Merge commit from fork

* fix(security): keep agent file tools out of the app state directory

The agent's read tools (read_file, grep, glob, ls) resolved model-supplied
paths against a root list whose first entry was the whole data directory.
That directory holds the session store, the auth database, the app
encryption key and the settings file, so prompt-injected content could ask
for any of them. No approval prompt stood in the way: reads are classified
read_workspace and pass the untrusted-context gate untouched, which is
correct for reading a workspace and wrong for reading the app's own state.

The agent gets data/agent_workspace/ instead, and the subprocess cwd and
HOME move with it so bash and read_file agree on where scratch files live.

The deny itself is a property of the path, not of the root it arrived
through, because three routes reach the same bytes and closing only the
first leaves the other two working:

  - the default root list
  - a workspace bound at or above the data directory, which vet_workspace
    accepted and chat_routes auto-binds from a path named in the message
  - a tool_path_extra_roots setting covering the data directory

_resolve_search_root also returned the workspace root unchecked when the
path was empty, so a bare ls enumerated the directory whatever the deny
list said. It now resolves that case through the same guards.

A containment rule rather than a filename deny list, so state files added
later are covered without anyone remembering to list them, and so a user's
own settings.json or app.db inside a real workspace is not caught.

Four directories of user content stay readable, because the application
hands their paths to the model and tells it to open them: the chat upload
manifest, downloaded mail attachments, personal docs (which covers the
runbook) and personal uploads.

* fix: enforce state deny during recursive file search

* fix: bound protected filesystem searches

* fix(security): reject inode aliases and workspace redirects

* fix(security): harden partitioned agent searches

* fix(security): report fallback worker exits promptly

* fix(security): clean up search readers and retain relative data roots

---------

Co-authored-by: RaresKeY <158580472+RaresKeY@users.noreply.github.com>

launcher.py

@@ -14,6 +14,13 @@ import threading
 import time
 import webbrowser
 
+# PyInstaller multiprocessing children re-enter this executable with a private
+# bootstrap argument. Consume it before splash/UI or application imports so a
+# spawn-based worker does not relaunch the full desktop application.
+if __name__ == "__main__":
+    import multiprocessing
+    multiprocessing.freeze_support()
+
 # Define a dummy NullWriter to suppress standard stream crashes (isatty etc.) in GUI mode
 class NullWriter:
     def write(self, text):

setup.py

@@ -16,7 +16,7 @@ sys.path.insert(0, BASE_DIR)
 from src.constants import (
     DATA_DIR, AUTH_FILE, UPLOAD_DIR, PERSONAL_DIR, PERSONAL_UPLOADS_DIR,
     TTS_CACHE_DIR, GENERATED_IMAGES_DIR, DEEP_RESEARCH_DIR, CHROMA_DIR,
-    RAG_DIR, MEMORY_VECTORS_DIR, PASSWORD_MIN_LENGTH,
+    RAG_DIR, MEMORY_VECTORS_DIR, AGENT_WORKSPACE_DIR, PASSWORD_MIN_LENGTH,
 )
 from core.auth import RESERVED_USERNAMES
 
@@ -31,6 +31,7 @@ DIRS = [
     CHROMA_DIR,
     RAG_DIR,
     MEMORY_VECTORS_DIR,
+    AGENT_WORKSPACE_DIR,
     os.path.join(BASE_DIR, "logs"),
 ]
 

src/agent_tools/filesystem_tools.py

@@ -3,8 +3,8 @@ import json
 import os
 import re
 import difflib
-import fnmatch
 import shutil
+import time
 from typing import Optional, Dict, Any, Tuple, List
 
 from src.constants import MAX_READ_CHARS, MAX_DIFF_LINES, MAX_OUTPUT_CHARS
@@ -16,6 +16,8 @@ _CODENAV_SKIP_DIRS = frozenset({
 })
 _CODENAV_MAX_HITS = 200
 _CODENAV_MAX_LINE = 400
+_GREP_TIMEOUT_SECONDS = 20
+_GREP_STDERR_PREFIX = 20_000
 
 
 def _glob_to_regex(pat: str) -> "re.Pattern":
@@ -42,6 +44,113 @@ def _glob_to_regex(pat: str) -> "re.Pattern":
             i += 1
     return re.compile("".join(out))
 
