"""Pure tool-call loop guardrail primitives. The controller in this module is intentionally side-effect free: it tracks per-turn tool-call observations and returns decisions. Runtime code owns whether those decisions become warning guidance, synthetic tool results, or controlled turn halts. """ from __future__ import annotations import hashlib import json from dataclasses import dataclass, field from typing import Any, Mapping from utils import safe_json_loads from agent.tool_result_classification import file_mutation_result_landed IDEMPOTENT_TOOL_NAMES = frozenset( { "read_file", "search_files", "web_search", "web_extract", "session_search", "browser_snapshot", "browser_console", "browser_get_images", "mcp_filesystem_read_file", "mcp_filesystem_read_text_file", "mcp_filesystem_read_multiple_files", "mcp_filesystem_list_directory", "mcp_filesystem_list_directory_with_sizes", "mcp_filesystem_directory_tree", "mcp_filesystem_get_file_info", "mcp_filesystem_search_files", } ) MUTATING_TOOL_NAMES = frozenset( { "terminal", "execute_code", "write_file", "patch", "todo", "memory", "skill_manage", "browser_click", "browser_type", "browser_press", "browser_scroll", "browser_navigate", "send_message", "cronjob", "delegate_task", "process", } ) @dataclass(frozen=True) class ToolCallGuardrailConfig: """Thresholds for per-turn tool-call loop detection. Warnings are enabled by default and never prevent tool execution. Hard stops are explicit opt-in so interactive CLI/TUI sessions get a gentle nudge unless the user enables circuit-breaker behavior in config.yaml. """ warnings_enabled: bool = True hard_stop_enabled: bool = False exact_failure_warn_after: int = 2 exact_failure_block_after: int = 5 same_tool_failure_warn_after: int = 3 same_tool_failure_halt_after: int = 8 no_progress_warn_after: int = 2 no_progress_block_after: int = 5 idempotent_tools: frozenset[str] = field(default_factory=lambda: IDEMPOTENT_TOOL_NAMES) mutating_tools: frozenset[str] = field(default_factory=lambda: MUTATING_TOOL_NAMES) loop_caps: "LoopCapConfig" = field(default_factory=lambda: LoopCapConfig()) @classmethod def from_mapping(cls, data: Mapping[str, Any] | None) -> "ToolCallGuardrailConfig": """Build config from the `tool_loop_guardrails` config.yaml section.""" if not isinstance(data, Mapping): return cls() warn_after = data.get("warn_after") if not isinstance(warn_after, Mapping): warn_after = {} hard_stop_after = data.get("hard_stop_after") if not isinstance(hard_stop_after, Mapping): hard_stop_after = {} defaults = cls() return cls( warnings_enabled=_as_bool(data.get("warnings_enabled"), defaults.warnings_enabled), hard_stop_enabled=_as_bool(data.get("hard_stop_enabled"), defaults.hard_stop_enabled), exact_failure_warn_after=_positive_int( warn_after.get("exact_failure", data.get("exact_failure_warn_after")), defaults.exact_failure_warn_after, ), same_tool_failure_warn_after=_positive_int( warn_after.get("same_tool_failure", data.get("same_tool_failure_warn_after")), defaults.same_tool_failure_warn_after, ), no_progress_warn_after=_positive_int( warn_after.get("idempotent_no_progress", data.get("no_progress_warn_after")), defaults.no_progress_warn_after, ), exact_failure_block_after=_positive_int( hard_stop_after.get("exact_failure", data.get("exact_failure_block_after")), defaults.exact_failure_block_after, ), same_tool_failure_halt_after=_positive_int( hard_stop_after.get("same_tool_failure", data.get("same_tool_failure_halt_after")), defaults.same_tool_failure_halt_after, ), no_progress_block_after=_positive_int( hard_stop_after.get("idempotent_no_progress", data.get("no_progress_block_after")), defaults.no_progress_block_after, ), loop_caps=LoopCapConfig.from_mapping(data.get("loop_caps")), ) # Default session-wide caps, matching Claude Code's v2.1.212 runaway-loop # Per-turn (per-agent-loop) caps on runaway-prone tool calls. Counts reset at # the start of every agent loop (reset_for_turn), so the limit is "within a # single turn" rather than cumulative over the whole session. A single loop # issuing dozens of web searches or spawning dozens of subagents is already # pathological, so the defaults are deliberately low. _DEFAULT_MAX_WEB_SEARCHES_PER_TURN = 50 _DEFAULT_MAX_SUBAGENTS_PER_TURN = 50 @dataclass(frozen=True) class LoopCapConfig: """Per-turn caps on runaway-prone tool calls. Inspired by Claude Code v2.1.212 (Week 29, July 2026), which added caps on WebSearch calls and subagent spawns to stop runaway search / delegation loops. Here the caps count *within a single agent loop* (one turn): the counters reset in ``reset_for_turn`` at the start of every ``run_conversation``, so a legitimate multi-turn session is never starved, but a single turn that spirals into an unbounded search / delegation loop is stopped. Semantics differ from the per-turn loop *detector* above (which keys on repeated identical/failing calls): these caps are a hard ceiling on the