Bug Description

With terminal.persistent_shell: true (the default), a single tool call that deletes the shell's current working directory wedges the terminal environment for the rest of the gateway's lifetime. Every subsequent subprocess.Popen raises FileNotFoundError: [Errno 2] on the cwd kwarg before the requested command runs.

Same root-cause class as #6607 and #4982, but in tools/environments/local.py — the executor for every shell command run by the agent.

Environment

• Hermes Agent: v0.11.0 (2026.4.23), commit ff687c01 (v2026.4.23-948-gff687c01)
• OS: Linux (Ubuntu 24.04, kernel 6.17)
• Python: 3.12.13
• Terminal backend: local, persistent_shell: true, lifetime_seconds: 300

Related

• [Bug]: checkpoint_manager subprocess.run cwd can point to non-existent directory, causing FileNotFoundError #6607 — same root-cause class in checkpoint_manager._run_git
• fix: guard os.getcwd() against deleted CWD (FileNotFoundError) #4982 — same root-cause class in cli.py via os.getcwd()
• [Bug]: terminal tool passes invalid workdir to Docker exec and overrides configured terminal.cwd #4669 — workdir validation for the docker terminal backend
• Skill auto-creation learns from transient failures, causing persistent tool avoidance (learned helplessness) #6051 — downstream skill-curator damage when this wedge is misdiagnosed as an environmental restriction

Steps to Reproduce

In a single agent session with the local terminal backend:

mkdir -p /tmp/wedge-repro
cd /tmp/wedge-repro
rm -rf /tmp/wedge-repro
pwd

The final pwd fails with FileNotFoundError: [Errno 2] No such file or directory: '/tmp/wedge-repro'. Every terminal/file-tool call after this point fails the same way until the gateway is restarted.

Expected Behavior

A command leaving self.cwd at a non-existent path should not block all subsequent commands.

Actual Behavior

Observed Behavior

~/.hermes/logs/errors.log:

ERROR tools.terminal_tool: Execution failed after 3 retries -
  Command: pwd - Error: FileNotFoundError: [Errno 2] No such file or directory:
  '/tmp/wedge-repro' - Task: default, Backend: local

The path in the error is the cwd kwarg passed to Popen, not the command target. Stack trace:

File ".../tools/file_operations.py", line 383, in _exec
    result = self.env.execute(command, cwd=effective_cwd, **kwargs)
File ".../tools/environments/base.py", line 758, in execute
    proc = self._run_bash(...)
File ".../tools/environments/local.py", line 356, in _run_bash
    proc = subprocess.Popen(args, ..., cwd=self.cwd)
FileNotFoundError: [Errno 2] No such file or directory: '/tmp/wedge-repro'

terminal_tool retries 3× with ~5s timeout each, so each failure takes ~14–15s of wall clock.

Recovery requires restarting the gateway. There is no in-process recovery: cd / and similar fail with the same error because they too are Popen calls with the dead cwd.

Affected Component

Tools (terminal, file ops, web, code execution, etc.)

Messaging Platform (if gateway-related)

N/A (CLI only)

Debug Report

Ubuntu 24.04

Python Version

3.12.13

Hermes Version

0.11.0

Additional Logs / Traceback (optional)

Root Cause Analysis (optional)

Root Cause

In tools/environments/local.py:

• Line 356–367: _run_bash passes cwd=self.cwd to subprocess.Popen.

• Line 392–400: _update_cwd assigns self.cwd = cwd_path from a post-command marker file with no existence check:

  def _update_cwd(self, result: dict):
      try:
          with open(self._cwd_file) as f:
              cwd_path = f.read().strip()
          if cwd_path:
              self.cwd = cwd_path     # no Path(cwd_path).is_dir() guard
      except (OSError, FileNotFoundError):
          pass

When bash deletes its own cwd, it stays alive with an unlinked directory. The marker writes back the now-deleted path, self.cwd is updated to it, and every subsequent Popen fails with ENOENT before exec.

Proposed Fix (optional)

Workaround

Avoid cd <dir> followed by rm -rf <dir> in the same persistent_shell session. Use absolute paths for delete, or cd .. first. Or set terminal.persistent_shell: false.

Possible Fixes

Two possible fixes:

1. Validate before Popen in _run_bash: if Path(self.cwd).is_dir() is false, fall back to a known-good cwd (~, or terminal.cwd from config) and log a warning. Mirrors _safe_getcwd() in fix: guard os.getcwd() against deleted CWD (FileNotFoundError) #4982.

2. Validate at update time in _update_cwd: only assign self.cwd = cwd_path if Path(cwd_path).is_dir(). Keep the previous cwd otherwise.

(1) also handles the case where the cwd is deleted out-of-band by another process.

Are you willing to submit a PR for this?

• I'd like to fix this myself and submit a PR