Problem

When multiple gateway sessions share a single LocalEnvironment (the default — all subagents map to "default" task_id), a subagent that cd's into a temp directory poisons self.cwd for ALL sessions if that directory is later deleted.

subprocess.Popen(cwd='/tmp/deleted_dir') raises FileNotFoundError at the Python level — before bash even starts — making terminal and file tools completely unusable across all concurrent sessions until gateway restart.

The retry logic (3 retries with exponential backoff) just retries the same broken Popen call, making it worse by delaying the error response.

Reproduction scenario (from a user's debug report):

1. Gateway running with Telegram, spawning delegate_task subagents
2. Subagent cd's into /tmp/dr3_4_a1_46hd1m5g (a temp sandbox)
3. Temp dir deleted by script cleanup
4. ALL sessions get: FileNotFoundError: [Errno 2] No such file or directory: '/tmp/dr3_4_a1_46hd1m5g'
5. Every terminal/file operation fails with 3 retries each — gateway is bricked

Fix

Two changes in tools/environments/local.py:

1. _run_bash() — validate cwd before Popen

Before passing cwd to subprocess.Popen, check os.path.isdir(self.cwd). If the directory no longer exists, reset self.cwd to user's home (or /) as a safe fallback. The shell-level cd -- <path> in _wrap_command handles the logical directory switch — Popen's cwd just needs to be a valid launch point for bash.

2. _update_cwd() — don't restore stale paths from cwd file

After a command runs, _update_cwd() reads the cwd tracking file. If the recorded path no longer exists on disk, reset to fallback instead of re-setting self.cwd to the stale path. This prevents the exit-126 loop where every command fails because the stale path keeps getting re-read from the file.

Recovery behavior

After the fix:

• First command after deletion: returns exit 126 with clear No such file or directory message (model can understand and adapt)
• All subsequent commands: succeed from home directory (self-healing)
• No Python crash, no retry storms, no gateway bricking

Test plan

E2E verified:

env = LocalEnvironment(cwd=temp_dir, timeout=30)
env.execute('echo hello')  # works
shutil.rmtree(temp_dir)    # simulate cleanup
env.execute('echo test')   # rc=126, no crash, cwd resets to ~
env.execute('echo ok')     # rc=0, recovered

Existing test suite: 50 passed (test_base_environment + test_local_env_blocklist + test_terminal_tool), 973 passed in full tools/ suite (2 pre-existing failures unrelated to this change).