total count of a tool within the turn and fire regardless of ``hard_stop_enabled``. A value of ``0`` disables the cap (unlimited). """ max_web_searches: int = _DEFAULT_MAX_WEB_SEARCHES_PER_TURN max_subagents: int = _DEFAULT_MAX_SUBAGENTS_PER_TURN @classmethod def from_mapping(cls, data: Mapping[str, Any] | None) -> "LoopCapConfig": """Build config from the ``tool_loop_guardrails.loop_caps`` section.""" if not isinstance(data, Mapping): return cls() defaults = cls() return cls( max_web_searches=_non_negative_int( data.get("max_web_searches"), defaults.max_web_searches ), max_subagents=_non_negative_int( data.get("max_subagents"), defaults.max_subagents ), ) @dataclass(frozen=True) class ToolCallSignature: """Stable, non-reversible identity for a tool name plus canonical args.""" tool_name: str args_hash: str @classmethod def from_call(cls, tool_name: str, args: Mapping[str, Any] | None) -> "ToolCallSignature": canonical = canonical_tool_args(args or {}) return cls(tool_name=tool_name, args_hash=_sha256(canonical)) def to_metadata(self) -> dict[str, str]: """Return public metadata without raw argument values.""" return {"tool_name": self.tool_name, "args_hash": self.args_hash} @dataclass(frozen=True) class ToolGuardrailDecision: """Decision returned by the tool-call guardrail controller.""" action: str = "allow" # allow | warn | block | halt code: str = "allow" message: str = "" tool_name: str = "" count: int = 0 signature: ToolCallSignature | None = None @property def allows_execution(self) -> bool: return self.action in {"allow", "warn"} @property def should_halt(self) -> bool: return self.action in {"block", "halt"} def to_metadata(self) -> dict[str, Any]: data: dict[str, Any] = { "action": self.action, "code": self.code, "message": self.message, "tool_name": self.tool_name, "count": self.count, } if self.signature is not None: data["signature"] = self.signature.to_metadata() return data def canonical_tool_args(args: Mapping[str, Any]) -> str: """Return sorted compact JSON for parsed tool arguments.""" if not isinstance(args, Mapping): raise TypeError(f"tool args must be a mapping, got {type(args).__name__}") return json.dumps( args, ensure_ascii=False, sort_keys=True, separators=(",", ":"), default=str, ) def classify_tool_failure(tool_name: str, result: str | None) -> tuple[bool, str]: """Safety-fallback classifier used only when callers don't pass ``failed``. Mirrors ``agent.display._detect_tool_failure`` exactly so the guardrail never disagrees with the CLI's user-visible ``[error]`` tag. Production callers in ``run_agent.py`` always pass an explicit ``failed=`` derived from ``_detect_tool_failure``; this function exists so standalone callers (tests, tooling) still get consistent behavior. """ if result is None: return False, "" if file_mutation_result_landed(tool_name, result): return False, "" if tool_name == "terminal": data = safe_json_loads(result) if isinstance(data, dict): exit_code = data.get("exit_code") if exit_code is not None and exit_code != 0: return True, f" [exit {exit_code}]" return False, "" if tool_name == "memory": data = safe_json_loads(result) if isinstance(data, dict): if data.get("success") is False and "exceed the limit" in data.get("error", ""): return True, " [full]" lower = result[:500].lower() if '"error"' in lower or '"failed"' in lower or result.startswith("Error"): return True, " [error]" return False, "" class ToolCallGuardrailController: """Per-turn controller for repeated failed/non-progressing tool calls.""" def __init__(self, config: ToolCallGuardrailConfig | None = None): self.config = config or ToolCallGuardrailConfig() self.reset_for_turn() def reset_for_turn(self) -> None: self._exact_failure_counts: dict[ToolCallSignature, int] = {} self._same_tool_failure_counts: dict[str, int] = {} self._no_progress: dict[ToolCallSignature, tuple[str, int]] = {} self._halt_decision: ToolGuardrailDecision | None = None # Per-turn runaway-loop cap counters. Reset every turn (this method # runs at the start of each run_conversation), so the caps bound a # single agent loop rather than accumulating across the session. self._turn_web_search_count = 0 self._turn_subagent_count = 0 @property def halt_decision(self) -> ToolGuardrailDecision | None: return self._halt_decision def before_call(self, tool_name: str, args: Mapping[str, Any] | None) -> ToolGuardrailDecision: signature = ToolCallSignature.from_call(tool_name, _coerce_args(args)) # ── Per-turn runaway-loop caps ────────────────────────────────── # These are hard ceilings on how many times a runaway-prone tool may # be called within a single agent loop (turn). They apply regardless # of hard_stop_enabled (which only governs the per-turn loop detector). # We block BEFORE the call runs once the count is already at the cap, # then increment for an allowed call so the (cap+1)-th is refused. cap_block = self._check_loop_cap(tool_name, _coerce_args(args), signature) if