+
+def _python_grep_worker(payload: dict, output_queue) -> None:
+    """Spawn-safe fallback grep worker used when ripgrep is unavailable.
+
+    Keep this at module scope: a frozen Windows executable cannot safely be
+    relaunched as ``sys.executable -c ...``, while multiprocessing can invoke a
+    top-level target through its frozen-process bootstrap.
+    """
+    try:
+        flags = re.IGNORECASE if payload["ignore_case"] else 0
+        try:
+            regex = re.compile(payload["pattern"], flags)
+            glob_regex = (
+                _glob_to_regex(payload["glob"].replace("\\", "/"))
+                if payload["glob"]
+                else None
+            )
+        except re.error as exc:
+            output_queue.put(("error", f"grep: bad pattern: {exc}"))
+            return
+
+        requested_root = payload["root"]
+        skip_dirs = set(payload["skip_dirs"])
+        sensitive = {name.casefold() for name in payload["sensitive_names"]}
+        max_hits = payload["max_hits"]
+        hits = 0
+
+        def within(path: str, root: str) -> bool:
+            try:
+                return os.path.commonpath(
+                    [os.path.normcase(path), os.path.normcase(root)]
+                ) == os.path.normcase(root)
+            except ValueError:
+                return False
+
+        def safe_file(path: str, target: str) -> Optional[str]:
+            if os.path.islink(path):
+                return None
+            canonical = os.path.realpath(path)
+            if not within(canonical, requested_root) or not within(canonical, target):
+                return None
+            parts = [part.casefold() for part in canonical.split(os.sep)]
+            if any(part in sensitive for part in parts):
+                return None
+            try:
+                if not os.path.isfile(canonical) or os.stat(canonical).st_nlink > 1:
+                    return None
+            except OSError:
+                return None
+            return canonical
+
+        for target in payload["targets"]:
+            if hits >= max_hits:
+                break
+            if os.path.isfile(target):
+                file_iter = iter((target,))
+            else:
+                def walk_files():
+                    for directory, dirnames, filenames in os.walk(
+                        target, followlinks=False
+                    ):
+                        dirnames[:] = [
+                            name
+                            for name in dirnames
+                            if name not in skip_dirs
+                            and name.casefold() not in sensitive
+                            and not os.path.islink(os.path.join(directory, name))
+                        ]
+                        for name in filenames:
+                            yield os.path.join(directory, name)
+
+                file_iter = walk_files()
+
+            for candidate in file_iter:
+                path = safe_file(candidate, target)
+                if path is None:
+                    continue
+                relative = os.path.relpath(path, requested_root).replace(os.sep, "/")
+                if glob_regex and not (
+                    glob_regex.fullmatch(relative)
+                    or glob_regex.fullmatch(os.path.basename(path))
+                ):
+                    continue
+                try:
+                    with open(path, "r", encoding="utf-8", errors="strict") as handle:
+                        for number, line in enumerate(handle, 1):
+                            if regex.search(line):
+                                output_queue.put((
+                                    "match",
+                                    path,
+                                    number,
+                                    line.rstrip()[:_CODENAV_MAX_LINE],
+                                ))
+                                hits += 1
+                                if hits >= max_hits:
+                                    break
+                except (UnicodeDecodeError, OSError):
+                    continue
+                if hits >= max_hits:
+                    break
+        output_queue.put(("done",))
+    except BaseException as exc:
+        try:
+            output_queue.put(("error", f"grep: fallback worker failed: {exc}"))
+        except BaseException:
+            pass
+
 def _unified_diff(old: str, new: str, path: str) -> Optional[Dict[str, Any]]:
     if old == new:
         return None
@@ -407,7 +516,11 @@ def _apply_patch_hunks(original: str, hunks: List[List[str]], label: str) -> str
 
 class LsTool:
     async def execute(self, content: str, ctx: dict) -> dict:
-        from src.tool_execution import _resolve_tool_path, _resolve_search_root, _truncate
+        from src.tool_execution import (
+            _is_denied_tool_path,
+            _resolve_search_root,
+            _truncate,
+        )
         raw_path = ""
         _s = (content or "").strip()
         if _s.startswith("{"):
@@ -431,6 +544,8 @@ class LsTool:
                     for entry in it:
                         if entry.name.startswith("."):
                             continue
+                        if _is_denied_tool_path(os.path.realpath(entry.path)):
+                            continue
                         try:
                             is_dir = entry.is_dir(follow_symlinks=False)
                             size = entry.stat(follow_symlinks=False).st_size if not is_dir else 0
@@ -458,7 +573,8 @@ class GlobTool:
     async def execute(self, content: str, ctx: dict) -> dict:
         from src.tool_execution import (
             _SENSITIVE_BASENAMES,
-            _is_sensitive_path,
+            _can_traverse_tool_path,
+            _is_denied_tool_path,
             _resolve_tool_path,
             _resolve_search_root,
             _truncate,
@@ -507,7 +623,7 @@ class GlobTool:
                 # .ssh/id_rsa, …) falls through to the walk, which skips it —
                 # otherwise glob would surface secret paths that read_file /
                 # grep already refuse to touch.
-                if inside and os.path.exists(cand) and not _is_sensitive_path(cand):
+                if inside and os.path.exists(cand) and not _is_denied_tool_path(cand):
                     return [cand], None
                 # Literal not at exact path — fall through to walk so
                 # e.g. "foo.py" still matches at any depth (like rglob).
@@ -517,13 +633,18 @@ class GlobTool:
             cap = _CODENAV_MAX_HITS * 5
             try:
                 for dp, dns, fns in os.walk(base):
+                    if not _can_traverse_tool_path(os.path.realpath(dp)):
+                        dns[:] = []
+                        continue
                     # Prune skipped dirs before descending (unlike rglob which
                     # descends first then filters — fatal on large node_modules).
                     # Sensitive dirs (.ssh, .gnupg, …) are pruned too so glob
                     # never enumerates the keys/tokens inside them.
                     dns[:] = [
                         d for d in dns
-                        if d not in _CODENAV_SKIP_DIRS and d not in _SENSITIVE_BASENAMES
+                        if d not in _CODENAV_SKIP_DIRS
+                        and d not in _SENSITIVE_BASENAMES
+                        and _can_traverse_tool_path(os.path.realpath(os.path.join(dp, d)))
                     ]
                     for name in fns + dns:
                         full = os.path.join(dp, name)
@@ -531,7 +652,7 @@ class GlobTool:
                         if regex.fullmatch(rel) or regex.fullmatch(name):
                             # Skip deny-listed sensitive files (.env, id_rsa,
                             # known_hosts, …) the same way grep does.
-                            if _is_sensitive_path(os.path.realpath(full)):
+                            if _is_denied_tool_path(os.path.realpath(full)):
                                 continue
                             try:
                                 mtime = os.stat(full).st_mtime
@@ -558,9 +679,12 @@ class GlobTool:
 class GrepTool:
     async def execute(self, content: str, ctx: dict) -> dict:
         from src.tool_execution import (
+            _SENSITIVE_BASENAMES,
             _SENSITIVE_FILE_PATTERNS,
+            _agent_readable_data_subdirs,
+            _is_denied_tool_path,
             _is_sensitive_path,
-            _resolve_tool_path,
+            _path_within,
             _resolve_search_root,
             _truncate,
         )
@@ -589,64 +713,307 @@ class GrepTool:
             return {"error": f"grep: {e}", "exit_code": 1}
 
