"""Persistent session goals — the Ralph loop for Hermes. A goal is a free-form user objective that stays active across turns. After each turn completes, a small judge call asks an auxiliary model "is this goal satisfied by the assistant's last response?". If not, Hermes feeds a continuation prompt back into the same session and keeps working until the goal is done, turn budget is exhausted, the user pauses/clears it, or the user sends a new message (which takes priority and pauses the goal loop). State is persisted in SessionDB's ``state_meta`` table keyed by ``goal:`` so ``/resume`` picks it up. Design notes / invariants: - The continuation prompt is just a normal user message appended to the session via ``run_conversation``. No system-prompt mutation, no toolset swap — prompt caching stays intact. - Judge failures are fail-OPEN: ``continue``. A broken judge must not wedge progress; the turn budget is the backstop. - When a real user message arrives mid-loop it preempts the continuation prompt and also pauses the goal loop for that turn (we still re-judge after, so if the user's message happens to complete the goal the judge will say ``done``). - This module has zero hard dependency on ``cli.HermesCLI`` or the gateway runner — both wire the same ``GoalManager`` in. Nothing in this module touches the agent's system prompt or toolset. """ from __future__ import annotations import json import logging import re import time from dataclasses import dataclass, field, asdict from datetime import datetime, timezone from typing import Any, Dict, List, Optional, Tuple logger = logging.getLogger(__name__) # ────────────────────────────────────────────────────────────────────── # Constants & defaults # ────────────────────────────────────────────────────────────────────── DEFAULT_MAX_TURNS = 20 DEFAULT_JUDGE_TIMEOUT = 30.0 # Judge output budget. The freeform judge returns a one-line JSON verdict, but # reasoning models (deepseek-v4, qwq, etc.) burn tokens on hidden reasoning # before emitting the visible JSON — and the first /goal turn's prompt is # larger than later turns, which pushes total reply length past tight caps. # 200 tokens (the original default) reliably truncated the JSON on reasoning # models, leaving '{"done": true, "reason": "The agent successfully' and # triggering the auto-pause. 4096 covers reasoning + verdict on every model # we've live-tested; override via auxiliary.goal_judge.max_tokens for # specifically constrained setups. DEFAULT_JUDGE_MAX_TOKENS = 4096 # Cap how much of the last response + recent messages we send to the judge. _JUDGE_RESPONSE_SNIPPET_CHARS = 4000 # After this many consecutive judge *parse* failures (empty output / non-JSON), # the loop auto-pauses and points the user at the goal_judge config. API / # transport errors do NOT count toward this — those are transient. This guards # against small models (e.g. deepseek-v4-flash) that cannot follow the strict # JSON reply contract; without it the loop runs until the turn budget is # exhausted with every reply shaped like `judge returned empty response` or # `judge reply was not JSON`. DEFAULT_MAX_CONSECUTIVE_PARSE_FAILURES = 3 CONTINUATION_PROMPT_TEMPLATE = ( "[Continuing toward your standing goal]\n" "Goal: {goal}\n\n" "Continue working toward this goal. Take the next concrete step. " "If you believe the goal is complete, state so explicitly and stop. " "If you are blocked and need input from the user, say so clearly and stop." ) # Used when the goal carries a structured completion contract. The contract # block tells the agent exactly what "done" means, how to prove it, what not # to break, what's in scope, and when to stop and ask — so it targets the # verification surface instead of declaring victory loosely. CONTINUATION_PROMPT_WITH_CONTRACT_TEMPLATE = ( "[Continuing toward your standing goal]\n" "Goal: {goal}\n\n" "Completion contract:\n" "{contract_block}\n\n" "Continue working toward the outcome above. Take the next concrete step. " "Stay within the stated boundaries and do not violate the constraints. " "Before claiming the goal is done, satisfy the Verification criterion and " "show the concrete evidence (command output, file contents, test result). " "If you hit the stated stop condition or are otherwise blocked and need " "user input, say so clearly and stop." ) # Used when the user has added one or more /subgoal criteria. Surfaced # to the agent verbatim so it sees what to target on the next turn, # and surfaced to the judge so the verdict considers them too. CONTINUATION_PROMPT_WITH_SUBGOALS_TEMPLATE = ( "[Continuing toward your standing goal]\n" "Goal: {goal}\n\n" "Additional criteria the user added mid-loop:\n" "{subgoals_block}\n\n" "Continue working toward the goal AND all additional criteria. Take " "the next concrete step. If you believe the goal and every " "additional criterion are complete, state so explicitly and stop. " "If you are blocked and need input from the user, say so clearly " "and stop." ) JUDGE_SYSTEM_PROMPT = ( "You are a strict judge evaluating whether an autonomous agent has " "achieved a user's stated goal. You receive the goal text, the agent's " "most recent response, and — when present — a list of background " "processes the agent has running. Decide one of three verdicts.\n\n" "DONE — the goal is fully satisfied:\n" "- The response explicitly confirms the goal was completed, OR\n" "- The response clearly shows the final deliverable was produced, OR\n" "- The response explains the goal is unachievable / blocked / needs " "user input (treat this as DONE with reason describing the block).\n\n" "WAIT — the goal is NOT done, but the next step is to wait for async " "work to finish rather than act again. Choose this ONLY when the agent's " "progress is genuinely gated on something running on its own:\n" "- A background process listed below is still running AND the response " "shows the agent is waiting on its result (e.g. a CI poller, build, " "test run, deploy). If the process has a session id, return it in " "``wait_on_session`` — that releases when the process exits OR its " "watch_patterns trigger fires (use this for a long-lived watcher that " "signals mid-run and may never exit). Otherwise return its pid in " "``wait_on_pid`` (releases on exit only).\n" "- The agent says it is rate-limited / backing off / must wait a fixed " "period — return seconds in ``wait_for_seconds``.\n" "Picking WAIT parks the loop without burning a turn; it resumes " "automatically when the pid exits or the time elapses. Do NOT pick WAIT " "just because work remains — only when re-poking now would be pure " "busy-work because the agent can't progress until the async thing " "finishes.\n\n" "CONTINUE — not done, and there is a concrete next step the agent can " "take right now. This is the default when in doubt.\n\n" "Reply ONLY with a single JSON object on one line. Shapes:\n" '{"verdict": "done", "reason": ""}\n' '{"verdict": "continue", "reason": ""}\n' '{"verdict": "wait", "wait_on_session": "", "reason": ""}\n' '{"verdict": "wait", "wait_on_pid": , "reason": ""}\n' '{"verdict": "wait", "wait_for_seconds": , "reason": ""}\n' "The legacy shape {\"done\": , \"reason\": \"...\"} is still " "accepted (true=done, false=continue)." ) # Rendered into the judge prompt when the agent has background processes # running. Gives the judge the context it needs to decide WAIT vs CONTINUE # (and which pid to wait on) without it having to probe anything itself. JUDGE_BACKGROUND_BLOCK_TEMPLATE = ( "Background processes the agent currently has running (it may be waiting " "on one of these):\n{background_lines}\n\n" ) JUDGE_USER_PROMPT_TEMPLATE = ( "Goal:\n{goal}\n\n" "Agent's most recent response:\n{response}\n\n" "{background_block}" "Current time: {current_time}\n\n" "Is the goal satisfied — done, continue, or wait?" ) # Used when the user has added /subgoal criteria. The judge must # evaluate ALL of them being met, not just the original goal. JUDGE_USER_PROMPT_WITH_SUBGOALS_TEMPLATE = ( "Goal:\n{goal}\n\n" "Additional criteria the user added mid-loop (all must also be " "satisfied for the goal to be DONE):\n{subgoals_block}\n\n" "Agent's most recent response:\n{response}\n\n" "{background_block}" "Current time: {current_time}\n\n" "Decision: For each numbered criterion above, find concrete " "evidence in the agent's response that the criterion is " "satisfied. Do not accept generic phrases like 'all requirements " "met' or 'implying it was done' — require specific evidence (a " "file contents excerpt, an output line, a command result). If " "ANY criterion lacks specific evidence in the response, the goal " "is NOT done — return CONTINUE (or WAIT if blocked on a listed " "background process).