feat(settings): add a tool approval mode with an explicit off (YOLO) option (#6304)

* feat: add optional tool approval mode

Admin-configurable tool_approval_mode setting (ask/auto) for the
post-external-context tool approval gate.

- ask (default): existing approval flow, gate stays on
- auto: bypasses the approval gate for all runs - same effect as
  ODYSSEUS_TOOL_APPROVAL_GATE=0. Documented honestly in the settings
  row copy and in THREAT_MODEL.md, per review on #6304.
- value validated server-side ({ask, auto}, anything else -> 400)
- no new env var: the setting wires to the existing gate control
- tests: setting validation + _tool_approval_gate_bypassed behavior

* feat: surface deployment gate state on Tool approval row

Follow-up to StressTestor's re-review: when ODYSSEUS_TOOL_APPROVAL_GATE is
set to a falsy value, the Tool approval row still read 'Always ask' even though
the gate was off, misleading admins the other way.

GET /api/auth/settings now returns tool_approval_gate_env_disabled (derived
from TOOL_APPROVAL_GATE_ENABLED) for admins, and the settings row renders a
red notice naming the env var and explaining that the gate is off for every
run regardless of the stored setting. Non-admin/scrubbed responses do not
expose the deployment state. Test covers both env values and the scrub path.

* refactor: rename tool approval mode auto to off

The gate-off value is now "off" ("Off (YOLO)" in the UI) so the auto-approve name stays available for the future judged mode.

---------

Co-authored-by: Matej <matej@local>

THREAT_MODEL.md

@@ -64,7 +64,7 @@ External content that reaches the LLM is treated as untrusted via `src/prompt_se
 
 `src/tool_capabilities.py` carries a second layer: once untrusted content has entered a run, `ToolRunSecurityContext.decision_for()` blocks tools that execute code, mutate state, or cause external side effects until the user authorises the action separately.
 
-It is on unless `ODYSSEUS_TOOL_APPROVAL_GATE` is set to a falsy value (`0`/`false`/`no`/`off`). Request authority narrows which tool families a turn may use, but within an admitted family it does not bind the exact action: a request to read email admits `send_email` and `delete_email`, and agent processes inherit the host network. Until those are covered by their own boundaries, this gate is what stops an injected instruction from reaching them. Turning it off lets a successful injection reach `bash`, `host_shell`, `send_email` and `delete_email` without a separate confirmation.
+It is on unless `ODYSSEUS_TOOL_APPROVAL_GATE` is set to a falsy value (`0`/`false`/`no`/`off`). The admin-only `tool_approval_mode` setting (Settings → Agent Tools, `off` = "Off (YOLO)") is the same knob exposed in the UI: with `off`, every run sets `approval_gate_bypassed` and this gate stops asking, exactly as if the env var were off. Request authority narrows which tool families a turn may use, but within an admitted family it does not bind the exact action: a request to read email admits `send_email` and `delete_email`, and agent processes inherit the host network. Until those are covered by their own boundaries, this gate is what stops an injected instruction from reaching them. Turning it off — by either knob — lets a successful injection reach `bash`, `host_shell`, `send_email` and `delete_email` without a separate confirmation.
 
 Three exemptions apply even when the gate is on, all deliberate:
 

routes/auth_routes.py

@@ -18,6 +18,7 @@ from src.constants import DEEP_RESEARCH_DIR, MEMORY_FILE, PASSWORD_MIN_LENGTH, S
 from src.auth_helpers import _auth_disabled
 from src.rate_limiter import RateLimiter
 from src.settings_scrub import scrub_settings
+from src import tool_capabilities
 from src.settings import (
     load_settings as _load_settings,
     save_settings as _save_settings,
@@ -753,6 +754,11 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         user = _get_current_user(request)
         settings = without_retired_settings(_load_settings())
         if _auth_disabled() or (user and auth_manager.is_admin(user)):
+            # Deployment-level truth for the Tool approval row: a falsy
+            # ODYSSEUS_TOOL_APPROVAL_GATE disables the approval gate for every
+            # run regardless of the stored tool_approval_mode, so surface it to
+            # admins instead of letting the row mislead them the other way.
+            settings["tool_approval_gate_env_disabled"] = not tool_capabilities.TOOL_APPROVAL_GATE_ENABLED
             return settings
         return scrub_settings(settings)
 