cap_block is not None: return cap_block if not self.config.hard_stop_enabled: return ToolGuardrailDecision(tool_name=tool_name, signature=signature) exact_count = self._exact_failure_counts.get(signature, 0) if exact_count >= self.config.exact_failure_block_after: decision = ToolGuardrailDecision( action="block", code="repeated_exact_failure_block", message=( f"Blocked {tool_name}: the same tool call failed {exact_count} " "times with identical arguments. Stop retrying it unchanged; " "change strategy or explain the blocker." ), tool_name=tool_name, count=exact_count, signature=signature, ) self._halt_decision = decision return decision if self._is_idempotent(tool_name): record = self._no_progress.get(signature) if record is not None: _result_hash, repeat_count = record if repeat_count >= self.config.no_progress_block_after: decision = ToolGuardrailDecision( action="block", code="idempotent_no_progress_block", message=( f"Blocked {tool_name}: this read-only call returned the same " f"result {repeat_count} times. Stop repeating it unchanged; " "use the result already provided or try a different query." ), tool_name=tool_name, count=repeat_count, signature=signature, ) self._halt_decision = decision return decision return ToolGuardrailDecision(tool_name=tool_name, signature=signature) def after_call( self, tool_name: str, args: Mapping[str, Any] | None, result: str | None, *, failed: bool | None = None, ) -> ToolGuardrailDecision: args = _coerce_args(args) signature = ToolCallSignature.from_call(tool_name, args) if failed is None: failed, _ = classify_tool_failure(tool_name, result) if failed: exact_count = self._exact_failure_counts.get(signature, 0) + 1 self._exact_failure_counts[signature] = exact_count self._no_progress.pop(signature, None) same_count = self._same_tool_failure_counts.get(tool_name, 0) + 1 self._same_tool_failure_counts[tool_name] = same_count if self.config.hard_stop_enabled and same_count >= self.config.same_tool_failure_halt_after: decision = ToolGuardrailDecision( action="halt", code="same_tool_failure_halt", message=( f"Stopped {tool_name}: it failed {same_count} times this turn. " "Stop retrying the same failing tool path and choose a different approach." ), tool_name=tool_name, count=same_count, signature=signature, ) self._halt_decision = decision return decision if self.config.warnings_enabled and exact_count >= self.config.exact_failure_warn_after: return ToolGuardrailDecision( action="warn", code="repeated_exact_failure_warning", message=( f"{tool_name} has failed {exact_count} times with identical arguments. " "This looks like a loop; inspect the error and change strategy " "instead of retrying it unchanged." ), tool_name=tool_name, count=exact_count, signature=signature, ) if self.config.warnings_enabled and same_count >= self.config.same_tool_failure_warn_after: return ToolGuardrailDecision( action="warn", code="same_tool_failure_warning", message=_tool_failure_recovery_hint(tool_name, same_count), tool_name=tool_name, count=same_count, signature=signature, ) return ToolGuardrailDecision(tool_name=tool_name, count=exact_count, signature=signature) self._exact_failure_counts.pop(signature, None) self._same_tool_failure_counts.pop(tool_name, None) if not self._is_idempotent(tool_name): self._no_progress.pop(signature, None) return ToolGuardrailDecision(tool_name=tool_name, signature=signature) result_hash = _result_hash(result) previous = self._no_progress.get(signature) repeat_count = 1 if previous is not None and previous[0] == result_hash: repeat_count = previous[1] + 1 self._no_progress[signature] = (result_hash, repeat_count) if self.config.warnings_enabled and repeat_count >= self.config.no_progress_warn_after: return ToolGuardrailDecision( action="warn", code="idempotent_no_progress_warning", message=( f"{tool_name} returned the same result {repeat_count} times. " "Use the result already provided or change the query instead of " "repeating it unchanged." ), tool_name=tool_name, count=repeat_count, signature=signature, ) return ToolGuardrailDecision(tool_name=tool_name, count=repeat_count, signature=signature) def _is_idempotent(self, tool_name: str) -> bool: if tool_name in self.config.mutating_tools: return False return tool_name in self.config.idempotent_tools def _check_loop_cap( self, tool_name: str, args: Mapping[str, Any], signature: ToolCallSignature, ) -> ToolGuardrailDecision | None: """Enforce and advance the per-turn runaway-loop counters. Returns a ``block`` decision when the cap is already reached, otherwise increments the relevant counter for the allowed call and returns ``None``. A cap of 0 disables that limit entirely. Counters reset each turn via ``reset_for_turn``. """ caps = self.config.loop_caps if tool_name == "web_search": cap = caps.max_web_searches if cap and self._turn_web_search_count >= cap: decision = ToolGuardrailDecision( action="block", code="loop_web_search_cap", message=( f"Blocked