\n\n" "Is the goal AND every additional criterion satisfied?" ) # Used when the goal carries a structured completion contract. The judge # decides DONE strictly against the Verification criterion and refuses to # accept completion when a constraint was violated. JUDGE_USER_PROMPT_WITH_CONTRACT_TEMPLATE = ( "Goal:\n{goal}\n\n" "Completion contract (the authoritative definition of done):\n" "{contract_block}\n\n" "Agent's most recent response:\n{response}\n\n" "{background_block}" "Current time: {current_time}\n\n" "Decision rules:\n" "- The goal is DONE only when the Verification criterion is satisfied AND " "the response shows concrete evidence of it (a command result, file " "contents excerpt, test/benchmark output) — not a claim like 'done' or " "'all tests pass' without evidence.\n" "- If any stated Constraint was violated, the goal is NOT done — CONTINUE.\n" "- If the response shows the agent is waiting on a listed background " "process to satisfy the Verification criterion (e.g. CI is the " "verification and it's still running), return WAIT on that process " "instead of re-poking — re-poking now would be pure busy-work.\n" "- If the response explains the work is blocked / unachievable / needs " "user input (e.g. the stated Stop condition was hit), treat it as DONE " "with the reason describing the block.\n" "- Otherwise the goal is NOT done — CONTINUE.\n\n" "Is the goal satisfied per its completion contract — done, continue, or wait?" ) # System prompt for /goal draft — turns a plain-language objective into a # structured completion contract the user can review before activating. # Adapted from Codex's "let Codex draft the goal" guidance. DRAFT_CONTRACT_SYSTEM_PROMPT = ( "You turn a user's plain-language objective into a structured completion " "contract for an autonomous coding agent. The contract has five fields:\n" "- outcome: the single end state that must be true when done\n" "- verification: the specific test / command / artifact that PROVES the " "outcome (must be concrete and checkable)\n" "- constraints: what must NOT change or regress\n" "- boundaries: which files, dirs, tools, or systems are in scope\n" "- stop_when: the condition under which the agent should stop and ask " "for human input instead of pushing on\n\n" "Infer sensible, specific values from the objective and any project " "context implied by it. Prefer concrete verification (a named test " "command, a build, a benchmark) over vague phrases. Keep each field to " "one or two sentences. If a field genuinely cannot be inferred, use an " "empty string for it.\n\n" "Reply ONLY with a single JSON object on one line:\n" '{"outcome": "...", "verification": "...", "constraints": "...", ' '"boundaries": "...", "stop_when": "..."}' ) # ────────────────────────────────────────────────────────────────────── # Completion contract # ────────────────────────────────────────────────────────────────────── # The five contract fields, in display order. Adapted from OpenAI Codex's # "strong goal" guidance: a durable objective works best when it names what # "done" means, how to prove it, what must not regress, what tools/paths are # in bounds, and when to stop and ask. A bare free-form goal (no contract) # stays fully supported — every field defaults empty and is simply omitted # from the prompts when unset. _CONTRACT_FIELDS = ("outcome", "verification", "constraints", "boundaries", "stop_when") # Human labels for rendering and for the inline `field: value` parser. _CONTRACT_LABELS = { "outcome": "Outcome", "verification": "Verification", "constraints": "Constraints", "boundaries": "Boundaries", "stop_when": "Stop when blocked", } # Inline-input aliases the user may type before a value, mapped to the # canonical field name. e.g. `verify: tests pass` or `done when: ...`. _CONTRACT_ALIASES = { "outcome": "outcome", "goal": "outcome", "done": "outcome", "done when": "outcome", "verification": "verification", "verify": "verification", "verified by": "verification", "evidence": "verification", "proof": "verification", "constraints": "constraints", "constraint": "constraints", "preserve": "constraints", "must not": "constraints", "do not change": "constraints", "boundaries": "boundaries", "boundary": "boundaries", "scope": "boundaries", "allowed": "boundaries", "files": "boundaries", "stop when": "stop_when", "stop_when": "stop_when", "blocked": "stop_when", "stop if blocked": "stop_when", "give up when": "stop_when", } @dataclass class GoalContract: """Optional structured completion contract for a goal. Each field is free-form prose the user (or :func:`draft_contract`) supplies. Empty fields are omitted everywhere — a goal with no contract behaves exactly like the original free-form goal. The contract is woven into both the continuation prompt (so the agent targets the verification surface and respects constraints) and the judge prompt (so "done" is decided against evidence, not vibes). """ outcome: str = "" verification: str = "" constraints: str = "" boundaries: str = "" stop_when: str = "" def is_empty(self) -> bool: return not any(getattr(self, f).strip() for f in _CONTRACT_FIELDS) def to_dict(self) -> Dict[str, str]: return {f: getattr(self, f) for f in _CONTRACT_FIELDS} @classmethod def from_dict(cls, data: Optional[Dict[str, Any]]) -> "GoalContract": if not isinstance(data, dict): return cls() return cls(**{f: str(data.get(f) or "").strip() for f in _CONTRACT_FIELDS}) def render_block(self) -> str: """Render non-empty contract fields as a labelled block. Empty contract → empty string (callers skip the section entirely).""" lines = [] for f in _CONTRACT_FIELDS: val = getattr(self, f).strip() if val: lines.append(f"- {_CONTRACT_LABELS[f]}: {val}") return "\n".join(lines) def parse_contract(text: str) -> Tuple[str, GoalContract]: """Split user-typed goal text into a headline + structured contract. Supports inline ``field: value`` lines so power users can type a full contract in one shot, e.g.:: Migrate auth to JWT verify: the auth test suite passes constraints: keep the public /login response shape unchanged boundaries: only touch services/auth and its tests stop when: a schema change needs product sign-off The first non-field line(s) become the goal headline; recognized ``field:`` lines populate the contract. Lines for the same field are joined. Unrecognized prefixes stay part of the headline, so a plain free-form goal with an incidental colon (``Fix bug: the parser``) is NOT mangled — only lines whose prefix matches a known alias are pulled out. Returns ``(headline, contract)``. """ if not text: return "", GoalContract() headline_parts: List[str] = [] fields: Dict[str, List[str]] = {f: [] for f in _CONTRACT_FIELDS} for raw_line in text.splitlines(): line = raw_line.strip() if not line: continue matched = False if ":" in line: prefix, _, value = line.partition(":") key = _CONTRACT_ALIASES.get(prefix.strip().lower()) if key is not None and value.strip(): fields[key].append(value.strip()) matched = True if not matched: headline_parts.append(line) headline = " ".join(headline_parts).strip() contract = GoalContract( **{f: " ".join(v).strip() for f, v in fields.items()} ) # If a headline was given but no explicit `outcome:` field, the headline # IS the outcome — don't duplicate it into the contract block (the goal # text already