core/agent: let YOLO mode skip YOLO-tagged soft confirms

`--yes` / `/yolo on|off|once` clears the prompt of a `YOLO`-tagged tool: interpreter runs (`run_argv`, `run_argv_batch`, `open_pty`), the file mutations (overwrite / copy / move / delete) and `fetch_local`'s cleartext and mutating requests. A tool without the tag still asks, and hard path/command denylists stay untouched.

`classify_danger` now answers with a `DangerReason` — the reason text plus `yolo_skip` — so one call can refuse that skip: a command that runs through `sudo` / `su` / `doas` / `pkexec` (as the program or a wrapper) keeps its confirm whatever YOLO mode says. It never adds a reason of its own, so the program the denylist gate already asks about is not asked twice.

`prepare_batch_confirm` leaves the covered calls out, so a skipped call neither prompts on its own nor takes a slot in the round's aggregated prompt, and one escalating step keeps the whole batch in front of the human.

src/plyngent/agent/__init__.py

@@ -30,6 +30,7 @@ from .notices import skill_tools_notice as skill_tools_notice
 from .notices import truncated_turn_notice as truncated_turn_notice
 from .tools import BatchConfirmDecision as BatchConfirmDecision
 from .tools import DangerClassifier as DangerClassifier
+from .tools import DangerReason as DangerReason
 from .tools import ToolBatchConfirmHook as ToolBatchConfirmHook
 from .tools import ToolConfirmHook as ToolConfirmHook
 from .tools import ToolConfirmRequest as ToolConfirmRequest

src/plyngent/agent/tools.py

@@ -16,8 +16,22 @@ from msgspec import Struct
 from plyngent.lmproto.openai_compatible.model import ChatImagePart, ToolFunction, ToolFunctionItem
 from plyngent.typedef import JSONSchema  # noqa: TC001
 
+
+@dataclass(frozen=True, slots=True)
+class DangerReason:
+    """Why a call needs a soft confirm, and whether YOLO mode may drop that prompt.
+
+    ``yolo_skip`` is false for a call that reaches rights the session does not
+    have (``sudo`` / ``doas`` / ``pkexec`` and friends): the human answers it
+    however YOLO mode is set.
+    """
+
+    text: str
+    yolo_skip: bool = True
+
+
 type ToolHandler = Callable[..., Any | Awaitable[Any]]
-type DangerClassifier = Callable[[str, Mapping[str, object]], str | None]
+type DangerClassifier = Callable[[str, Mapping[str, object]], str | DangerReason | None]
 # Returns True to allow, False to deny, or a non-empty str as denial reason for the model.
 type ToolConfirmResult = bool | str
 type ToolConfirmHook = Callable[
@@ -313,6 +327,13 @@ def _denied_error(name: str, reason: str, comment: str | None) -> str:
     return f"error: tool {name!r} denied by user confirm ({reason})"
 
 
+def _as_danger_reason(raw: str | DangerReason | None) -> DangerReason | None:
+    """Normalize a classifier answer; a bare string is a reason YOLO mode may drop."""
+    if raw is None:
+        return None
+    return raw if isinstance(raw, DangerReason) else DangerReason(text=raw)
+
+
 def _write_path_key(value: str) -> str:
     """Comparable key for one path argument of a writing tool.
 
@@ -559,7 +580,9 @@ class ToolRegistry:
 