@@ -769,6 +775,9 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
             "agent_max_tool_calls": (0, 1000),  # 0 = unlimited
             "auto_compact_threshold_percent": (50, 95),
         }
+        _VALID_VALUES = {
+            "tool_approval_mode": {"ask", "off"},
+        }
         for key in DEFAULT_SETTINGS:
             if key in RETIRED_SETTING_KEYS:
                 continue
@@ -782,6 +791,15 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
                 except (TypeError, ValueError):
                     raise HTTPException(400, f"{key} must be an integer")
                 val = max(lo, min(val, hi))
+
+            if key in _VALID_VALUES:
+                allowed = _VALID_VALUES[key]
+                if not isinstance(val, str) or val not in allowed:
+                    raise HTTPException(
+                        400,
+                        f"{key} must be one of {sorted(allowed)}",
+                    )
+
             current[key] = val
         _save_settings(current)
         return without_retired_settings(current)

src/agent_loop.py

@@ -14,12 +14,12 @@ import csv
 import difflib
 import html
 import json
+import logging
 import os
 import re
 import shlex
 import shutil
 import time
-import logging
 import hashlib
 from src.web_recovery import WebRecoveryBudget
 from itertools import count
@@ -16994,6 +16994,20 @@ def _detect_runaway_call(call_freq, threshold=15):
     sig = next((s for s, n in call_freq.items() if n >= threshold), None)
     return sig.split(":", 1)[0] if sig else None
 
+def _tool_approval_gate_bypassed(
+    exact_approval: Optional[ExactToolApproval] = None,
+) -> bool:
+    """Return whether the post-external-context tool approval gate is bypassed.
+
+    ``tool_approval_mode=off`` (admin setting, Settings → Agent Tools) bypasses
+    the gate on every run, the same effect as ``ODYSSEUS_TOOL_APPROVAL_GATE=0``.
+    A chat-session grant bypasses it only within the scope of its own task.
+    """
+    return (
+        get_setting("tool_approval_mode", "ask") == "off"
+        or bool(exact_approval and exact_approval.allow_remaining_actions)
+    )
+
 
 def _tool_result_signature(tool_result_records: list[dict]) -> str:
     """Return a bounded fingerprint of the observable result of a tool batch.
@@ -21437,9 +21451,7 @@ async def stream_agent_loop(
             )
             or messages_contain_external_untrusted_context(messages)
         ),
-        approval_gate_bypassed=bool(
-            exact_approval and exact_approval.allow_remaining_actions
-        ),
+        approval_gate_bypassed=_tool_approval_gate_bypassed(exact_approval),
         delegated_credential=bool(delegated_credential),
     )
     _has_tui_host_bridge = _tui_host_bridge_is_usable(client_runtime_context)

src/agent_runtime/effect_log.py

@@ -47,6 +47,7 @@ EFFECTS_DIR = os.path.join(DATA_DIR, "effects")
 _RUN_ID = re.compile(r"[a-f0-9]{32}")
 _TYPES = {"claim": EffectClaim, "outcome": EffectOutcome, "observation": Observation}
 _VERSION = 1
+_O_BINARY = getattr(os, "O_BINARY", 0)  # Windows: raw os.open() is text-mode without it
 
