feat: /goal — persistent cross-turn goals (Ralph loop) (#18262)
Add a standing-goal slash command that keeps Hermes working toward a user-stated objective across turns until it is achieved, paused, or the turn budget runs out. Our take on the Ralph loop — cf. Codex CLI 0.128.0's /goal. After each turn, a lightweight auxiliary-model judge call asks 'is this goal satisfied by the assistant's last response?'. If not, and we're under the turn budget (default 20), Hermes feeds a continuation prompt back into the same session as a normal user message. Any real user message preempts the continuation loop automatically. Judge failures fail OPEN (continue) so a flaky judge never wedges progress — the turn budget is the real backstop. ### Commands - `/goal <text>` — set a standing goal (kicks off the first turn) - `/goal` or `/goal status` — show current state - `/goal pause` — pause the continuation loop - `/goal resume` — resume (resets turn counter) - `/goal clear` — drop the goal Works on both CLI and gateway platforms via the central CommandDef registry. ### Design invariants preserved - **Prompt cache**: continuation prompts are regular user-role messages appended to history. No system-prompt mutation, no toolset swap. - **Role alternation**: continuation is a user turn, never injected mid-tool-loop. - **Session persistence**: goal state lives in SessionDB.state_meta keyed by `goal:<session_id>`, so `/resume` picks it up. - **Mid-run safety**: on the gateway, `/goal status|pause|clear` are allowed mid-run (control-plane only); setting a new goal requires `/stop` first so we don't race a second continuation prompt against the current turn. ### Files - `hermes_cli/goals.py` (new, 380 lines) — GoalManager + judge + state - `hermes_cli/commands.py` — CommandDef entry - `hermes_cli/config.py` — `goals.max_turns` default - `hermes_cli/web_server.py` — dashboard category merge - `cli.py` — /goal handler + post-turn continuation hook in process_loop - `gateway/run.py` — /goal handler + post-turn continuation hook wrapping _handle_message_with_agent - `tests/hermes_cli/test_goals.py` (new, 26 tests) — judge parsing, fail-open semantics, lifecycle, persistence, budget exhaustion - `website/docs/reference/slash-commands.md` — docs entry
This commit is contained in:
Teknium
173
cli.py
173
cli.py
@@ -6540,6 +6540,8 @@ class HermesCLI:
|
||||
# No active run — treat as a normal next-turn message.
|
||||
self._pending_input.put(payload)
|
||||
_cprint(f" No agent running; queued as next turn: {payload[:80]}{'...' if len(payload) > 80 else ''}")
|
||||
elif canonical == "goal":
|
||||
self._handle_goal_command(cmd_original)
|
||||
elif canonical == "skin":
|
||||
self._handle_skin_command(cmd_original)
|
||||
elif canonical == "voice":
|
||||
@@ -7020,6 +7022,166 @@ class HermesCLI:
|
||||
print(" status Show current browser mode")
|
||||
print()
|
||||
|
||||
# ────────────────────────────────────────────────────────────────
|
||||
# /goal — persistent cross-turn goals (Ralph-style loop)
|
||||
# ────────────────────────────────────────────────────────────────
|
||||
def _get_goal_manager(self):
|
||||
"""Return the GoalManager bound to the current session_id.
|
||||
|
||||
Cached on ``self._goal_manager`` and rebound lazily when
|
||||
``session_id`` changes (e.g. after /new or a compression-driven
|
||||
session split).
|
||||
"""
|
||||
try:
|
||||
from hermes_cli.goals import GoalManager
|
||||
from hermes_cli.config import load_config
|
||||
except Exception as exc:
|
||||
logging.debug("goal manager unavailable: %s", exc)
|
||||
return None
|
||||
|
||||
sid = getattr(self, "session_id", None) or ""
|
||||
if not sid:
|
||||
return None
|
||||
|
||||
existing = getattr(self, "_goal_manager", None)
|
||||
if existing is not None and getattr(existing, "session_id", None) == sid:
|
||||
return existing
|
||||
|
||||
try:
|
||||
cfg = load_config() or {}
|
||||
goals_cfg = cfg.get("goals") or {}
|
||||
max_turns = int(goals_cfg.get("max_turns", 20) or 20)
|
||||
except Exception:
|
||||
max_turns = 20
|
||||
|
||||
mgr = GoalManager(session_id=sid, default_max_turns=max_turns)
|
||||
self._goal_manager = mgr
|
||||
return mgr
|
||||
|
||||
def _handle_goal_command(self, cmd: str) -> None:
|
||||
"""Dispatch /goal subcommands: set / status / pause / resume / clear."""