web_search: this turn has already made {cap} " "web searches, the per-turn limit. This looks like a " "runaway search loop. Work with the results you already " "have and give the user your answer." ), tool_name=tool_name, count=self._turn_web_search_count, signature=signature, ) self._halt_decision = decision return decision self._turn_web_search_count += 1 return None if tool_name == "delegate_task": cap = caps.max_subagents if not cap: return None spawn_count = _subagent_spawn_count(args) if self._turn_subagent_count >= cap: decision = ToolGuardrailDecision( action="block", code="loop_subagent_cap", message=( f"Blocked delegate_task: this turn has already spawned " f"{self._turn_subagent_count} subagents (limit {cap}). " "This looks like a runaway delegation loop. Finish the " "work with the results you have and answer the user." ), tool_name=tool_name, count=self._turn_subagent_count, signature=signature, ) self._halt_decision = decision return decision self._turn_subagent_count += spawn_count return None return None def toolguard_synthetic_result(decision: ToolGuardrailDecision) -> str: """Build a synthetic role=tool content string for a blocked tool call.""" return json.dumps( { "error": decision.message, "guardrail": decision.to_metadata(), }, ensure_ascii=False, ) def append_toolguard_guidance(result: str, decision: ToolGuardrailDecision) -> str: """Append runtime guidance to the current tool result content.""" if decision.action not in {"warn", "halt"} or not decision.message: return result label = "Tool loop hard stop" if decision.action == "halt" else "Tool loop warning" suffix = ( f"\n\n[{label}: " f"{decision.code}; count={decision.count}; {decision.message}]" ) return (result or "") + suffix def _tool_failure_recovery_hint(tool_name: str, count: int) -> str: """Action-oriented guidance for recovering from repeated tool failures.""" common = ( f"{tool_name} has failed {count} times this turn. This looks like a loop. " "Do not switch to text-only replies; keep using tools, but diagnose before retrying. " "First inspect the latest error/output and verify your assumptions. " ) if tool_name == "terminal": return common + ( "For terminal failures, run a small diagnostic such as `pwd && ls -la` " "in the same tool, then try an absolute path, a simpler command, a different " "working directory, or a different tool such as read_file/write_file/patch." ) return common + ( "Try different arguments, a narrower query/path, an absolute path when relevant, " "or a different tool that can make progress. If the blocker is external, report " "the blocker after one diagnostic attempt instead of repeating the same failing path." ) def _coerce_args(args: Mapping[str, Any] | None) -> Mapping[str, Any]: return args if isinstance(args, Mapping) else {} def _result_hash(result: str | None) -> str: parsed = safe_json_loads(result or "") if parsed is not None: try: canonical = json.dumps( parsed, ensure_ascii=False, sort_keys=True, separators=(",", ":"), default=str, ) except TypeError: canonical = str(parsed) else: canonical = result or "" return _sha256(canonical) def _as_bool(value: Any, default: bool) -> bool: if value is None: return default if isinstance(value, bool): return value if isinstance(value, (int, float)): return bool(value) if isinstance(value, str): lowered = value.strip().lower() if lowered in {"1", "true", "yes", "on", "enabled"}: return True if lowered in {"0", "false", "no", "off", "disabled"}: return False return default def _positive_int(value: Any, default: int) -> int: if value is None: return default try: parsed = int(value) except (TypeError, ValueError): return default return parsed if parsed >= 1 else default def _non_negative_int(value: Any, default: int) -> int: """Parse a session-cap value. 0 is a valid (disable) value; negatives and junk fall back to the default.""" if value is None: return default try: parsed = int(value) except (TypeError, ValueError): return default return parsed if parsed >= 0 else default def _subagent_spawn_count(args: Mapping[str, Any]) -> int: """How many subagents a single delegate_task call spawns. delegate_task runs in one of two modes: a batch (``tasks`` is a non-empty list, one child per item) or a single task (``goal``). Count the batch size when present, otherwise 1, so the session subagent cap reflects real spawns rather than delegate_task invocations. """ tasks = args.get("tasks") if isinstance(args, Mapping) else None if isinstance(tasks, list) and tasks: return len(tasks) return 1 def _sha256(value: str) -> str: # surrogatepass: tool results scraped from the web can carry unpaired # UTF-16 surrogates (e.g. half of a mathematical-bold pair); a strict # encode raises and takes down the whole conversation loop. The hash only # needs deterministic bytes, not valid UTF-8. return hashlib.sha256(value.encode("utf-8", "surrogatepass")).hexdigest()