         Mirrors the front half of :meth:`execute`: a call whose arguments fail to
         decode or miss a required argument is answered with an error there and
-        must not be put in front of the user now.
+        must not be put in front of the user now, and a call whose reason YOLO
+        mode covers runs unprompted, so it must not take a slot in the round's
+        aggregated prompt either.
         """
         definition = self._tools.get(name)
         if self._danger is None or definition is None:
@@ -567,8 +590,10 @@ class ToolRegistry:
         args = _decode_tool_args(arguments_json)
         if isinstance(args, str) or _missing_required_error(definition, args) is not None:
             return None
-        reason = self._danger(name, args)
-        return None if reason is None else ToolConfirmRequest(name=name, args=args, reason=reason)
+        reason = _as_danger_reason(self._danger(name, args))
+        if reason is None or (self._yolo and (definition.tags & ToolTag.YOLO) and reason.yolo_skip):
+            return None
+        return ToolConfirmRequest(name=name, args=args, reason=reason.text)
 
     def _batch_requests(
         self,
@@ -633,29 +658,32 @@ class ToolRegistry:
         tags: ToolTag,
         decision: BatchConfirmDecision | None = None,
     ) -> str | None:
-        """Soft-confirm gate driven by danger reason + tags + grants.
+        """Soft-confirm gate driven by danger reason + tags + grants + YOLO.
 
         Pipeline (soft gate only; hard denylists stay in tool handlers)::
 
             no soft reason → run
+            YOLO mode, (tags & YOLO) and a reason it may drop → allow
             batch decision for this call → apply it
             (tags & TRUSTABLE) and grant exists → allow
-            else → on_confirm (decorator YOLO tag does not override
-                   dangerous invocations — internal danger classification
-                   always takes precedence)
+            else → on_confirm (a tool without the YOLO tag, or a reason YOLO
+                   may not drop — a privilege escalation — still prompts
+                   under YOLO mode)
         """
         if self._danger is None:
             return None
-        reason = self._danger(name, args)
+        reason = _as_danger_reason(self._danger(name, args))
         if reason is None:
             return None
+        if self._yolo and (tags & ToolTag.YOLO) and reason.yolo_skip:
+            return None
 
         async with self._confirm_lock:
             if decision is not None:
-                return await self._apply_batch_decision(name, reason, decision, tags=tags)
+                return await self._apply_batch_decision(name, reason.text, decision, tags=tags)
             if self._grant_allows(name, tags=tags):
                 return None
-            return await self._prompt_soft_confirm(name, args, reason, tags=tags)
+            return await self._prompt_soft_confirm(name, args, reason.text, tags=tags)
 
     async def execute(
         self,

src/plyngent/tools/command_scan.py

@@ -16,7 +16,16 @@ from typing import TYPE_CHECKING
 if TYPE_CHECKING:
     from collections.abc import Sequence
 
-__all__ = ["CommandString", "UnwrappedCommand", "basename", "is_interpreter", "unwrap_command"]
+__all__ = [
+    "ESCALATION_PROGRAMS",
+    "CommandString",
+    "UnwrappedCommand",
+    "basename",
+    "dash_c_code",
+    "escalators_in_chain",
+    "is_interpreter",
+    "unwrap_command",
+]
 
 # Wrappers that run the program named on their own command line.
 _ENV_WRAPPERS: frozenset[str] = frozenset(
@@ -99,6 +108,10 @@ _INTERPRETERS: frozenset[str] = frozenset(
 # Trailing version of an interpreter: ``python3.12`` / ``node20`` / ``php8.2``.
 _VERSION_SUFFIX = re.compile(r"[0-9][0-9.]*$")
 
+# Privilege escalation: these run with rights the session itself does not have,
+# so a call that reaches one is never approved on the caller's behalf
+# (``danger._command_reason`` keeps the confirm whatever YOLO mode says).
+ESCALATION_PROGRAMS: frozenset[str] = frozenset({"doas", "pkexec", "run0", "su", "sudo", "systemd-run"})
 # ``FOO=bar`` in front of a program (env/sudo style).
 _ASSIGNMENT = re.compile(r"[A-Za-z_][A-Za-z0-9_]*=")
 
@@ -245,6 +258,31 @@ def is_interpreter(base: str) -> bool:
     return base in _INTERPRETERS or _VERSION_SUFFIX.sub("", base) in _INTERPRETERS
 