 
 def _fsync_directory(directory: str | os.PathLike) -> None:
@@ -243,7 +244,7 @@ class EffectLog:
         """Hold the file exclusively with every durable record merged."""
         assert self.path is not None
         _ensure_directory(self.path.parent)
-        flags = os.O_RDWR | os.O_APPEND | os.O_CREAT | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0)
+        flags = os.O_RDWR | os.O_APPEND | os.O_CREAT | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0) | _O_BINARY
         descriptor = os.open(self.path, flags, 0o600)
         try:
             if fcntl is not None:
@@ -358,7 +359,7 @@ class EffectLog:
         with self._lock:
             try:
                 descriptor = os.open(self.path, os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0)
-                                     | getattr(os, "O_CLOEXEC", 0))
+                                     | getattr(os, "O_CLOEXEC", 0) | _O_BINARY)
             except FileNotFoundError:
                 return
             except OSError:
@@ -394,7 +395,7 @@ class EffectLog:
         log = cls(run_id, directory=directory)
         assert log.path is not None
         try:
-            descriptor = os.open(log.path, os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0))
+            descriptor = os.open(log.path, os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0) | _O_BINARY)
         except FileNotFoundError:
             return log
         except OSError as error:
@@ -460,7 +461,7 @@ class EffectLog:
         temporary = target.with_suffix(".tmp")
         data = json.dumps({"run_id": self.run_id, "effect_id": effect_id}, sort_keys=True).encode()
         _ensure_directory(target.parent)
-        flags = os.O_WRONLY | os.O_CREAT | os.O_TRUNC | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0)
+        flags = os.O_WRONLY | os.O_CREAT | os.O_TRUNC | getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_CLOEXEC", 0) | _O_BINARY
         descriptor = os.open(temporary, flags, 0o600)
         try:
             view = memoryview(data)
@@ -478,7 +479,7 @@ class EffectLog:
             return None
         base = Path(directory) if directory is not None else effects_dir()
         try:
-            descriptor = os.open(base / f"launch-{generation}.json", os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0))
+            descriptor = os.open(base / f"launch-{generation}.json", os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0) | _O_BINARY)
             with os.fdopen(descriptor, "rb") as stream:
                 if os.fstat(stream.fileno()).st_nlink != 1:
                     return None

src/settings.py

@@ -120,6 +120,7 @@ DEFAULT_SETTINGS = {
     # Tune via Settings or by editing data/settings.json.
     "research_run_timeout_seconds": 1800,
     "agent_max_tool_calls": 0,
+    "tool_approval_mode": "ask",
     "agent_max_rounds": 20,  # per-message agent step cap (clamped 1..200)
     # Soft input-token budget for the agent loop. The DEFAULT value (6000) is the
     # "auto" sentinel: it means "scale the budget to the model's context window"

static/index.html

@@ -2463,6 +2463,22 @@
                 <label class="settings-label">Max steps per message</label>
                 <input id="set-agentMaxRounds" type="text" inputmode="numeric" placeholder="20" class="settings-select" style="width:120px;">
               </div>
+              <div class="settings-row">
+                <label class="settings-label">Tool approval</label>
+                <select id="set-toolApprovalMode" class="settings-select">
+                  <option value="ask">Always ask</option>
+                  <option value="off">Off (YOLO)</option>
+                </select>
+              </div>
+
+              <div class="admin-toggle-sub" style="margin-top:4px;">
+                Auto-approve turns off the tool approval gate (same control as the
+                ODYSSEUS_TOOL_APPROVAL_GATE env var). Bash, host_shell, send_email and
+                delete_email then run without a confirmation even after the agent has
+                seen untrusted external content (email, web pages, documents), so a
+                prompt-injected instruction can reach them directly.
+              </div>
+              <div id="set-toolApprovalEnvNote" class="admin-toggle-sub" style="margin-top:4px;display:none;color:var(--red);"></div>
               <div id="set-agentMsg" style="font-size:11px;color:color-mix(in srgb, var(--fg) 45%, transparent);"></div>
             </div>
           </div>

static/js/settings/agent.js

@@ -10,6 +10,7 @@ export async function initAgentSettings() {
   var toolsInput = el('set-agentMaxTools');
   var roundsInput = el('set-agentMaxRounds');
   var supInput = el('set-agentSupervisorLadder');
+  var approvalModeInput = el('set-toolApprovalMode');
   var msg = el('set-agentMsg');
   if (!toolsInput) return;
 