         def _grep():
-            import re as _re
+            import multiprocessing
-            import shutil
+            import queue
+            import subprocess
+            import threading
+
+            from src.constants import DATA_DIR
+
             rg = shutil.which("rg")
-            if rg:
+            real_root = os.path.realpath(root)
-                cmd = [rg, "--line-number", "--no-heading", "--color=never",
+            data_dir = os.path.realpath(DATA_DIR)
-                       "--max-count", str(max_hits)]
+            spans_state = _path_within(data_dir, real_root)
-                if ignore_case:
+
-                    cmd.append("--ignore-case")
+            def is_top_level_safe(path: str, *, partition_generated: bool) -> bool:
-                if glob_pat:
+                lexical = os.path.abspath(path)
-                    cmd += ["--glob", glob_pat]
+                if os.path.islink(lexical):
-                # --iglob (not --glob) so the exclusion is case-insensitive:
+                    return False
-                # on a case-insensitive filesystem "ID_RSA"/"Known_Hosts"
+                canonical = os.path.realpath(lexical)
-                # resolve to the same secret as their lowercase forms, and the
+                if not _path_within(canonical, real_root):
-                # Python fallback below already folds case via _is_sensitive_path.
+                    return False
-                for _pat in _SENSITIVE_FILE_PATTERNS:
+                if partition_generated and os.path.basename(lexical) in _CODENAV_SKIP_DIRS:
-                    cmd += ["--iglob", f"!*{_pat}*"]
+                    return False
-                for _d in _CODENAV_SKIP_DIRS:
+                if _is_sensitive_path(canonical) or _is_denied_tool_path(canonical):
-                    cmd += ["--glob", f"!**/{_d}/**"]
+                    return False
-                cmd += ["--regexp", pattern, root]
+                return True
+
+            def safe_targets() -> tuple[list[str], Optional[str]]:
+                candidates: list[tuple[str, bool]] = []
+                if not spans_state:
+                    # Preserve direct-root compatibility: skip-directory policy
+                    # prunes descendants, but an explicitly requested allowed
+                    # root named node_modules remains searchable.
+                    candidates.append((real_root, False))
+                else:
+                    current = real_root
+                    if current != data_dir:
+                        for part in os.path.relpath(data_dir, current).split(os.sep):
+                            try:
+                                with os.scandir(current) as entries:
+                                    for entry in entries:
+                                        if entry.name != part:
+                                            # Reject a sibling link lexically before
+                                            # canonicalizing or treating it as a target.
+                                            if entry.is_symlink():
+                                                continue
+                                            candidates.append((entry.path, True))
+                            except OSError as exc:
+                                return [], f"grep: {exc}"
+                            current = os.path.join(current, part)
+                    for readable in _agent_readable_data_subdirs():
+                        if (
+                            _path_within(readable, data_dir)
+                            and _path_within(readable, real_root)
+                            and os.path.exists(readable)
+                        ):
+                            candidates.append((readable, True))
+
+                targets: list[str] = []
+                seen: set[str] = set()
+                for candidate, partition_generated in candidates:
+                    if not is_top_level_safe(
+                        candidate, partition_generated=partition_generated
+                    ):
+                        continue
+                    canonical = os.path.realpath(candidate)
+                    if canonical not in seen:
+                        seen.add(canonical)
+                        targets.append(canonical)
+                return targets, None
+
+            targets, target_error = safe_targets()
+            if target_error:
+                return None, target_error
+
+            base = real_root if os.path.isdir(real_root) else os.path.dirname(real_root)
+            deadline = time.monotonic() + _GREP_TIMEOUT_SECONDS
+            lines: list[str] = []
+
+            def parse_rg_result(raw: str) -> Optional[str]:
                 try:
-                    import subprocess
+                    record = json.loads(raw)
-                    p = subprocess.run(cmd, capture_output=True, text=True, timeout=20)
+                except (TypeError, json.JSONDecodeError):
-                    lines = [ln for ln in (p.stdout or "").splitlines() if ln][:max_hits]
+                    return None
-                    return lines, None
+                if record.get("type") != "match":
-                except subprocess.TimeoutExpired:
+                    return None
-                    return None, "grep: timed out"
+                data = record.get("data") or {}
-                except Exception as _e:
+                path = (data.get("path") or {}).get("text")
-                    return None, f"grep: {_e}"
+                text_value = (data.get("lines") or {}).get("text")
-            try:
+                number = data.get("line_number")
-                rx = _re.compile(pattern, _re.IGNORECASE if ignore_case else 0)
+                if not isinstance(path, str) or not isinstance(text_value, str):
-            except _re.error as _e:
+                    return None
-                return None, f"grep: bad pattern: {_e}"
+                absolute = path if os.path.isabs(path) else os.path.join(base, path)
-            hits = []
+                canonical = os.path.realpath(absolute)
-            if os.path.isfile(root):
+                if not _path_within(canonical, real_root) or _is_denied_tool_path(canonical):
-                file_iter = [root]
+                    return None
-            else:
+                return f"{os.path.abspath(absolute)}:{number}:{text_value.rstrip()[:_CODENAV_MAX_LINE]}"
-                file_iter = []
+
-                for dp, dns, fns in os.walk(root):
+            def run_rg(cmd: list[str]) -> Optional[str]:
-                    dns[:] = [d for d in dns if d not in _CODENAV_SKIP_DIRS]
+                try:
-                    for fn in fns:
+                    process = subprocess.Popen(
-                        if glob_pat and not fnmatch.fnmatch(fn, glob_pat):
+                        cmd,
+                        cwd=base,
+                        stdin=subprocess.DEVNULL,
+                        stdout=subprocess.PIPE,
+                        stderr=subprocess.PIPE,
+                        text=True,
+                        bufsize=1,
+                    )
+                except Exception as exc:
+                    return f"grep: {exc}"
+                output: queue.Queue[Optional[str]] = queue.Queue(maxsize=max_hits + 2)
+                stderr_prefix: list[str] = []
+                stderr_size = 0
+                stop_reader = threading.Event()
+
+                def enqueue_stdout(value: Optional[str]) -> bool:
+                    # The consumer stops at the result cap or deadline. Never
+                    # leave a producer blocked on its bounded queue afterward.
+                    while not stop_reader.is_set():
+                        try:
+                            output.put(value, timeout=0.05)
+                            return True
+                        except queue.Full:
                             continue
-                        file_iter.append(os.path.join(dp, fn))
+                    return False
-            for fp in file_iter:
+
-                if len(hits) >= max_hits:
+                def read_stdout() -> None:
-                    break
+                    assert process.stdout is not None
-                if _is_sensitive_path(os.path.realpath(fp)):
+                    try:
-                    continue
+                        for line in process.stdout:
+                            if not enqueue_stdout(line.rstrip("\n")):
+                                break
+                    finally:
+                        enqueue_stdout(None)
+