 
+def dash_c_code(command: Sequence[str]) -> tuple[int, str] | None:
+    """Index and text of the ``-c`` code in an interpreter argv, if present.
+
+    Only interpreters hand ``-c`` a block of code; for anything else it is a
+    plain flag (``grep -c``, ``od -c``) with nothing to review.
+    """
+    for index, part in enumerate(command[1:], start=1):
+        if part == "-c":
+            return index, command[index + 1] if index + 1 < len(command) else ""
+        # Combined short forms are uncommon; only exact -c is supported.
+    return None
+
+
+def escalators_in_chain(argv: Sequence[str]) -> tuple[str, ...]:
+    """Escalation programs *argv* runs: the program itself or a wrapper around it.
+
+    ``sudo -u root python …`` and ``doas bash`` both count, whatever program they
+    finally run. Code an interpreter only carries as text is out of scope here —
+    this reads the command line, not what the text would do.
+    """
+    scan = unwrap_command(argv)
+    names = (scan.base, *scan.wrappers)
+    return tuple(dict.fromkeys(name for name in names if name in ESCALATION_PROGRAMS))
+
+
 def unwrap_command(argv: Sequence[str]) -> UnwrappedCommand:
     """Resolve *argv* to the program behind any wrappers and ``run`` launchers.
 

src/plyngent/tools/danger.py

@@ -4,7 +4,8 @@ import json
 import shlex
 from typing import TYPE_CHECKING, cast
 
-from plyngent.tools.command_scan import is_interpreter, unwrap_command
+from plyngent.agent.tools import DangerReason
+from plyngent.tools.command_scan import dash_c_code, escalators_in_chain, is_interpreter, unwrap_command
 from plyngent.tools.workspace import AccessMode, WorkspaceError, resolve_path
 
 if TYPE_CHECKING:
@@ -23,19 +24,6 @@ def _as_argv(args: Mapping[str, object], *, key: str = "command") -> list[str] |
     return out
 
 
-def _dash_c_code(command: Sequence[str]) -> tuple[int, str] | None:
-    """Index and text of the ``-c`` code in an interpreter argv, if present.
-
-    Only interpreters hand ``-c`` a block of code; for anything else it is a
-    plain flag (``grep -c``, ``od -c``) with nothing to review.
-    """
-    for index, part in enumerate(command[1:], start=1):
-        if part == "-c":
-            return index, command[index + 1] if index + 1 < len(command) else ""
-        # Combined short forms are uncommon; only exact -c is supported.
-    return None
-
-
 def _indent_block(text: str, *, prefix: str = "  ") -> str:
     """Indent every line of *text* (for multi-line -c code in confirm prompts)."""
     if not text:
@@ -51,7 +39,7 @@ def _format_argv_for_confirm(argv: Sequence[str], *, code_index: int | None) ->
     return shlex.join(parts)
 
 
-def _command_reason(argv: Sequence[str], *, via: str) -> str | None:
+def _command_reason(argv: Sequence[str], *, via: str) -> DangerReason | None:
     """Confirm a command that runs code or detaches, whatever wraps it.
 