carries it), so leave outcome empty in that case. return headline, contract # ────────────────────────────────────────────────────────────────────── # Dataclass # ────────────────────────────────────────────────────────────────────── @dataclass class GoalState: """Serializable goal state stored per session.""" goal: str status: str = "active" # active | paused | done | cleared turns_used: int = 0 max_turns: int = DEFAULT_MAX_TURNS created_at: float = 0.0 last_turn_at: float = 0.0 last_verdict: Optional[str] = None # "done" | "continue" | "skipped" last_reason: Optional[str] = None paused_reason: Optional[str] = None # why we auto-paused (budget, etc.) consecutive_parse_failures: int = 0 # judge-output parse failures in a row # User-added criteria appended mid-loop via the /subgoal command. # When non-empty the judge prompt and continuation prompt both # include them so the agent works toward them and the judge factors # them into the verdict. Backwards-compatible: defaults to empty so # old state_meta rows load unchanged. subgoals: List[str] = field(default_factory=list) # Wait barrier: when the agent is blocked on long-running async work # (CI poller, build, test run, deploy, rate-limit cooldown) the goal loop # PARKS instead of being re-poked every turn into busy-work. Two barrier # kinds, set automatically by the judge (which now sees the live # background-process list and can return a ``wait`` verdict) or manually # via ``/goal wait``: # • ``waiting_on_pid`` — park until that process exits. # • ``waiting_on_session`` — park until that process_registry session's # OWN trigger fires: it exits, OR (if it has watch_patterns) its # pattern matches. Covers long-lived watchers/servers that signal # mid-run via a trigger and may never exit. Preferred over raw pid # when the agent set up a watch_patterns/notify_on_complete process. # • ``waiting_until`` — park until this wall-clock epoch (time backoff). # While ANY is active, ``evaluate_after_turn`` short-circuits to # should_continue=False without burning a turn or calling the judge. The # barrier auto-clears when the pid exits / the trigger fires / the deadline # passes, then the next turn resumes normal judging. Cleared by that, # ``/goal unwait``, pause, resume, or clear. Backwards-compatible: old # state_meta rows load with no barrier. waiting_on_pid: Optional[int] = None waiting_on_session: Optional[str] = None waiting_until: float = 0.0 waiting_reason: Optional[str] = None waiting_since: float = 0.0 # Optional structured completion contract (outcome / verification / # constraints / boundaries / stop_when). Empty by default; a goal with # no contract behaves exactly like the original free-form goal. contract: GoalContract = field(default_factory=GoalContract) def to_json(self) -> str: data = asdict(self) # asdict already recursed GoalContract into a plain dict. return json.dumps(data, ensure_ascii=False) @classmethod def from_json(cls, raw: str) -> "GoalState": data = json.loads(raw) raw_subgoals = data.get("subgoals") or [] subgoals: List[str] = [] if isinstance(raw_subgoals, list): subgoals = [str(s).strip() for s in raw_subgoals if str(s).strip()] return cls( goal=data.get("goal", ""), status=data.get("status", "active"), turns_used=int(data.get("turns_used", 0) or 0), max_turns=int(data.get("max_turns", DEFAULT_MAX_TURNS) or DEFAULT_MAX_TURNS), created_at=float(data.get("created_at", 0.0) or 0.0), last_turn_at=float(data.get("last_turn_at", 0.0) or 0.0), last_verdict=data.get("last_verdict"), last_reason=data.get("last_reason"), paused_reason=data.get("paused_reason"), consecutive_parse_failures=int(data.get("consecutive_parse_failures", 0) or 0), subgoals=subgoals, waiting_on_pid=(int(data["waiting_on_pid"]) if data.get("waiting_on_pid") else None), waiting_on_session=(str(data["waiting_on_session"]) if data.get("waiting_on_session") else None), waiting_until=float(data.get("waiting_until", 0.0) or 0.0), waiting_reason=data.get("waiting_reason"), waiting_since=float(data.get("waiting_since", 0.0) or 0.0), contract=GoalContract.from_dict(data.get("contract")), ) # --- contract helpers ------------------------------------------------- def has_contract(self) -> bool: return self.contract is not None and not self.contract.is_empty() # --- subgoals helpers ------------------------------------------------- def render_subgoals_block(self) -> str: """Render the subgoals as a numbered ``- N. text`` block. Empty when no subgoals exist.""" if not self.subgoals: return "" return "\n".join(f"- {i}. {text}" for i, text in enumerate(self.subgoals, start=1)) # ────────────────────────────────────────────────────────────────────── # Persistence (SessionDB state_meta) # ────────────────────────────────────────────────────────────────────── def _meta_key(session_id: str) -> str: return f"goal:{session_id}" _DB_CACHE: Dict[str, Any] = {} def _get_session_db() -> Optional[Any]: """Return a SessionDB instance for the current HERMES_HOME. SessionDB has no built-in singleton, but opening a new connection per /goal call would thrash the file. We cache one instance per ``hermes_home`` path so profile switches still pick up the right DB. Defensive against import/instantiation failures so tests and non-standard launchers can still use the GoalManager. """ try: from hermes_constants import get_hermes_home from hermes_state import SessionDB home = str(get_hermes_home()) except Exception as exc: # pragma: no cover logger.debug("GoalManager: SessionDB bootstrap failed (%s)", exc) return None cached = _DB_CACHE.get(home) if cached is not None: return cached try: db = SessionDB() except Exception as exc: # pragma: no cover logger.debug("GoalManager: SessionDB() raised (%s)", exc) return None _DB_CACHE[home] = db return db def load_goal(session_id: str) -> Optional[GoalState]: """Load the goal for a session, or None if none exists.""" if not session_id: return None db = _get_session_db() if db is None: return None try: raw = db.get_meta(_meta_key(session_id)) except Exception as exc: logger.debug("GoalManager: get_meta failed: %s", exc) return None if not raw: return None try: return GoalState.from_json(raw) except Exception as exc: logger.warning("GoalManager: could not parse stored goal for %s: %s", session_id, exc) return None def save_goal(session_id: str, state: GoalState) -> None: """Persist a goal to SessionDB. No-op if DB unavailable.""" if not session_id: return db = _get_session_db() if db is None: return try: db.set_meta(_meta_key(session_id), state.to_json()) except Exception as exc: logger.debug("GoalManager: set_meta failed: %s", exc) def clear_goal(session_id: str) -> None: """Mark a goal cleared in the DB (preserved for audit, status=cleared).""" state = load_goal(session_id) if state is None: return state.status = "cleared" save_goal(session_id, state) def migrate_goal_to_session(old_session_id: str, new_session_id: str, *, reason: str = "") -> bool: """Carry a persistent /goal from a parent session to its continuation. Context compression rotates ``session_id`` to a fresh child session, but ``load_goal`` does a flat ``goal:`` lookup with no parent-lineage walk — so an active goal silently dies at the compaction boundary (#33618). Copy the goal onto the new session and archive the old row as ``cleared`` so exactly one active goal row exists per logical conversation (avoids the "two active goals" hazard of a pure copy). Returns True when a goal was migrated, False when there was nothing to migrate or the DB was unavailable. Best-effort and never raises — a failure here must not block compression. """ if not old_session_id or not new_session_id or old_session_id == new_session_id: return False try: state = load_goal(old_session_id) if state is None or getattr(state, "status", None) == "cleared": return False # Don't clobber a goal already set on the child (e.g. a resumed # lineage that re-established its own goal). if load_goal(new_session_id) is not None: return False save_goal(new_session_id, state) # Archive the parent's row so it isn't