+                def read_stderr() -> None:
+                    nonlocal stderr_size
+                    assert process.stderr is not None
+                    while True:
+                        chunk = process.stderr.read(4096)
+                        if not chunk:
+                            break
+                        if stderr_size < _GREP_STDERR_PREFIX:
+                            kept = chunk[:_GREP_STDERR_PREFIX - stderr_size]
+                            stderr_prefix.append(kept)
+                            stderr_size += len(kept)
+
+                stdout_thread = threading.Thread(target=read_stdout, daemon=True)
+                stderr_thread = threading.Thread(target=read_stderr, daemon=True)
+                stdout_thread.start()
+                stderr_thread.start()
+                timed_out = False
+                capped = False
                 try:
-                    with open(fp, "r", encoding="utf-8", errors="strict") as f:
+                    while len(lines) < max_hits:
-                        for i, line in enumerate(f, 1):
+                        remaining = deadline - time.monotonic()
-                            if rx.search(line):
+                        if remaining <= 0:
-                                hits.append(f"{fp}:{i}:{line.rstrip()[:_CODENAV_MAX_LINE]}")
+                            timed_out = True
-                                if len(hits) >= max_hits:
+                            break
-                                    break
+                        try:
-                except (UnicodeDecodeError, OSError):
+                            raw = output.get(timeout=remaining)
-                    continue
+                        except queue.Empty:
-            return hits, None
+                            timed_out = True
+                            break
+                        if raw is None:
+                            break
+                        parsed = parse_rg_result(raw)
+                        if parsed and parsed not in lines:
+                            lines.append(parsed)
+                    capped = len(lines) >= max_hits
+                finally:
+                    stop_reader.set()
+                    if (timed_out or capped) and process.poll() is None:
+                        process.terminate()
+                    try:
+                        remaining = max(0.01, deadline - time.monotonic())
+                        return_code = process.wait(timeout=min(1, remaining))
+                    except subprocess.TimeoutExpired:
+                        process.kill()
+                        return_code = process.wait()
+                    stdout_thread.join()
+                    stderr_thread.join()
+                if timed_out:
+                    return "grep: timed out"
+                if not capped and return_code not in (0, 1):
+                    detail = "".join(stderr_prefix).strip()
+                    return f"grep: {detail or f'process exited {return_code}'}"
+                return None
+
+            if rg:
+                # Validate even when policy filtering leaves no search targets.
+                if not targets:
+                    error = run_rg([rg, "--json", "--no-config", "--regexp", pattern])
+                    return (None, error) if error else ([], None)
+                relative_targets = [os.path.relpath(target, base) for target in targets]
+                for offset in range(0, len(relative_targets), 128):
+                    if len(lines) >= max_hits:
+                        break
+                    cmd = [
+                        rg, "--json", "--no-config", "--no-follow",
+                        "--max-count", str(max_hits - len(lines)),
+                        "--max-columns", str(_CODENAV_MAX_LINE),
+                        "--max-columns-preview",
+                    ]
+                    if ignore_case:
+                        cmd.append("--ignore-case")
+                    if glob_pat:
+                        cmd += ["--glob", glob_pat]
+                    for sensitive_pattern in _SENSITIVE_FILE_PATTERNS:
+                        cmd += ["--iglob", f"!{sensitive_pattern}"]
+                    for skipped_dir in _CODENAV_SKIP_DIRS:
+                        cmd += ["--glob", f"!**/{skipped_dir}/**"]
+                    cmd += ["--regexp", pattern, "--", *relative_targets[offset:offset + 128]]
+                    error = run_rg(cmd)
+                    if error:
+                        return None, error
+                return lines, None
+
+            # This runs inside asyncio.to_thread(), so forking would clone a
+            # multithreaded process and can deadlock. Spawn is platform-safe and
+            # PyInstaller-compatible via launcher's early freeze_support().
+            payload = {
+                "root": real_root,
+                "targets": targets,
+                "pattern": pattern,
+                "ignore_case": ignore_case,
+                "glob": glob_pat,
+                "max_hits": max_hits,
+                "skip_dirs": tuple(_CODENAV_SKIP_DIRS),
+                "sensitive_names": tuple(
+                    set(_SENSITIVE_BASENAMES) | set(_SENSITIVE_FILE_PATTERNS)
+                ),
+            }
+            try:
+                context = multiprocessing.get_context("spawn")
+                output_queue = context.Queue(maxsize=max_hits + 2)
+                worker = context.Process(
+                    target=_python_grep_worker, args=(payload, output_queue)
+                )
+                worker.start()
+            except Exception as exc:
+                try:
+                    output_queue.close()
+                except (NameError, OSError, ValueError):
+                    pass
+                return None, f"grep: could not start fallback worker: {exc}"
+            error = None
+            completed = False
+            try:
+                while len(lines) < max_hits:
+                    remaining = deadline - time.monotonic()
+                    if remaining <= 0:
+                        error = "grep: timed out"
+                        break
+                    try:
+                        # Keep queue waits short enough to observe a spawn
+                        # worker that dies during bootstrap/import before it
+                        # can enqueue either an error or the done sentinel.
+                        record = output_queue.get(timeout=min(0.05, remaining))
+                    except queue.Empty:
+                        if worker.is_alive():
+                            continue
+                        worker.join(timeout=0)
+                        try:
+                            # A multiprocessing queue's feeder can make the
+                            # final record visible at process-exit time. Give
+                            # that record precedence over the exit status.
+                            remaining = deadline - time.monotonic()
+                            record = output_queue.get(
+                                timeout=min(0.05, max(0, remaining))
+                            )
+                        except queue.Empty:
+                            error = f"grep: fallback worker exited {worker.exitcode}"
+                            break
+                    if record[0] == "done":
+                        completed = True
+                        break
+                    if record[0] == "error":
+                        error = record[1]
+                        break
+                    _, path, number, text_value = record
+                    canonical = os.path.realpath(path)
+                    if not _path_within(canonical, real_root) or _is_denied_tool_path(canonical):
+                        continue
+                    rendered = f"{path}:{number}:{text_value}"
+                    if rendered not in lines:
+                        lines.append(rendered)
+            finally:
+                if completed:
+                    worker.join(timeout=min(1, max(0.01, deadline - time.monotonic())))
+                if worker.is_alive():
+                    worker.terminate()
+                    worker.join(timeout=1)
+                if worker.is_alive():
+                    worker.kill()
+                    worker.join()
+                output_queue.close()
+            if error:
+                return None, error
+            if worker.exitcode not in (0, None) and len(lines) < max_hits:
+                return None, f"grep: fallback worker exited {worker.exitcode}"
+            return lines, None
 