     Multi-line reason (shown inside the CLI confirm box). ``via`` is a short
@@ -65,6 +53,11 @@ def _command_reason(argv: Sequence[str], *, via: str) -> str | None:
     (``env -S "…"``, ``su -c "…"``) hides code the same way, so it is confirmed
     too, with the text it carries printed under its own ``<wrapper> <option>:``
     label.
+
+    A command that escalates privileges (``sudo``/``doas``/``pkexec`` as the
+    program or a wrapper) is confirmed with ``yolo_skip=False``. It never adds a
+    reason of its own: the denylist gate in the handler always asks about the
+    escalation, so a second prompt here would only repeat it.
     """
     scan = unwrap_command(argv)
     interpreter = is_interpreter(scan.base)
@@ -75,7 +68,10 @@ def _command_reason(argv: Sequence[str], *, via: str) -> str | None:
     findings = list(dict.fromkeys(findings))
     if not findings:
         return None
-    code = _dash_c_code(scan.command) if interpreter else None
+    escalators = escalators_in_chain(argv)
+    if escalators:
+        findings.append(f"privilege escalation {', '.join(escalators)}")
+    code = dash_c_code(scan.command) if interpreter else None
     code_index = scan.offset + code[0] if code is not None else None
     display = _format_argv_for_confirm(argv, code_index=code_index)
     reason = f"{via}: {' + '.join(findings)} — review before allow\n  argv: {display}"
@@ -83,7 +79,7 @@ def _command_reason(argv: Sequence[str], *, via: str) -> str | None:
         reason += f"\n  command:\n{_indent_block(code[1])}"
     for span in scan.command_strings:
         reason += f"\n  {span.label}:\n{_indent_block(span.text)}"
-    return reason
+    return DangerReason(text=reason, yolo_skip=not escalators)
 
 
 def _write_file_reason(args: Mapping[str, object]) -> str | None:
@@ -139,7 +135,7 @@ def _delete_path_reason(args: Mapping[str, object]) -> str | None:
     return f"delete path {path!r}{extra}"
 
 
-def _run_argv_reason(args: Mapping[str, object]) -> str | None:
+def _run_argv_reason(args: Mapping[str, object]) -> DangerReason | None:
     argv = _as_argv(args, key="argv")
     if argv is None:
         return None
@@ -161,8 +157,12 @@ def _batch_step_argv(item: object) -> list[str] | None:
     return parts or None
 
 
-def _run_argv_batch_reason(args: Mapping[str, object]) -> str | None:
-    """One confirm for the whole batch if any step runs code or detaches."""
+def _run_argv_batch_reason(args: Mapping[str, object]) -> DangerReason | None:
+    """One confirm for the whole batch if any step runs code or detaches.
+
+    YOLO mode may drop it only when every step's reason allows that, so one
+    escalating step keeps the whole batch in front of the human.
+    """
     raw = args.get("steps")
     if isinstance(raw, str):
         try:
@@ -181,10 +181,13 @@ def _run_argv_batch_reason(args: Mapping[str, object]) -> str | None:
     if not risky:
         return None
     header = f"run_argv_batch: {len(risky)} risky step(s) (review before allow)"
-    return header + "\n" + "\n".join(risky)
+    return DangerReason(
+        text=header + "\n" + "\n".join(reason.text for reason in risky),
+        yolo_skip=all(reason.yolo_skip for reason in risky),
+    )
 
 
-def _open_pty_reason(args: Mapping[str, object]) -> str | None:
+def _open_pty_reason(args: Mapping[str, object]) -> DangerReason | None:
     argv = _as_argv(args)
     if argv is None:
         return None
@@ -204,14 +207,18 @@ def _fetch_reason(args: Mapping[str, object]) -> str | None:
     return fetch_soft_reason(method, url, body)
 
 
-def classify_danger(name: str, args: Mapping[str, object]) -> str | None:  # noqa: PLR0911
-    """Return a short reason if ``name``/``args`` need user confirm, else ``None``.
+def classify_danger(name: str, args: Mapping[str, object]) -> str | DangerReason | None:  # noqa: PLR0911
+    """Return why ``name``/``args`` need a user confirm, else ``None``.
 
     Hard denylists (paths/commands) still raise independently. This only covers
     soft confirms for mutating tools and risky command launches — interpreters
     (however invoked) and detached runners such as ``nohup``, resolved past
     environment wrappers and ``<launcher> run``.
     Private/loopback fetch_local targets use a separate policy grant (not YOLO).
+
+    A reason also says whether YOLO mode may drop its prompt
+    (:attr:`DangerReason.yolo_skip`); a command that escalates privileges
+    refuses.
     """
     if name == "delete_path":
         return _delete_path_reason(args)