double-counted as active. clear_goal(old_session_id) logger.debug( "GoalManager: migrated goal %s -> %s (%s)", old_session_id, new_session_id, reason or "rotation", ) return True except Exception as exc: # pragma: no cover - defensive logger.debug("GoalManager: goal migration failed: %s", exc) return False # ────────────────────────────────────────────────────────────────────── # Judge # ────────────────────────────────────────────────────────────────────── def _truncate(text: str, limit: int) -> str: if not text: return "" if len(text) <= limit: return text return text[:limit] + "… [truncated]" def _pid_alive(pid: int) -> bool: """Return True if a process with ``pid`` is currently alive. Delegates to ``gateway.status._pid_exists`` — the canonical, cross-platform, footgun-safe liveness check (psutil with a ctypes / POSIX fallback). Critically this avoids ``os.kill(pid, 0)``, which on Windows is NOT a no-op: it routes to ``CTRL_C_EVENT`` and hard-kills the target's console process group (bpo-14484). Any error resolves to False (treat unknown as dead) so a stale barrier never wedges the loop — the worst case is the goal resumes one turn early, which is safe. """ if not pid or pid <= 0: return False try: from gateway.status import _pid_exists return bool(_pid_exists(int(pid))) except Exception: pass # Last-resort fallback if gateway.status is unavailable: psutil directly. try: import psutil # type: ignore return bool(psutil.pid_exists(int(pid))) except Exception: return False def _session_waiting(session_id: str) -> bool: """Whether a goal parked on a process_registry session should stay parked. Delegates to ``process_registry.is_session_waiting`` — True while the session is running and (if it has watch_patterns) its trigger hasn't fired. Fail-safe: any import/registry error yields False (don't wait) so a stale barrier can never wedge the loop. """ if not session_id: return False try: from tools.process_registry import process_registry return bool(process_registry.is_session_waiting(session_id)) except Exception: return False _JSON_OBJECT_RE = re.compile(r"\{.*?\}", re.DOTALL) def _goal_judge_max_tokens() -> int: """Resolve auxiliary.goal_judge.max_tokens, falling back to the default. ``load_config()`` is cached on the config file's (mtime, size), so calling this once per judge turn is cheap. A non-positive or non-int value falls back to the default rather than crashing the goal loop. """ try: from hermes_cli.config import load_config cfg = load_config() value = ( (cfg.get("auxiliary") or {}) .get("goal_judge", {}) .get("max_tokens", DEFAULT_JUDGE_MAX_TOKENS) ) value = int(value) if value > 0: return value except Exception: pass return DEFAULT_JUDGE_MAX_TOKENS def _parse_judge_response(raw: str) -> Tuple[str, str, bool, Optional[Dict[str, Any]]]: """Parse the judge's reply. Fail-open on unusable output. Returns ``(verdict, reason, parse_failed, wait_directive)`` where: - ``verdict`` is ``"done"``, ``"continue"``, or ``"wait"``. - ``parse_failed`` is True when the judge returned output that couldn't be interpreted as the expected JSON verdict (empty body, prose, malformed JSON). Callers use it to auto-pause after N consecutive parse failures so a weak judge model doesn't silently burn the budget. - ``wait_directive`` is set only for ``verdict == "wait"``: a dict with ``{"pid": int}`` or ``{"seconds": int}`` (whichever the judge supplied). ``None`` otherwise. If a wait verdict carries neither a usable pid nor seconds, it is downgraded to ``continue`` (can't park on nothing). Accepts both the new ``{"verdict": ...}`` shape and the legacy ``{"done": }`` shape. """ if not raw: return "continue", "judge returned empty response", True, None text = raw.strip() # Strip markdown code fences the model may wrap JSON in. if text.startswith("```"): text = text.strip("`") # Peel off leading json/JSON/etc tag nl = text.find("\n") if nl != -1: text = text[nl + 1:] # First try: parse the whole blob. data: Optional[Dict[str, Any]] = None try: data = json.loads(text) except Exception: # Second try: pull the first JSON object out. match = _JSON_OBJECT_RE.search(text) if match: try: data = json.loads(match.group(0)) except Exception: data = None if not isinstance(data, dict): return "continue", f"judge reply was not JSON: {_truncate(raw, 200)!r}", True, None reason = str(data.get("reason") or "").strip() or "no reason provided" # Determine verdict — prefer the explicit "verdict" field, fall back to # the legacy "done" boolean. verdict_raw = data.get("verdict") if isinstance(verdict_raw, str): verdict = verdict_raw.strip().lower() else: done_val = data.get("done") if isinstance(done_val, str): done = done_val.strip().lower() in {"true", "yes", "1", "done"} else: done = bool(done_val) verdict = "done" if done else "continue" if verdict not in {"done", "continue", "wait"}: verdict = "continue" if verdict != "wait": return verdict, reason, False, None # Wait verdict: extract a concrete directive (pid or seconds). Accept a # few key spellings the model might emit. def _first_int(*keys: str) -> Optional[int]: for k in keys: v = data.get(k) if v is None: continue try: iv = int(v) if iv > 0: return iv except (TypeError, ValueError): continue return None # Prefer a session-id directive (releases on the process's own trigger — # exit OR watch-pattern match), then pid (exit only), then seconds. sess = data.get("wait_on_session") or data.get("session_id") or data.get("wait_session") if isinstance(sess, str) and sess.strip(): return "wait", reason, False, {"session_id": sess.strip()} pid = _first_int("wait_on_pid", "pid", "wait_pid") if pid is not None: return "wait", reason, False, {"pid": pid} seconds = _first_int("wait_for_seconds", "seconds", "wait_seconds") if seconds is not None: return "wait", reason, False, {"seconds": seconds} # Wait with no usable target — can't park on nothing; treat as continue. return "continue", f"{reason} (wait verdict had no target — continuing)", False, None def _render_background_block(background_processes: Optional[List[Dict[str, Any]]]) -> str: """Render the live background-process list for the judge prompt. Each entry is a ``process_registry.list_sessions()`` dict. Only RUNNING processes are worth showing (an exited one is nothing to wait on). Returns an empty string when there's nothing running, so the judge prompt is byte-identical to the no-background case (no behavior change for the common path). """ if not background_processes: return "" lines: List[str] = [] for p in background_processes: if not isinstance(p, dict): continue if p.get("status") == "exited": continue pid = p.get("pid") if not pid: continue cmd = _truncate(str(p.get("command") or "").replace("\n", " ").strip(), 120) uptime = p.get("uptime_seconds") tail = _truncate(str(p.get("output_preview") or "").replace("\n", " ").strip(), 120) sid = p.get("session_id") line = f"- pid {pid}" if sid: line += f" / session {sid}" line += f": {cmd}" if uptime is not None: line += f" (running {uptime}s)" # Surface the process's own trigger so the judge can wait on a # mid-run signal (watch-pattern) or completion, not just exit. wps = p.get("watch_patterns") if wps: hit = " [already matched]" if p.get("watch_hit") else "" line += f" | watch_patterns={wps}{hit}" elif p.get("notify_on_complete"): line += " | notify_on_complete" if tail: line += f" | recent output: {tail}" lines.append(line) if not lines: return "" return JUDGE_BACKGROUND_BLOCK_TEMPLATE.format(background_lines="\n".join(lines)) def judge_goal( goal: str, last_response: str, *, timeout: float = DEFAULT_JUDGE_TIMEOUT, subgoals: Optional[List[str]] = None, background_processes: Optional[List[Dict[str, Any]]] = None, contract: Optional[GoalContract] = None, ) -> Tuple[str, str, bool, Optional[Dict[str, Any]]]: """Ask the auxiliary model whether the goal is satisfied. Returns ``(verdict, reason, parse_failed, wait_directive)`` where verdict is ``"done"``, ``"continue"``, ``"wait"``, or ``"skipped"`` (when the judge couldn't be reached). ``wait_directive`` is set only