         lines, err = await asyncio.to_thread(_grep)
         if err:

src/app_initializer.py

@@ -2,10 +2,11 @@
 """Initialize all application components and dependencies."""
 import os
 import logging
+import stat
 from typing import Dict, Any
 
 from src.constants import (
-    DATA_DIR, PERSONAL_DIR, RUNBOOK_DIR, UPLOAD_DIR,
+    DATA_DIR, PERSONAL_DIR, RUNBOOK_DIR, UPLOAD_DIR, AGENT_WORKSPACE_DIR,
     SESSIONS_FILE, DEFAULT_HOST, OPENAI_API_KEY
 )
 from src.memory import MemoryManager
@@ -31,6 +32,34 @@ def create_directories():
     for directory in (DATA_DIR, PERSONAL_DIR, RUNBOOK_DIR, UPLOAD_DIR):
         os.makedirs(directory, exist_ok=True)
 
+    # The model-controlled workspace must be a real child of DATA_DIR.  Never
+    # follow a pre-existing symlink here: it would silently move the default
+    # native-file root outside the application volume before any resolver runs.
+    data_root = os.path.realpath(os.path.abspath(os.path.expanduser(DATA_DIR)))
+    workspace = os.path.abspath(os.path.expanduser(AGENT_WORKSPACE_DIR))
+    expected_workspace = os.path.join(data_root, "agent_workspace")
+    # Validate the real parent so a supported DATA_DIR bind/symlink works, but
+    # require the fixed internal carve-out name and reject a link at the model-
+    # controlled workspace entry itself.
+    if (
+        os.path.basename(workspace) != "agent_workspace"
+        or os.path.realpath(os.path.dirname(workspace)) != data_root
+    ):
+        raise RuntimeError("agent workspace must be the canonical child of DATA_DIR")
+    if os.path.lexists(workspace):
+        mode = os.lstat(workspace).st_mode
+        if stat.S_ISLNK(mode) or not stat.S_ISDIR(mode):
+            raise RuntimeError("agent workspace must be a real directory")
+    else:
+        os.mkdir(workspace, 0o700)
+    resolved_workspace = os.path.realpath(workspace)
+    if resolved_workspace != expected_workspace:
+        raise RuntimeError("agent workspace must be the canonical child of DATA_DIR")
+    try:
+        os.chmod(workspace, 0o700)
+    except OSError:
+        pass
+
 def initialize_managers(base_dir: str, rag_manager=None) -> Dict[str, Any]:
     """
     Initialize all manager and handler instances.

src/constants.py

@@ -54,6 +54,11 @@ GALLERY_DIR = os.path.join(DATA_DIR, "gallery")
 GALLERY_UPLOADS_DIR = os.path.join(DATA_DIR, "gallery_uploads")
 MEMORY_VECTORS_DIR = os.path.join(DATA_DIR, "memory_vectors")
 
