},
|
|
{"role": "assistant", "content": "y"},
|
|
]
|
|
# Claim there were N tool msgs, but the tail has none — simulates
|
|
# the interrupt-cancelled case.
|
|
agent._apply_pending_steer_to_tool_results(messages, num_tool_msgs=2)
|
|
# Messages untouched
|
|
assert messages[-1]["content"] == "y"
|
|
# And the steer is back in pending so the fallback can grab it
|
|
assert agent._pending_steer == "ping"
|
|
|
|
|
|
class TestSteerThreadSafety:
|
|
def test_concurrent_steer_calls_preserve_all_text(self):
|
|
agent = _bare_agent()
|
|
N = 200
|
|
|
|
def worker(idx: int) -> None:
|
|
agent.steer(f"note-{idx}")
|
|
|
|
threads = [threading.Thread(target=worker, args=(i,)) for i in range(N)]
|
|
for t in threads:
|
|
t.start()
|
|
for t in threads:
|
|
t.join()
|
|
|
|
text = agent._drain_pending_steer()
|
|
assert text is not None
|
|
# Every single note must be preserved — none dropped by the lock.
|
|
lines = text.split("\n")
|
|
assert len(lines) == N
|
|
assert set(lines) == {f"note-{i}" for i in range(N)}
|
|
|
|
|
|
class TestSteerClearedOnInterrupt:
|
|
def test_clear_interrupt_drops_pending_steer(self):
|
|
"""A hard interrupt supersedes any pending steer — the agent's
|
|
next tool iteration won't happen, so delivering the steer later
|
|
would be surprising."""
|
|
agent = _bare_agent()
|
|
# Minimal surface needed by clear_interrupt()
|
|
agent._interrupt_requested = True
|
|
agent._interrupt_message = None
|
|
agent._interrupt_thread_signal_pending = False
|
|
agent._execution_thread_id = None
|
|
agent._tool_worker_threads = None
|
|
agent._tool_worker_threads_lock = None
|
|
|
|
agent.steer("will be dropped")
|
|
agent._pending_redirect = "also drop this"
|
|
assert agent._pending_steer == "will be dropped"
|
|
|
|
agent.clear_interrupt()
|
|
assert agent._pending_steer is None
|
|
assert agent._pending_redirect is None
|
|
|
|
|
|
class TestPreApiCallSteerDrain:
|
|
"""Test that steers arriving during an API call are drained before the
|
|
next API call — not deferred until the next tool batch. This is the
|
|
fix for the scenario where /steer sent during model thinking only lands
|
|
after the agent is completely done."""
|
|
|
|
def test_pre_api_drain_injects_into_last_tool_result(self):
|
|
"""If a steer is pending when the main loop starts building
|
|
api_messages, it should be injected into the last tool result
|
|
in the messages list."""
|
|
agent = _bare_agent()
|
|
# Simulate messages after a tool batch completed
|
|
messages = [
|
|
{"role": "user", "content": "do something"},
|
|
{"role": "assistant", "content": "ok", "tool_calls": [
|
|
{"id": "tc1", "function": {"name": "terminal", "arguments": "{}"}}
|
|
]},
|
|
{"role": "tool", "content": "output here", "tool_call_id": "tc1"},
|
|
]
|
|
# Steer arrives during API call (set after tool execution)
|
|
agent.steer("focus on error handling")
|
|
# Simulate what the pre-API-call drain does:
|
|
_pre_api_steer = agent._drain_pending_steer()
|
|
assert _pre_api_steer == "focus on error handling"
|
|
# Inject into last tool msg (mirrors the new code in run_conversation)
|
|
for _si in range(len(messages) - 1, -1, -1):
|
|
if messages[_si].get("role") == "tool":
|
|
messages[_si]["content"] += format_steer_marker(_pre_api_steer)
|
|
break
|
|
assert STEER_MARKER_OPEN in messages[-1]["content"]
|
|
assert "focus on error handling" in messages[-1]["content"]
|
|
assert agent._pending_steer is None
|
|
|
|
def test_pre_api_drain_restashes_when_no_tool_message(self):
|
|
"""If there are no tool results yet (first iteration), the steer
|
|
should be put back into _pending_steer for the post-tool drain."""
|
|
agent = _bare_agent()
|
|
messages = [
|
|
{"role": "user", "content": "hello"},
|
|
]
|
|
agent.steer("early steer")
|
|
_pre_api_steer = agent._drain_pending_steer()
|
|
assert _pre_api_steer == "early steer"
|
|
# No tool message found — put it back
|
|
found = False
|
|
for _si in range(len(messages) - 1, -1, -1):
|
|
if messages[_si].get("role") == "tool":
|
|
found = True
|
|
break
|
|
assert not found
|
|
# Restash
|
|
agent._pending_steer = _pre_api_steer
|
|
assert agent._pending_steer == "early steer"
|
|
|
|
def test_pre_api_drain_finds_tool_msg_past_assistant(self):
|
|
"""The pre-API drain should scan backwards past a non-tool message
|
|
(e.g., if an assistant message was somehow appended after tools)
|
|
and still find the tool result."""
|
|
agent = _bare_agent()
|
|
messages = [
|
|
{"role": "user", "content": "do something"},
|
|
{"role": "assistant", "content": "let me check", "tool_calls": [
|
|
{"id": "tc1", "function": {"name": "web_search", "arguments": "{}"}}
|
|
]},
|
|
{"role": "tool", "content": "search results", "tool_call_id": "tc1"},
|
|
]
|
|
agent.steer("change approach")
|
|
_pre_api_steer = agent._drain_pending_steer()
|
|
assert _pre_api_steer is not None
|
|
for _si in range(len(messages) - 1, -1, -1):
|
|
if messages[_si].get("role") == "tool":
|
|
messages[_si]["content"] += format_steer_marker(_pre_api_steer)
|
|
break
|
|
assert "change approach" in messages[2]["content"]
|
|
|
|
|
|
class TestSteerMarkerContract:
|
|
def test_system_prompt_note_describes_the_real_marker(self):
|
|
"""The system-prompt note tells the model which marker to trust; it
|
|
must reference the exact open/close the injector emits, or the model
|
|
trusts a marker that never appears (and vice-versa)."""
|
|
from agent.prompt_builder import STEER_CHANNEL_NOTE, STEER_MARKER_CLOSE
|
|
|
|
emitted = format_steer_marker("hi")
|
|
assert STEER_MARKER_OPEN in emitted and STEER_MARKER_CLOSE in emitted
|
|
assert STEER_MARKER_OPEN in STEER_CHANNEL_NOTE and STEER_MARKER_CLOSE in STEER_CHANNEL_NOTE
|
|
|
|
def test_marker_no_longer_uses_the_distrusted_label(self):
|
|
"""Regression: the bare 'User guidance:' line read as tool content and
|
|
got refused as injection — it must not come back."""
|
|
assert "User guidance:" not in format_steer_marker("hi")
|
|
|
|
|
|
class TestSteerCommandRegistry:
|
|
def test_steer_in_command_registry(self):
|
|
"""The /steer slash command must be registered so it reaches all
|
|
platforms (CLI, gateway, TUI autocomplete, Telegram/Slack menus).
|
|
"""
|
|
from hermes_cli.commands import resolve_command
|
|
|
|
cmd = resolve_command("steer")
|
|
assert cmd is not None
|
|
assert cmd.name == "steer"
|
|
assert cmd.category == "Session"
|
|
assert cmd.args_hint == "<prompt>"
|
|
|
|
def test_steer_in_bypass_set(self):
|
|
"""When the agent is running, /steer MUST bypass the Level-1
|
|
base-adapter queue so it reaches the gateway runner's /steer
|
|
handler. Otherwise it would be queued as user text and only
|
|
delivered at turn end — defeating the whole point.
|
|
"""
|
|
from hermes_cli.commands import ACTIVE_SESSION_BYPASS_COMMANDS, should_bypass_active_session
|
|
|
|
assert "steer" in ACTIVE_SESSION_BYPASS_COMMANDS
|
|
assert should_bypass_active_session("steer") is True
|
|
|
|
|
|
if __name__ == "__main__": # pragma: no cover
|
|
pytest.main([__file__, "-v"])
|