for ``"wait"`` (``{"pid": int}`` or ``{"seconds": int}``); ``None`` otherwise. ``parse_failed`` is True only when the judge call succeeded but its output was unusable (empty or non-JSON). API/transport errors return False — they are transient and should fail-open silently. Callers use this flag to auto-pause after N consecutive parse failures (see ``DEFAULT_MAX_CONSECUTIVE_PARSE_FAILURES``). ``subgoals`` is an optional list of user-added criteria (from ``/subgoal``) factored into the verdict. ``background_processes`` is the live ``process_registry.list_sessions()`` snapshot; when the agent is waiting on one (a CI poller, build, etc.) the judge can return a ``wait`` verdict naming its pid, parking the loop instead of re-poking. ``contract`` is an optional structured completion contract; when present the judge decides DONE strictly against its Verification criterion and refuses completion when a Constraint was violated. All three are additive — a contract, subgoals, and a background-process list can coexist in one judge prompt; when none are set, behavior is identical to the original free-form judge. This is deliberately fail-open: any error returns ``("continue", ..., False, None)`` so a broken judge doesn't wedge progress — the turn budget and the consecutive-parse-failures auto-pause are the backstops. """ if not goal.strip(): return "skipped", "empty goal", False, None if not last_response.strip(): # No substantive reply this turn — almost certainly not done yet. return "continue", "empty response (nothing to evaluate)", False, None try: from agent.auxiliary_client import get_auxiliary_extra_body, get_text_auxiliary_client except Exception as exc: logger.debug("goal judge: auxiliary client import failed: %s", exc) return "continue", "auxiliary client unavailable", False, None try: client, model = get_text_auxiliary_client("goal_judge") except Exception as exc: logger.debug("goal judge: get_text_auxiliary_client failed: %s", exc) return "continue", "auxiliary client unavailable", False, None if client is None or not model: return "continue", "no auxiliary client configured", False, None # Build the prompt. Priority: contract > subgoals > plain. When both a # contract and subgoals exist, the subgoals are appended into the # contract block as extra criteria so the judge sees a single source of # truth. clean_subgoals = [s.strip() for s in (subgoals or []) if s and s.strip()] background_block = _render_background_block(background_processes) current_time = datetime.now(tz=timezone.utc).astimezone().strftime("%Y-%m-%d %H:%M:%S %Z") if contract is not None and not contract.is_empty(): contract_block = contract.render_block() if clean_subgoals: extra = "\n".join( f"- Extra criterion {i}: {text}" for i, text in enumerate(clean_subgoals, start=1) ) contract_block = f"{contract_block}\n{extra}" prompt = JUDGE_USER_PROMPT_WITH_CONTRACT_TEMPLATE.format( goal=_truncate(goal, 2000), contract_block=_truncate(contract_block, 2500), response=_truncate(last_response, _JUDGE_RESPONSE_SNIPPET_CHARS), background_block=background_block, current_time=current_time, ) elif clean_subgoals: subgoals_block = "\n".join( f"- {i}. {text}" for i, text in enumerate(clean_subgoals, start=1) ) prompt = JUDGE_USER_PROMPT_WITH_SUBGOALS_TEMPLATE.format( goal=_truncate(goal, 2000), subgoals_block=_truncate(subgoals_block, 2000), response=_truncate(last_response, _JUDGE_RESPONSE_SNIPPET_CHARS), background_block=background_block, current_time=current_time, ) else: prompt = JUDGE_USER_PROMPT_TEMPLATE.format( goal=_truncate(goal, 2000), response=_truncate(last_response, _JUDGE_RESPONSE_SNIPPET_CHARS), background_block=background_block, current_time=current_time, ) try: resp = client.chat.completions.create( model=model, messages=[ {"role": "system", "content": JUDGE_SYSTEM_PROMPT}, {"role": "user", "content": prompt}, ], temperature=0, max_tokens=_goal_judge_max_tokens(), timeout=timeout, extra_body=get_auxiliary_extra_body() or None, ) except Exception as exc: logger.info("goal judge: API call failed (%s) — falling through to continue", exc) return "continue", f"judge error: {type(exc).__name__}", False, None try: raw = resp.choices[0].message.content or "" except Exception: raw = "" verdict, reason, parse_failed, wait_directive = _parse_judge_response(raw) logger.info( "goal judge: verdict=%s reason=%s%s", verdict, _truncate(reason, 120), f" wait={wait_directive}" if wait_directive else "", ) return verdict, reason, parse_failed, wait_directive def gather_background_processes(task_id: Optional[str] = None) -> List[Dict[str, Any]]: """Return the live background-process snapshot for the goal judge. Thin, fail-safe wrapper over ``process_registry.list_sessions(task_id)``. Returns only RUNNING processes (an exited one is nothing to wait on) and never raises — any import/registry failure yields ``[]`` so the goal loop degrades to its pre-wait-barrier behavior (judge just won't see processes). The drivers (CLI + gateway) call this and pass the result into ``GoalManager.evaluate_after_turn(background_processes=...)``. """ try: from tools.process_registry import process_registry sessions = process_registry.list_sessions(task_id=task_id) or [] except Exception as exc: logger.debug("gather_background_processes failed: %s", exc) return [] return [s for s in sessions if isinstance(s, dict) and s.get("status") != "exited"] def draft_contract(objective: str, *, timeout: float = DEFAULT_JUDGE_TIMEOUT) -> Optional[GoalContract]: """Expand a plain-language objective into a structured completion contract. Uses the ``goal_judge`` auxiliary task (main-model-first, cache-safe — it is a side LLM call, not a conversation turn). Returns a populated :class:`GoalContract` on success, or ``None`` when the auxiliary client is unavailable or the model's reply can't be parsed. Callers fall back to a bare free-form goal in that case, so a missing/weak aux model never blocks setting a goal. """ objective = (objective or "").strip() if not objective: return None try: from agent.auxiliary_client import get_auxiliary_extra_body, get_text_auxiliary_client except Exception as exc: logger.debug("goal draft: auxiliary client import failed: %s", exc) return None try: client, model = get_text_auxiliary_client("goal_judge") except Exception as exc: logger.debug("goal draft: get_text_auxiliary_client failed: %s", exc) return None if client is None or not model: return None try: resp = client.chat.completions.create( model=model, messages=[ {"role": "system", "content": DRAFT_CONTRACT_SYSTEM_PROMPT}, {"role": "user", "content": f"Objective:\n{_truncate(objective, 4000)}"}, ], temperature=0, max_tokens=_goal_judge_max_tokens(), timeout=timeout, extra_body=get_auxiliary_extra_body() or None, ) except Exception as exc: logger.info("goal draft: API call failed (%s)", exc) return None try: raw = resp.choices[0].message.content or "" except Exception: raw = "" data = _extract_json_object(raw) if not isinstance(data, dict): logger.debug("goal draft: reply was not JSON: %r", _truncate(raw, 200)) return None contract = GoalContract.from_dict(data) return None if contract.is_empty() else contract def _extract_json_object(raw: str) -> Optional[Dict[str, Any]]: """Best-effort: pull the first JSON object out of a model reply. Shares the fence-stripping + first-object fallback logic used by the judge parser, but returns the dict (or None) rather than a verdict. """ if not raw: return None text = raw.strip() if text.startswith("```"): text = text.strip("`") nl = text.find("\n") if nl != -1: text = text[nl + 1:] try: data = json.loads(text) except Exception: match = _JSON_OBJECT_RE.search(text) if not match: return None try: data = json.loads(match.group(0)) except Exception: return None return data if isinstance(data, dict) else None # ────────────────────────────────────────────────────────────────────── # GoalManager — the orchestration surface CLI + gateway talk to # ────────────────────────────────────────────────────────────────────── class GoalManager: """Per-session