@@ -19,6 +20,21 @@ export async function initAgentSettings() {
     if (settings.agent_max_tool_calls) toolsInput.value = settings.agent_max_tool_calls;
     if (roundsInput && settings.agent_max_rounds) roundsInput.value = settings.agent_max_rounds;
     if (supInput) supInput.checked = !!settings.agent_supervisor_ladder;
+    if (approvalModeInput) {
+      approvalModeInput.value =
+        settings.tool_approval_mode === 'off' ? 'off' : 'ask';
+    }
+    var approvalEnvNote = el('set-toolApprovalEnvNote');
+    if (approvalEnvNote) {
+      if (settings.tool_approval_gate_env_disabled) {
+        approvalEnvNote.textContent =
+          'Forced by deployment: ODYSSEUS_TOOL_APPROVAL_GATE is set to a falsy value on this server, so the approval gate is off for every run regardless of this setting.';
+        approvalEnvNote.style.display = 'block';
+      } else {
+        approvalEnvNote.textContent = '';
+        approvalEnvNote.style.display = 'none';
+      }
+    }
   } catch (e) {}
 
   // Clamp + coerce a raw input to an int in [lo, hi]; falls back to `dflt`
@@ -37,6 +53,10 @@ export async function initAgentSettings() {
     var payload = { agent_max_tool_calls: tools };
     if (rounds != null) payload.agent_max_rounds = rounds;
     if (supInput) payload.agent_supervisor_ladder = !!supInput.checked;
+    if (approvalModeInput) {
+      payload.tool_approval_mode =
+        approvalModeInput.value === 'off' ? 'off' : 'ask';
+    }
     try {
       await _postSettings(payload);
       msg.textContent = (tools > 0 ? 'Limit: ' + tools + ' tool calls' : 'Unlimited tool calls') +
@@ -49,6 +69,7 @@ export async function initAgentSettings() {
   toolsInput.addEventListener('change', save);
   if (roundsInput) roundsInput.addEventListener('change', save);
   if (supInput) supInput.addEventListener('change', save);
+  if (approvalModeInput) approvalModeInput.addEventListener('change', save);
   var cur = parseInt(toolsInput.value, 10) || 0;
   var curR = roundsInput ? (parseInt(roundsInput.value, 10) || 20) : null;
   msg.textContent = (cur > 0 ? 'Limit: ' + cur + ' tool calls' : 'Unlimited tool calls') +

tests/test_effect_journal_persistence.py

@@ -344,3 +344,13 @@ def test_hardlinked_effect_state_is_control_plane_without_scanning_the_store(tmp
     elsewhere.write_text("y")
     os.link(elsewhere, workspace / "pnpm-style.txt")
     assert resources._control_plane_path(str(workspace / "pnpm-style.txt")) is True
+
+
+def test_consecutive_claims_write_lf_only_records(tmp_path, target):
+    # Raw os.open() is text-mode on Windows unless O_BINARY is set; the \r\n it
+    # wrote made the next append read a stray empty line as a corrupt tail.
+    log = EffectLog(RUN, directory=tmp_path / "fx")
+    claim(log, target, "e1", "act-1")
+    claim(log, target, "e2", "act-2")
+    assert b"\r" not in log.path.read_bytes()
+    assert len(records(log.path)) == 2

tests/test_retired_settings_interfaces.py

@@ -3,6 +3,7 @@
 import asyncio
 import json
 from types import SimpleNamespace
+from fastapi import HTTPException
 
 import pytest
 
@@ -117,3 +118,62 @@ def test_manage_settings_tombstones_legacy_fallback_key(monkeypatch):
 