+# The only part of DATA_DIR the agent's file tools and subprocesses may touch.
+# Everything else under DATA_DIR is application state (session store, auth
+# database, encryption key, settings), and the agent has no business reading it.
+AGENT_WORKSPACE_DIR = os.path.join(DATA_DIR, "agent_workspace")
+
 # Paths with an intentional dedicated env override, defaulting under DATA_DIR.
 MAIL_ATTACHMENTS_DIR = os.getenv("ODYSSEUS_MAIL_ATTACHMENTS_DIR", os.path.join(DATA_DIR, "mail-attachments"))
 # `or` (not os.getenv's default arg) so a PRESENT-but-EMPTY value falls back to

src/tool_execution.py

@@ -15,6 +15,7 @@ import logging
 import os
 import pathlib
 import re
+import stat
 import sys
 import time
 from typing import Any, Awaitable, Callable, Dict, Optional, Tuple
@@ -30,7 +31,12 @@ from src.tool_security import (
 from src.tool_capabilities import ToolRunSecurityContext, blocked_tool_result
 from src.tool_approvals import ExactToolApproval
 from src.tool_policy import ToolPolicy
-from src.constants import MAX_OUTPUT_CHARS, MAX_READ_CHARS, MAX_DIFF_LINES, DATA_DIR
+from src.constants import (
+    MAX_OUTPUT_CHARS,
+    MAX_READ_CHARS,
+    MAX_DIFF_LINES,
+    AGENT_WORKSPACE_DIR,
+)
 from src.tool_utils import _truncate, get_mcp_manager
 
 
@@ -46,11 +52,11 @@ _MISSING_TOOL_SECURITY_CONTEXT = _MissingToolSecurityContext()
 NO_TOOL_SECURITY_CONTEXT = _NoToolSecurityContext()
 
 # Persistent working directory for agent subprocesses.
-# Resolves to <repo_root>/data, which is the bind-mounted volume in Docker
+# Resolves to <repo_root>/data/agent_workspace, inside the bind-mounted volume
-# (/app/data) and the local data directory for manual installs.
+# in Docker (/app/data), so files survive a rebuild as before. The subdirectory
-# Using this as cwd and HOME prevents the agent from silently creating files
+# rather than data/ itself keeps agent scratch files and dotfiles out of the
-# in ephemeral container layers that are lost on the next rebuild.
+# directory holding the session store and the auth database.
-_AGENT_WORKDIR = DATA_DIR
+_AGENT_WORKDIR = AGENT_WORKSPACE_DIR
 
 
 
@@ -66,10 +72,15 @@ _AGENT_WORKDIR = DATA_DIR
 #   1. Sensitive-subpath deny list — checked FIRST. Blocks .ssh,
 #      .gnupg, shell rc files, token/env files even if the root above
 #      them is on the allowlist.
-#   2. Allowlist — only the directories the agent legitimately needs
+#   2. Application-state deny (_is_app_state_path) - DATA_DIR holds the
-#      (project data/, system tmp). $HOME is NOT on the default list.
+#      session store, auth database, app key and settings, so only
-#   3. Opt-in extra roots — admin can add broader roots via the
+#      _agent_readable_data_subdirs() is readable inside it.
-#      "tool_path_extra_roots" setting (list of path strings).
+#   3. Allowlist - only the directories the agent legitimately needs
+#      (its data/ workspace, user content, system tmp). $HOME is NOT on
+#      the default list.
+#   4. Opt-in extra roots - admin can add broader roots via the
+#      "tool_path_extra_roots" setting. These cannot re-open DATA_DIR;
+#      rule 2 is independent of which root a path arrived through.
 # ---------------------------------------------------------------------------
 
 _SENSITIVE_BASENAMES: set[str] = {
@@ -116,6 +127,184 @@ def _is_sensitive_path(resolved: str) -> bool:
     return filename in _SENSITIVE_FILE_PATTERNS_CF
 