goal state + continuation decisions. The CLI and gateway each hold one ``GoalManager`` per live session. Methods: - ``set(goal)`` — start a new standing goal. - ``clear()`` — remove the active goal. - ``pause()`` / ``resume()`` — explicit user controls. - ``status()`` — printable one-liner. - ``evaluate_after_turn(last_response)`` — call the judge, update state, and return a decision dict the caller uses to drive the next turn. - ``next_continuation_prompt()`` — the canonical user-role message to feed back into ``run_conversation``. """ def __init__(self, session_id: str, *, default_max_turns: int = DEFAULT_MAX_TURNS): self.session_id = session_id self.default_max_turns = int(default_max_turns or DEFAULT_MAX_TURNS) self._state: Optional[GoalState] = load_goal(session_id) # --- introspection ------------------------------------------------ @property def state(self) -> Optional[GoalState]: return self._state def is_active(self) -> bool: return self._state is not None and self._state.status == "active" def has_goal(self) -> bool: return self._state is not None and self._state.status in {"active", "paused"} def has_contract(self) -> bool: return self._state is not None and self._state.has_contract() def status_line(self) -> str: s = self._state if s is None or s.status in {"cleared",}: return "No active goal. Set one with /goal ." turns = f"{s.turns_used}/{s.max_turns} turns" sub = f", {len(s.subgoals)} subgoal{'s' if len(s.subgoals) != 1 else ''}" if s.subgoals else "" con = ", contract" if self.has_contract() else "" meta = f"{turns}{sub}{con}" if s.status == "active": if s.waiting_on_session and _session_waiting(s.waiting_on_session): wr = s.waiting_reason or f"session {s.waiting_on_session}" return f"⏳ Goal (parked on {wr}, {meta}): {s.goal}" if s.waiting_on_pid and _pid_alive(s.waiting_on_pid): wr = s.waiting_reason or f"pid {s.waiting_on_pid}" return f"⏳ Goal (parked on {wr}, {meta}): {s.goal}" if s.waiting_until and time.time() < s.waiting_until: remaining = int(s.waiting_until - time.time()) wr = s.waiting_reason or f"{remaining}s" return f"⏳ Goal (parked {remaining}s — {wr}, {meta}): {s.goal}" return f"⊙ Goal (active, {meta}): {s.goal}" if s.status == "paused": extra = f" — {s.paused_reason}" if s.paused_reason else "" return f"⏸ Goal (paused, {meta}{extra}): {s.goal}" if s.status == "done": return f"✓ Goal done ({meta}): {s.goal}" return f"Goal ({s.status}, {meta}): {s.goal}" # --- mutation ----------------------------------------------------- def set(self, goal: str, *, max_turns: Optional[int] = None, contract: Optional[GoalContract] = None) -> GoalState: goal = (goal or "").strip() if not goal: raise ValueError("goal text is empty") state = GoalState( goal=goal, status="active", turns_used=0, max_turns=int(max_turns) if max_turns else self.default_max_turns, created_at=time.time(), last_turn_at=0.0, contract=contract if contract is not None else GoalContract(), ) self._state = state save_goal(self.session_id, state) return state def set_contract(self, contract: GoalContract) -> Optional[GoalState]: """Attach or replace the completion contract on the active goal. Returns the updated state, or None when there is no goal to attach to. """ if self._state is None: return None self._state.contract = contract or GoalContract() save_goal(self.session_id, self._state) return self._state def pause(self, reason: str = "user-paused") -> Optional[GoalState]: if not self._state: return None self._state.status = "paused" self._state.paused_reason = reason # A wait barrier is meaningless once paused — drop it. self._state.waiting_on_pid = None self._state.waiting_on_session = None self._state.waiting_until = 0.0 self._state.waiting_reason = None self._state.waiting_since = 0.0 save_goal(self.session_id, self._state) return self._state def resume(self, *, reset_budget: bool = True) -> Optional[GoalState]: if not self._state: return None self._state.status = "active" self._state.paused_reason = None # Resuming starts fresh — clear any stale barrier. self._state.waiting_on_pid = None self._state.waiting_on_session = None self._state.waiting_until = 0.0 self._state.waiting_reason = None self._state.waiting_since = 0.0 if reset_budget: self._state.turns_used = 0 save_goal(self.session_id, self._state) return self._state def clear(self) -> None: if self._state is None: return self._state.status = "cleared" save_goal(self.session_id, self._state) self._state = None def mark_done(self, reason: str) -> None: if not self._state: return self._state.status = "done" self._state.last_verdict = "done" self._state.last_reason = reason save_goal(self.session_id, self._state) # --- /subgoal user controls --------------------------------------- def add_subgoal(self, text: str) -> str: """Append a user-added criterion to the active goal. Requires ``has_goal()``; raises ``RuntimeError`` otherwise. Returns the cleaned text so the caller can show it back to the user. """ if self._state is None or not self.has_goal(): raise RuntimeError("no active goal") text = (text or "").strip() if not text: raise ValueError("subgoal text is empty") self._state.subgoals.append(text) save_goal(self.session_id, self._state) return text def remove_subgoal(self, index_1based: int) -> str: """Remove a subgoal by 1-based index. Returns the removed text.""" if self._state is None or not self.has_goal(): raise RuntimeError("no active goal") idx = int(index_1based) - 1 if idx < 0 or idx >= len(self._state.subgoals): raise IndexError( f"index out of range (1..{len(self._state.subgoals)})" ) removed = self._state.subgoals.pop(idx) save_goal(self.session_id, self._state) return removed def clear_subgoals(self) -> int: """Wipe all subgoals. Returns the previous count.""" if self._state is None or not self.has_goal(): raise RuntimeError("no active goal") prev = len(self._state.subgoals) self._state.subgoals = [] save_goal(self.session_id, self._state) return prev def render_subgoals(self) -> str: """Public helper for the /subgoal slash command.""" if self._state is None: return "(no active goal)" if not self._state.subgoals: return "(no subgoals — use /subgoal to add criteria)" return self._state.render_subgoals_block() # --- /goal wait barrier ------------------------------------------- def wait_on(self, pid: int, reason: str = "") -> GoalState: """Park the goal loop on a background process PID. While the PID is alive, ``evaluate_after_turn`` returns ``should_continue=False`` without burning a turn or calling the judge — the loop quiesces instead of re-poking the agent into busy work. The barrier auto-clears when the process exits. Requires an active goal. For a process with a watch_patterns/notify_on_complete trigger, prefer ``wait_on_session`` so a mid-run trigger (not just exit) releases the barrier. """ if self._state is None or self._state.status != "active": raise RuntimeError("no active goal to park") pid = int(pid) if pid <= 0: raise ValueError("pid must be a positive integer") self._state.waiting_on_pid = pid self._state.waiting_on_session = None self._state.waiting_until = 0.0 self._state.waiting_reason = (reason or "").strip() or None self._state.waiting_since = time.time() save_goal(self.session_id, self._state) return self._state def wait_on_session(self, session_id: str, reason: str = "") -> GoalState: """Park the goal loop on a process_registry session's OWN trigger. Unlike ``wait_on`` (which releases only on PID exit), this releases when the session's trigger fires: it exits, OR — if it was started with ``watch_patterns`` — its pattern matches. This is the right barrier for a long-lived watcher/server/poller that signals mid-run and may never exit. Requires an active goal. """ if self._state is None or self._state.status != "active": raise RuntimeError("no active goal to park") session_id = str(session_id or "").strip() if not session_id: raise ValueError("session_id must be a non-empty string") self._state.waiting_on_session = session_id self._state.waiting_on_pid = None self._state.waiting_until = 0.0 self._state.waiting_reason = (reason or "").strip() or None self._state.waiting_since = time.time() save_goal(self.session_id, self._state) return self._state def wait_for_seconds(self, seconds: int, reason: str = "") -> GoalState: """Park the goal loop until ``seconds`` from now have elapsed. Time-based counterpart to ``wait_on`` — for backoff / cooldown waits where there's no process to track (e.g. the agent is rate-limited). The barrier auto-clears once the deadline passes. Requires an active goal. """ if self._state is None or self._state.status != "active": raise RuntimeError("no active goal to park") seconds = int(seconds) if seconds <= 0: raise ValueError("seconds must be a positive integer") self._state.waiting_on_pid = None self._state.waiting_on_session = None self._state.waiting_until = time.time() + seconds self._state.waiting_reason = (reason or "").strip() or None self._state.waiting_since = time.time() save_goal(self.session_id, self._state) return self._state def stop_waiting(self) -> bool: """Clear any active wait barrier (pid / session / time). Returns True if one was cleared.""" if self._state is None: return False if ( self._state.waiting_on_pid is None and self._state.waiting_on_session is None and not self._state.waiting_until ): return False self._state.waiting_on_pid = None self._state.waiting_on_session = None self._state.waiting_until = 0.0 self._state.waiting_reason = None self._state.waiting_since = 0.0 save_goal(self.session_id, self._state) return True def is_waiting(self) -> bool: """True iff a barrier is set AND not yet satisfied. Session barrier: active until the process exits or its watch-pattern trigger fires. Pid barrier: active while the process is alive. Time barrier: active until the deadline passes. Side effect: a satisfied barrier is cleared here (lazy auto-clear) so the next evaluation resumes normal judging. """ s = self._state if s is None: return False if s.waiting_on_session is not None: if _session_waiting(s.waiting_on_session): return True self.stop_waiting() # session exited or trigger fired return False if s.waiting_on_pid is not None: if _pid_alive(s.waiting_on_pid): return True self.stop_waiting() # process gone return False if s.waiting_until: if time.time() < s.waiting_until: return True self.stop_waiting() # deadline passed return False return False # --- the main entry point called after every turn ----------------- def evaluate_after_turn( self, last_response: str, *, user_initiated: bool = True, background_processes: Optional[List[Dict[str, Any]]] = None, ) -> Dict[str, Any]: """Run the judge and update state. Return a decision dict. ``user_initiated`` distinguishes a real user prompt (True) from a continuation prompt we fed ourselves (False). Both increment ``turns_used`` because both consume model budget. ``background_processes`` is the live ``process_registry.list_sessions()`` snapshot for this session. It's handed to the judge so it can decide to WAIT on an in-flight process (CI poller, build, ...) instead of re-poking the agent — the automatic counterpart to ``/goal wait``. Decision keys: - ``status``: current goal status after update - ``should_continue``: bool — caller should fire another turn - ``continuation_prompt``: str or None - ``verdict``: "done" | "continue" | "wait" | "skipped" | "inactive" - ``reason``: str - ``message``: user-visible one-liner to print/send """ state = self._state if state is None or state.status != "active": return { "status": state.status if state else None, "should_continue": False, "continuation_prompt": None, "verdict": "inactive", "reason": "no active goal", "message": "", } # Wait barrier: if the loop is parked (on a live process OR a time # deadline that hasn't passed), quiesce — do NOT burn a turn or call # the judge. Resumes automatically once the barrier clears. if self.is_waiting(): if state.waiting_on_session is not None: tgt = f"session {state.waiting_on_session}" elif state.waiting_on_pid is not None: tgt = f"pid {state.waiting_on_pid}" else: remaining = max(0, int(state.waiting_until - time.time())) tgt = f"{remaining}s remaining" reason = state.waiting_reason or tgt return { "status": "active", "should_continue": False, "continuation_prompt": None, "verdict": "waiting", "reason": reason, "message": f"⏳ Goal parked — waiting on {tgt}: {reason}", } # Count the turn that just finished. state.turns_used += 1 state.last_turn_at = time.time() verdict, reason, parse_failed, wait_directive = judge_goal( state.goal, last_response, subgoals=state.subgoals or None, background_processes=background_processes, contract=state.contract if state.has_contract() else None, ) state.last_verdict = verdict state.last_reason = reason # Track consecutive judge parse failures. Reset on any usable reply, # including API / transport errors (parse_failed=False) so a flaky # network doesn't trip the auto-pause meant for bad judge models. if parse_failed: state.consecutive_parse_failures += 1 else: state.consecutive_parse_failures = 0 # WAIT verdict: the judge decided the agent is blocked on async work # and re-poking now would be busy-work. Set the barrier and park — # the turn we just counted stands (the judge call happened), but no # continuation fires. The loop resumes automatically when the pid # exits or the deadline passes (next evaluate_after_turn falls through # the is_waiting() short-circuit once the barrier clears). if verdict == "wait" and wait_directive: if wait_directive.get("session_id"): self.wait_on_session(str(wait_directive["session_id"]), reason=reason) tgt = f"session {wait_directive['session_id']}" elif wait_directive.get("pid"): self.wait_on(int(wait_directive["pid"]), reason=reason) tgt = f"pid {wait_directive['pid']}" else: self.wait_for_seconds(int(wait_directive["seconds"]), reason=reason) tgt = f"{wait_directive['seconds']}s" return { "status": "active", "should_continue": False, "continuation_prompt": None, "verdict": "wait", "reason": reason, "message": f"⏳ Goal parked (judge) — waiting on {tgt}: {reason}", } if verdict == "done": state.status = "done" save_goal(self.session_id, state) return { "status": "done", "should_continue": False, "continuation_prompt": None, "verdict": "done", "reason": reason, "message": f"✓ Goal achieved: {reason}", } # Auto-pause when the judge model can't produce the expected JSON # verdict N turns in a row. Points the user at the goal_judge config # so they can route this side task to a model that follows the # contract (e.g. google/gemini-3-flash-preview). Without this guard, # weak judge models burn the entire turn budget returning prose or # empty strings. if state.consecutive_parse_failures >= DEFAULT_MAX_CONSECUTIVE_PARSE_FAILURES: state.status = "paused" state.paused_reason = ( f"judge model returned unparseable output {state.consecutive_parse_failures} turns in a row" ) save_goal(self.session_id, state) return { "status": "paused", "should_continue": False, "continuation_prompt": None, "verdict": "continue", "reason": reason, "message": ( f"⏸ Goal paused — the judge model ({state.consecutive_parse_failures} turns) " "isn't returning the required JSON verdict. Route the judge to a stricter " "model in ~/.hermes/config.yaml:\n" " auxiliary:\n" " goal_judge:\n" " provider: openrouter\n" " model: google/gemini-3-flash-preview\n" "Then /goal resume to continue." ), } if state.turns_used >= state.max_turns: state.status = "paused" state.paused_reason = f"turn budget exhausted ({state.turns_used}/{state.max_turns})" save_goal(self.session_id, state) return { "status": "paused", "should_continue": False, "continuation_prompt": None, "verdict": "continue", "reason": reason, "message": ( f"⏸ Goal paused — {state.turns_used}/{state.max_turns} turns used. " "Use /goal resume to keep going, or /goal clear to stop." ), } save_goal(self.session_id, state) return { "status": "active", "should_continue": True, "continuation_prompt": self.next_continuation_prompt(), "verdict": "continue", "reason": reason, "message": ( f"↻ Continuing toward goal ({state.turns_used}/{state.max_turns}): {reason}" ), } def next_continuation_prompt(self) -> Optional[str]: if not self._state or self._state.status != "active": return None # Contract takes priority: it carries the verification surface and # constraints the agent must target. Subgoals fold in as extra # criteria appended to the contract block. if self._state.has_contract(): contract_block = self._state.contract.render_block() if self._state.subgoals: extra = "\n".join( f"- Extra criterion {i}: {text}" for i, text in enumerate(self._state.subgoals, start=1) ) contract_block = f"{contract_block}\n{extra}" return CONTINUATION_PROMPT_WITH_CONTRACT_TEMPLATE.format( goal=self._state.goal, contract_block=contract_block, ) if self._state.subgoals: return CONTINUATION_PROMPT_WITH_SUBGOALS_TEMPLATE.format( goal=self._state.goal, subgoals_block=self._state.render_subgoals_block(), ) return CONTINUATION_PROMPT_TEMPLATE.format(goal=self._state.goal) def render_contract(self) -> str: """Public helper for the /goal show + /goal draft slash commands.""" if self._state is None: return "(no active goal)" if not self._state.has_contract(): return "(no completion contract — set one with /goal draft or inline field: value lines)" return self._state.contract.render_block() # ────────────────────────────────────────────────────────────────────── # Kanban worker goal loop # ────────────────────────────────────────────────────────────────────── # Continuation prompt fed back to a kanban goal-mode worker that has not # yet completed/blocked its task. The card's own acceptance criteria are # the goal — the worker already has the full task body in its first turn, # so we keep this short and point it back at the lifecycle contract. KANBAN_GOAL_CONTINUATION_TEMPLATE = ( "[Continuing toward this kanban task — judge says it is not done yet]\n" "Reason: {reason}\n\n" "Take the next concrete step toward completing the task. When the work " "is genuinely finished, call kanban_complete with a summary. If you are " "blocked and need human input, call kanban_block with a reason. Do not " "stop without calling one of them." ) # Fed when the judge believes the work is done but the worker never called # kanban_complete / kanban_block. One explicit nudge to terminate the task # the right way before the loop gives up. KANBAN_GOAL_FINALIZE_TEMPLATE = ( "[The work looks complete, but the task is still open]\n" "Reason: {reason}\n\n" "If the task is genuinely done, call kanban_complete now with a short " "summary of what you did. If something still blocks completion, call " "kanban_block with the reason instead." ) def run_kanban_goal_loop( *, task_id: str, goal_text: str, run_turn, task_status_fn, block_fn, max_turns: int = DEFAULT_MAX_TURNS, first_response: str = "", log=None, ) -> Dict[str, Any]: """Drive a kanban worker through a Ralph-style goal loop. The dispatcher spawns a goal-mode worker exactly like a normal worker (``hermes -p chat -q "work kanban task "``). The worker's first turn has already run by the time this is called; ``first_response`` is that turn's reply. From here we: 1. Check whether the worker already terminated the task (called ``kanban_complete`` / ``kanban_block``). If so, stop — nothing to do. 2. Otherwise judge the latest response against ``goal_text`` (the card's title + body). ``continue`` → feed a continuation prompt and run another turn IN THE SAME SESSION via ``run_turn``. ``done`` but the task is still open → one explicit "call kanban_complete" nudge. 3. When the turn budget is exhausted and the worker still hasn't terminated the task, ``block_fn`` is invoked so the card lands in a sticky ``blocked`` state for human review (NOT a silent exit). This function performs NO SessionDB persistence — a worker process is ephemeral, so the turn budget lives in a local counter. It is fully decoupled from the CLI for testability: callers inject ``run_turn`` (str -> str), ``task_status_fn`` (() -> str|None), and ``block_fn`` (reason: str -> None). Returns a decision dict: ``{"outcome", "turns_used", "reason"}`` where outcome is one of ``"completed_by_worker"``, ``"blocked_budget"``, ``"blocked_by_worker"``, or ``"stopped"``. """ def _log(msg: str) -> None: if log is not None: try: log(msg) except Exception: pass max_turns = int(max_turns or DEFAULT_MAX_TURNS) if max_turns < 1: max_turns = DEFAULT_MAX_TURNS last_response = first_response or "" # The first turn already consumed one unit of budget. turns_used = 1 nudged_to_finalize = False while True: # Did the worker terminate the task itself this turn? try: status = task_status_fn() except Exception as exc: _log(f"kanban goal loop: status check failed ({exc}); stopping") return {"outcome": "stopped", "turns_used": turns_used, "reason": "status check failed"} if status == "done": _log(f"kanban goal loop: task {task_id} completed by worker after {turns_used} turn(s)") return {"outcome": "completed_by_worker", "turns_used": turns_used, "reason": "worker completed the task"} if status == "blocked": _log(f"kanban goal loop: task {task_id} blocked by worker after {turns_used} turn(s)") return {"outcome": "blocked_by_worker", "turns_used": turns_used, "reason": "worker blocked the task"} if status not in ("running", "ready"): # Reclaimed / archived / unexpected — let the dispatcher own it. _log(f"kanban goal loop: task {task_id} status={status!r}; stopping") return {"outcome": "stopped", "turns_used": turns_used, "reason": f"status={status}"} # Still open — judge whether the latest response satisfies the card. # The kanban worker loop has no wait-barrier concept (workers finish # via kanban_complete / kanban_block, not by parking), so a WAIT # verdict is treated as CONTINUE here. verdict, reason, _parse_failed, _wait = judge_goal(goal_text, last_response) if verdict == "wait": verdict = "continue" _log(f"kanban goal loop: turn {turns_used}/{max_turns} verdict={verdict} reason={_truncate(reason, 120)}") if verdict == "done": if nudged_to_finalize: # Already asked once to call kanban_complete and it still # didn't — block for review rather than spin. _log(f"kanban goal loop: task {task_id} judged done but worker won't finalize; blocking") try: block_fn( f"Goal-mode worker's output looked complete but it never " f"called kanban_complete after a finalize nudge ({reason})." ) except Exception as exc: _log(f"kanban goal loop: block_fn failed ({exc})") return {"outcome": "blocked_budget", "turns_used": turns_used, "reason": "judged done, never finalized"} prompt = KANBAN_GOAL_FINALIZE_TEMPLATE.format(reason=_truncate(reason, 400)) nudged_to_finalize = True else: prompt = KANBAN_GOAL_CONTINUATION_TEMPLATE.format(reason=_truncate(reason, 400)) # Budget check BEFORE spending another turn. if turns_used >= max_turns: _log(f"kanban goal loop: task {task_id} exhausted {turns_used}/{max_turns} turns; blocking") try: block_fn( f"Goal-mode worker exhausted its turn budget " f"({turns_used}/{max_turns}) without completing the task. " f"Last judge verdict: {_truncate(reason, 300)}" ) except Exception as exc: _log(f"kanban goal loop: block_fn failed ({exc})") return {"outcome": "blocked_budget", "turns_used": turns_used, "reason": "turn budget exhausted"} # Run another turn in the same session. try: last_response = run_turn(prompt) or "" except Exception as exc: _log(f"kanban goal loop: run_turn failed ({exc}); stopping") return {"outcome": "stopped", "turns_used": turns_used, "reason": f"run_turn error: {type(exc).__name__}"} turns_used += 1 __all__ = [ "GoalState", "GoalContract", "GoalManager", "parse_contract", "draft_contract", "CONTINUATION_PROMPT_TEMPLATE", "CONTINUATION_PROMPT_WITH_SUBGOALS_TEMPLATE", "CONTINUATION_PROMPT_WITH_CONTRACT_TEMPLATE", "JUDGE_USER_PROMPT_TEMPLATE", "JUDGE_USER_PROMPT_WITH_SUBGOALS_TEMPLATE", "JUDGE_USER_PROMPT_WITH_CONTRACT_TEMPLATE", "DRAFT_CONTRACT_SYSTEM_PROMPT", "KANBAN_GOAL_CONTINUATION_TEMPLATE", "KANBAN_GOAL_FINALIZE_TEMPLATE", "DEFAULT_MAX_TURNS", "load_goal", "save_goal", "clear_goal", "migrate_goal_to_session", "judge_goal", "run_kanban_goal_loop", ]