     assert save_calls == []
     assert store["default_model_fallbacks"] == LEGACY_VALUE
+
+@pytest.mark.asyncio
+async def test_tool_approval_mode_accepts_only_supported_values(monkeypatch):
+    store = dict(settings_mod.DEFAULT_SETTINGS)
+
+    monkeypatch.setattr(auth_routes, "migrate_from_settings", lambda: None)
+    monkeypatch.setattr(auth_routes, "_load_settings", lambda: dict(store))
+
+    def save_settings(updated):
+        store.clear()
+        store.update(updated)
+
+    monkeypatch.setattr(auth_routes, "_save_settings", save_settings)
+
+    router = auth_routes.setup_auth_routes(_AuthManager())
+    set_settings = _route(router, "/api/auth/settings", "POST")
+
+    response = await set_settings(
+        _Request({"tool_approval_mode": "off"}, admin=True)
+    )
+    assert response["tool_approval_mode"] == "off"
+    assert store["tool_approval_mode"] == "off"
+
+    response = await set_settings(
+        _Request({"tool_approval_mode": "ask"}, admin=True)
+    )
+    assert response["tool_approval_mode"] == "ask"
+    assert store["tool_approval_mode"] == "ask"
+
+    with pytest.raises(HTTPException) as exc_info:
+        await set_settings(
+            _Request({"tool_approval_mode": "banana"}, admin=True)
+        )
+
+    assert exc_info.value.status_code == 400
+
+@pytest.mark.asyncio
+async def test_settings_expose_gate_env_kill_switch_to_admins_only(monkeypatch):
+    store = dict(settings_mod.DEFAULT_SETTINGS)
+
+    monkeypatch.setattr(auth_routes, "migrate_from_settings", lambda: None)
+    monkeypatch.setattr(auth_routes, "_load_settings", lambda: dict(store))
+
+    router = auth_routes.setup_auth_routes(_AuthManager())
+    get_settings = _route(router, "/api/auth/settings", "GET")
+
+    monkeypatch.setattr(auth_routes.tool_capabilities, "TOOL_APPROVAL_GATE_ENABLED", True)
+    response = await get_settings(_Request(admin=True))
+    assert response["tool_approval_gate_env_disabled"] is False
+
+    monkeypatch.setattr(auth_routes.tool_capabilities, "TOOL_APPROVAL_GATE_ENABLED", False)
+    response = await get_settings(_Request(admin=True))
+    assert response["tool_approval_gate_env_disabled"] is True
+    # The stored setting itself is untouched by the env kill-switch.
+    assert response["tool_approval_mode"] == "ask"
+
+    # Unauthenticated/scrubbed callers must not learn the deployment gate state.
+    response = await get_settings(_Request())
+    assert "tool_approval_gate_env_disabled" not in response

tests/test_tool_approval_task_scope.py

@@ -340,7 +340,7 @@ def test_route_context_agent_frontend_and_cache_bust_wire_the_contract():
     assert "_without_latest_matching_user_message(" not in helpers
     assert "selected_tools=approval_selected_tools" in agent
     assert "continuation_query=_retrieval_query or _last_user" in agent
-    assert "approval_gate_bypassed=bool(" in agent
+    assert "approval_gate_bypassed=_tool_approval_gate_bypassed(" in agent
     assert "['approve', 'approve_task', 'deny']" in frontend
     assert "input.value = label" not in frontend
     assert "const msg = approvalForSend ? '' : el('message').value;" in frontend

tests/test_tool_approvals.py

@@ -500,3 +500,37 @@ async def test_dispatcher_revalidates_sealed_workspace(monkeypatch, tmp_path):
 
     assert result["blocked"] is True
     assert result["policy"] == "exact_tool_approval"
+
+def test_tool_approval_mode_defaults_to_ask(monkeypatch):
+    import src.agent_loop as agent_loop
+
+    monkeypatch.setattr(
+        agent_loop,
+        "get_setting",
+        lambda key, default=None: (
+            "ask" if key == "tool_approval_mode" else default
+        ),
+    )
+
+    assert agent_loop._tool_approval_gate_bypassed() is False
+
+def test_tool_approval_gate_respects_setting(monkeypatch):
+    import src.agent_loop as agent_loop
+
+    monkeypatch.setattr(
+        agent_loop,
+        "get_setting",
+        lambda key, default=None: (
+            "ask" if key == "tool_approval_mode" else default
+        ),
+    )
+    assert agent_loop._tool_approval_gate_bypassed() is False
+
+    monkeypatch.setattr(
+        agent_loop,
+        "get_setting",
+        lambda key, default=None: (
+            "off" if key == "tool_approval_mode" else default
+        ),
+    )
+    assert agent_loop._tool_approval_gate_bypassed() is True