 
+def _path_within(resolved: str, root: str) -> bool:
+    """True when *resolved* is *root* itself or sits underneath it.
+
+    Use the platform's path-case rules.  This helper participates in allow
+    decisions, so unconditional case-folding would let a distinct ``/DATA``
+    tree masquerade as a descendant of ``/data`` on case-sensitive systems.
+    """
+    resolved, root = os.path.normcase(resolved), os.path.normcase(root)
+    if resolved == root:
+        return True
+    try:
+        if os.path.commonpath([resolved, root]) == root:
+            return True
+    except ValueError:
+        return False
+    # normcase is intentionally conservative about assumptions (notably on
+    # POSIX), so consult the filesystem when paths exist.  This recognizes a
+    # case alias on a case-insensitive volume without treating distinct
+    # case-sensitive paths as the same allow root.
+    if os.path.exists(root):
+        candidate = resolved
+        while True:
+            try:
+                if os.path.exists(candidate) and os.path.samefile(candidate, root):
+                    return True
+            except OSError:
+                pass
+            parent = os.path.dirname(candidate)
+            if parent == candidate:
+                break
+            candidate = parent
+    return False
+
+
+def _path_within_conservative(resolved: str, root: str) -> bool:
+    """Containment for deny decisions, folding case to fail closed."""
+    resolved, root = resolved.casefold(), root.casefold()
+    if resolved == root:
+        return True
+    try:
+        return os.path.commonpath([resolved, root]) == root
+    except ValueError:
+        return False
+
+
+def _agent_readable_data_subdirs() -> tuple[str, ...]:
+    """The only parts of DATA_DIR the agent's file tools may reach.
+
+    The agent's own scratch folder, plus the directories of user content whose
+    paths the application itself gives to the model, which it would then be
+    unable to open.  These normally live under DATA_DIR; the documented mail
+    attachment override may instead name a disjoint external directory:
+
+      UPLOAD_DIR            the chat upload manifest renders "path=<p>" and
+                            says to read it with read_file (agent_loop.py)
+      MAIL_ATTACHMENTS_DIR  download_attachment returns the path and its own
+                            description tells the model to read it
+      PERSONAL_DIR          GET /api/personal returns a path per file and is
+                            reachable through the app_api tool; RUNBOOK_DIR
+                            nests under it
+      PERSONAL_UPLOADS_DIR  indexed as a personal-docs directory, which
+                            manage_rag lists as an absolute path
+
+    Order matters: the first entry is roots[0], which _resolve_search_root uses
+    when grep/glob/ls are called with no path.
+    """
+    from src.constants import (
+        DATA_DIR,
+        MAIL_ATTACHMENTS_DIR,
+        PERSONAL_DIR,
+        PERSONAL_UPLOADS_DIR,
+        UPLOAD_DIR,
+    )
+    configured = (
+        (AGENT_WORKSPACE_DIR, "agent_workspace", False),
+        (UPLOAD_DIR, "uploads", False),
+        # This has a documented environment override and may legitimately
+        # live outside DATA_DIR, but it must never equal/contain DATA_DIR.
+        (MAIL_ATTACHMENTS_DIR, "mail-attachments", True),
+        (PERSONAL_DIR, "personal_docs", False),
+        (PERSONAL_UPLOADS_DIR, "personal_uploads", False),
+    )
+    configured_data_dir = os.path.abspath(os.path.expanduser(str(DATA_DIR)))
+    data_dir = os.path.realpath(configured_data_dir)
+    safe: list[str] = []
+    for raw, internal_name, external_ok in configured:
+        value = str(raw or "").strip()
+        # These paths are security-policy roots, not ordinary allowlist
+        # entries. Internal roles may inherit a relative DATA_DIR, but must
+        # still resolve to their exact canonical child below. External mail
+        # overrides require an absolute, disjoint directory.
+        if not value:
+            continue
+        expanded = os.path.abspath(os.path.expanduser(value))
+        # A policy root must not acquire an exemption by redirecting its final
+        # path component to protected state or to an unrelated external tree.
+        if os.path.islink(expanded):
+            continue
+        resolved = os.path.realpath(expanded)
+        if os.path.exists(resolved) and not os.path.isdir(resolved):
+            continue
+        expected_internal = os.path.join(data_dir, internal_name)
+        expected_configured = os.path.join(configured_data_dir, internal_name)
+        inside_data = (
+            os.path.normcase(expanded)
+            in {
+                os.path.normcase(expected_configured),
+                os.path.normcase(expected_internal),
+            }
+            and resolved == expected_internal
+        )
+        external_safe = (
+            external_ok
+            and os.path.isabs(os.path.expanduser(value))
+            and resolved != data_dir
+            and os.path.dirname(resolved) != resolved
+            and not _path_within(data_dir, resolved)
+            and not _path_within(resolved, data_dir)
+        )
+        if not (inside_data or external_safe) or _is_sensitive_path(resolved):
+            continue
+        safe.append(resolved)
+    return tuple(safe)
+
+
+def _is_app_state_path(resolved: str) -> bool:
+    """True for anything under DATA_DIR that is not agent-readable.
+
+    DATA_DIR holds the session store, the auth database, the app encryption key
+    and the settings file. A model-supplied path must not reach those through
+    any root, so this is checked in both resolvers rather than expressed as an
+    absence from the allowlist: a workspace bound at or above the data
+    directory, or an opt-in tool_path_extra_roots entry covering it, would
+    otherwise put them back in reach.
+
+    A containment rule rather than a filename deny list, so state files added
+    later are covered without anyone remembering to list them, and so a user's
+    own settings.json or app.db inside a real workspace is not caught.
+    """
+    from src.constants import DATA_DIR
+    if not _path_within_conservative(resolved, os.path.realpath(DATA_DIR)):
+        return False
+    return not any(
+        _path_within(resolved, d)
+        for d in _agent_readable_data_subdirs()
+    )
+
+
+def _is_hardlinked_regular_file(resolved: str) -> bool:
+    """Reject inode aliases that can smuggle DATA_DIR state into an allow root."""
+    try:
+        target = os.stat(resolved, follow_symlinks=False)
+    except OSError:
+        return False
+    return stat.S_ISREG(target.st_mode) and getattr(target, "st_nlink", 1) > 1
+
+
+def _is_denied_tool_path(resolved: str) -> bool:
+    """Apply every path deny to a canonical traversal result."""
+    return (
+        _is_sensitive_path(resolved)
+        or _is_app_state_path(resolved)
+        or _is_hardlinked_regular_file(resolved)
+    )
+
+
+def _can_traverse_tool_path(resolved: str) -> bool:
+    """Allow walking a denied state parent only to reach safe carve-outs."""
+    if _is_sensitive_path(resolved):
+        return False
+    if not _is_app_state_path(resolved):
+        return True
+    return any(
+        _path_within(readable, resolved)
+        for readable in _agent_readable_data_subdirs()
+    )
+
+
 def _tool_path_roots() -> list[str]:
     """Return the list of directory roots that read_file / write_file
     may touch. Default: project data/ + system temp dirs. Extra roots
@@ -123,9 +312,9 @@ def _tool_path_roots() -> list[str]:
     """
     roots: list[str] = []
 
-    # Project data directory — the agent's primary workspace.
+    # The agent's workspace plus the user-content directories inside data/.
-    from src.constants import DATA_DIR
+    # The rest of DATA_DIR is denied by _is_app_state_path.
-    roots.append(DATA_DIR)
+    roots.extend(_agent_readable_data_subdirs())
 
     # /tmp (and its macOS realpath /private/tmp).
     roots.append("/tmp")
@@ -193,6 +382,12 @@ def _resolve_tool_path(raw_path: str) -> str:
             f"path '{raw_path}' is inside a sensitive directory "
             f"(e.g. .ssh, .gnupg) or matches a sensitive filename"
         )
+    if _is_app_state_path(resolved):
+        raise ValueError(
+            f"path '{raw_path}' is inside the application state directory"
+        )
+    if _is_hardlinked_regular_file(resolved):
+        raise ValueError(f"path '{raw_path}' is a hard-linked file")
 