|
||||
parts = (cmd or "").strip().split(None, 1)
|
||||
arg = parts[1].strip() if len(parts) > 1 else ""
|
||||
|
||||
mgr = self._get_goal_manager()
|
||||
if mgr is None:
|
||||
_cprint(f" {_DIM}Goals unavailable (no active session).{_RST}")
|
||||
return
|
||||
|
||||
lower = arg.lower()
|
||||
|
||||
# Bare /goal or /goal status → show current state
|
||||
if not arg or lower == "status":
|
||||
_cprint(f" {mgr.status_line()}")
|
||||
return
|
||||
|
||||
if lower == "pause":
|
||||
state = mgr.pause(reason="user-paused")
|
||||
if state is None:
|
||||
_cprint(f" {_DIM}No goal set.{_RST}")
|
||||
else:
|
||||
_cprint(f" ⏸ Goal paused: {state.goal}")
|
||||
return
|
||||
|
||||
if lower == "resume":
|
||||
state = mgr.resume()
|
||||
if state is None:
|
||||
_cprint(f" {_DIM}No goal to resume.{_RST}")
|
||||
else:
|
||||
_cprint(f" ▶ Goal resumed: {state.goal}")
|
||||
_cprint(
|
||||
f" {_DIM}Send any message (or press Enter on an empty prompt "
|
||||
f"is a no-op; type 'continue' to kick it off).{_RST}"
|
||||
)
|
||||
return
|
||||
|
||||
if lower in ("clear", "stop", "done"):
|
||||
had = mgr.has_goal()
|
||||
mgr.clear()
|
||||
if had:
|
||||
_cprint(" ✓ Goal cleared.")
|
||||
else:
|
||||
_cprint(f" {_DIM}No active goal.{_RST}")
|
||||
return
|
||||
|
||||
# Otherwise treat the arg as the goal text.
|
||||
try:
|
||||
state = mgr.set(arg)
|
||||
except ValueError as exc:
|
||||
_cprint(f" Invalid goal: {exc}")
|
||||
return
|
||||
|
||||
_cprint(f" ⊙ Goal set ({state.max_turns}-turn budget): {state.goal}")
|
||||
_cprint(
|
||||
f" {_DIM}After each turn, a judge model will check if the goal is done. "
|
||||
f"Hermes keeps working until it is, you pause/clear it, or the budget is "
|
||||
f"exhausted. Use /goal status, /goal pause, /goal resume, /goal clear.{_RST}"
|
||||
)
|
||||
# Kick the loop off immediately so the user doesn't have to send a
|
||||
# separate message after setting the goal.
|
||||
try:
|
||||
self._pending_input.put(state.goal)
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
def _maybe_continue_goal_after_turn(self) -> None:
|
||||
"""Hook run after every CLI turn. Judges + maybe re-queues.
|
||||
|
||||
Safe to call when no goal is set — returns quickly.
|
||||
|
||||
Preemption is automatic: if a real user message is already in
|
||||
``_pending_input`` we skip judging (the user's new input takes
|
||||
priority and we'll re-judge after that turn). If judge says done,
|
||||
mark it done and tell the user. If judge says continue and we're
|
||||
under budget, push the continuation prompt onto the queue.
|
||||
"""
|
||||
mgr = self._get_goal_manager()
|
||||
if mgr is None or not mgr.is_active():
|
||||
return
|
||||
|
||||
# If a real user message is already queued, don't inject a
|
||||
# continuation prompt on top — let the user's turn go first.
|
||||
try:
|
||||
if getattr(self, "_pending_input", None) is not None \
|
||||
and not self._pending_input.empty():
|
||||
return
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
# Extract the agent's final response for this turn.
|
||||
last_response = ""
|
||||
try:
|
||||
hist = self.conversation_history or []
|
||||
for msg in reversed(hist):
|
||||
if msg.get("role") == "assistant":
|
||||
content = msg.get("content", "")
|
||||
if isinstance(content, list):
|
||||
# Multimodal content — flatten text parts.
|
||||
parts = [
|
||||
p.get("text", "")
|
||||
for p in content
|
||||
if isinstance(p, dict) and p.get("type") in ("text", "output_text")
|
||||
]
|
||||
last_response = "\n".join(t for t in parts if t)
|
||||
else:
|
||||
last_response = str(content or "")
|
||||
break
|
||||
except Exception:
|
||||
last_response = ""
|
||||
|
||||
decision = mgr.evaluate_after_turn(last_response, user_initiated=True)
|
||||
msg = decision.get("message") or ""
|
||||
if msg:
|
||||
_cprint(f" {msg}")
|
||||
|
||||
if decision.get("should_continue"):
|
||||
prompt = decision.get("continuation_prompt")
|
||||
if prompt:
|
||||
try:
|
||||
self._pending_input.put(prompt)
|
||||
except Exception as exc:
|
||||
logging.debug("goal continuation enqueue failed: %s", exc)
|
||||
|
||||
def _handle_skin_command(self, cmd: str):
|
||||
"""Handle /skin [name] — show or change the display skin."""
|
||||
try:
|
||||
@@ -11358,6 +11520,17 @@ class HermesCLI:
|
||||
|
||||
app.invalidate() # Refresh status line
|
||||
|
||||
# Goal continuation: if a standing goal is active, ask
|
||||
# the judge whether the turn satisfied it. If not, and
|
||||
# there's no real user message already queued, push the
|
||||
# continuation prompt back into _pending_input so the
|
||||
# next loop iteration picks it up naturally (and any
|
||||
# user input that arrives in between still preempts).
|
||||
try:
|
||||
self._maybe_continue_goal_after_turn()
|
||||
except Exception as _goal_exc:
|
||||
logging.debug("goal continuation hook failed: %s", _goal_exc)
|
||||
|
||||
# Continuous voice: auto-restart recording after agent responds.
|
||||
# Dispatch to a daemon thread so play_beep (sd.wait) and
|
||||
# AudioRecorder.start (lock acquire) never block process_loop —
|
||||
|
||||
Reference in New Issue
Block a user