     for root in _tool_path_roots():
         if resolved == root:
@@ -228,6 +423,12 @@ def _resolve_tool_path_in_workspace(workspace: str, raw_path: str) -> str:
             f"path '{raw_path}' is inside a sensitive directory "
             f"(e.g. .ssh, .gnupg) or matches a sensitive filename"
         )
+    if _is_app_state_path(resolved):
+        raise ValueError(
+            f"path '{raw_path}' is inside the application state directory"
+        )
+    if _is_hardlinked_regular_file(resolved):
+        raise ValueError(f"path '{raw_path}' is a hard-linked file")
     if resolved != base:
         # normcase so containment holds on case-insensitive filesystems
         # (Windows, default macOS): it lowercases on Windows and is a no-op on
@@ -277,6 +478,10 @@ def vet_workspace(raw: str) -> Optional[str]:
     resolved = os.path.realpath(os.path.expanduser(raw))
     if not os.path.isdir(resolved) or _is_sensitive_path(resolved):
         return None
+    # Refuse the bind rather than binding a workspace where every subsequent
+    # tool call would fail on the same deny list.
+    if _is_app_state_path(resolved):
+        return None
     # Reject filesystem roots: binding / (or a Windows drive/UNC root) as the
     # workspace would make every absolute path "inside" it, collapsing the
     # confinement into host-wide file access. A root is its own dirname, which
@@ -289,7 +494,13 @@ def vet_workspace(raw: str) -> Optional[str]:
 def agent_cwd() -> str:
     """Working directory for agent subprocesses (bash/python/background jobs):
     the active workspace when set, else the persistent data dir."""
-    return get_active_workspace() or _AGENT_WORKDIR
+    workspace = get_active_workspace()
+    if workspace:
+        return workspace
+    resolved = os.path.realpath(_AGENT_WORKDIR)
+    if resolved not in _agent_readable_data_subdirs():
+        raise RuntimeError("agent workspace is not a safe real directory")
+    return resolved
 
 
 def get_mcp_manager():
@@ -304,16 +515,22 @@ def _resolve_search_root(raw_path: str) -> str:
 
     With a workspace active, the workspace folder is the root and a supplied
     path is confined inside it. Otherwise an empty path defaults to the agent's
-    primary root (project data dir) and a supplied path is confined by the
+    primary root (its workspace under the project data dir) and a supplied path
-    global allowlist + sensitive-file policy.
+    is confined by the global allowlist + sensitive-file policy.
     """
     raw = (raw_path or "").strip()
     ws = get_active_workspace()
     if ws:
-        return os.path.realpath(ws) if not raw else _resolve_tool_path_in_workspace(ws, raw)
+        # Resolve the empty case as the workspace path rather than returning
+        # it directly: returned unchecked it skipped both deny lists, so a
+        # bare ls listed whatever the workspace was bound to.
+        return _resolve_tool_path_in_workspace(ws, raw or ws)
     if not raw:
         roots = _tool_path_roots()
-        return roots[0] if roots else os.path.realpath(".")
+        default_root = os.path.realpath(AGENT_WORKSPACE_DIR)
+        if default_root in roots and not _is_denied_tool_path(default_root):
+            return default_root
+        raise ValueError("default agent workspace is not a safe readable data subdirectory")
     return _resolve_tool_path(raw)
 
 logger = logging.getLogger(__name__)

tests/test_code_nav_tools.py

@@ -91,6 +91,17 @@ def test_grep_python_fallback_when_no_rg(repo, monkeypatch):
     assert ".git/config" not in r["output"]
 
 
+def test_grep_python_fallback_uses_relative_glob_paths(repo, monkeypatch):
+    monkeypatch.setattr(shutil, "which", lambda name: None)
+    r = _run(
+        "grep",
+        f'{{"pattern": "needle|python", "glob": "**/*.py", "path": "{repo}"}}',
+    )
+    assert r["exit_code"] == 0
+    assert "a.py" in r["output"]
+    assert "sub/deep/c.py" in r["output"]
+
+
 @pytest.mark.skipif(shutil.which("rg") is None, reason="targets the ripgrep fast-path")
 def test_grep_skips_case_variant_sensitive_files_rg(repo):
     """The rg fast-path must exclude deny-listed key files case-insensitively.

tests/test_launcher.py

@@ -1,12 +1,22 @@
 # tests/test_launcher.py
 import sys
 import os
+from pathlib import Path
 from unittest import mock
 import pytest
 
 from launcher import NullWriter, create_tray_image, on_open_browser, on_exit, open_browser
 
 
+def test_frozen_multiprocessing_bootstrap_precedes_gui_and_app_imports():
+    source = Path("launcher.py").read_text(encoding="utf-8")
+
+    freeze = source.index("multiprocessing.freeze_support()")
+    splash = source.index("if getattr(sys, 'frozen', False):")
+    app_import = source.index("from app import app")
+    assert freeze < splash < app_import
+
+
 def test_null_writer():
     writer = NullWriter()
     # writing and flushing should not raise any exceptions

tests/test_tool_path_confinement.py

@@ -161,12 +161,14 @@ def test_blocks_netrc():
         _resolve_tool_path("~/.netrc")
 
 
-def test_allows_project_data(tmp_path):
+def test_allows_agent_workspace(tmp_path):
-    """Paths under project data/ must resolve cleanly."""
+    """Paths under the agent's workspace in project data/ must resolve
+    cleanly. The rest of data/ is application state and is rejected;
+    tests/test_agent_state_dir_confinement.py covers that side."""
     from src.tool_execution import _resolve_tool_path
-    from src.constants import DATA_DIR
+    from src.constants import AGENT_WORKSPACE_DIR
-    target = os.path.join(DATA_DIR, "test-confinement-ok.txt")
+    target = os.path.join(AGENT_WORKSPACE_DIR, "test-confinement-ok.txt")
-    os.makedirs(DATA_DIR, exist_ok=True)
+    os.makedirs(AGENT_WORKSPACE_DIR, exist_ok=True)
     with open(target, "w") as f:
         f.write("ok")
     try: