Job Manager

A MoleditPy plugin for running calculations over SSH, locally, inside WSL, or through a LAN Job Server. Upload inputs, submit jobs, monitor progress, fetch results and open them in the editor. It also runs as a standalone desktop application.

AI agents can submit calculations through the local HTTP API, monitor jobs and retrieve results. Multiple agents can feed the same host-side helper queue, which schedules jobs within configured CPU and memory budgets to help prevent overloading a workstation.

Features

• Submit calculations to SLURM, PBS/Torque, SGE/UGE, a background process, or MoleditPy Job Server. Native Windows jobs use PowerShell; WSL jobs use bash inside the selected Linux distribution.
• Manage workstation resources with a built-in queue that schedules by physical cores and memory. Pause the queue, change its limits, chain jobs, or schedule a later start. Jobs continue with MoleditPy closed.
• Reuse inputs and settings. Save command templates and resource presets, submit files already on the host, or relay a predecessor's outputs into a new job. ORCA, Gaussian, Psi4, NWChem, Q-Chem and GAMESS inputs can supply their own core and memory requests.
• Monitor jobs and hosts. Sort and filter jobs, browse remote files and logs, and view CPU, memory and load in the Host Monitor. Desktop notifications, task-bar progress, a tray icon and Slack/Discord/Teams webhooks are available.
• Collect results. Download outputs automatically beside the input file and open them through the installed result analyzer. Save job lists as SQLite .pmejdb or legacy .pmejbs, export CSV, or rebuild a read-only list from existing result folders.
• Submit from AI agents and other tools. The optional local HTTP API lets agents and scripts submit jobs, inspect status and logs, and retrieve results. Agents using the same helper queue share its resource budgets; jobs wait until their requested cores and memory fit. ORCA and Gaussian Input Generator Pro provide a Submit to Cluster... button.

Instance discovery writes one small registration file on startup and removes it on shutdown. Liveness checks use an in-memory loopback socket, with no periodic heartbeat writes. Window activation requests write only when requested. After updating from 3.3.0 or earlier, quit all existing Job Manager monitors and tray processes before restarting; the old file-heartbeat protocol is no longer used.

The file Text Viewer has an Open in MoleditPy button. Passing the source file to the editor closes the viewer and its originating File Browser; cancelling or refusing the open leaves both windows available. Plain .txt files retain the existing protection against OpenBabel's text-to-molecules import.

Job Server connections reuse expiring authentication state in memory across monitoring and file operations, avoiding a new authentication record on every poll. Credentials and certificate trust remain checked; bearer tokens are not saved to disk.

Detailed instructions: Standard workflow.

Requirements

• MoleditPy 5.x, or Python 3.10+ with PySide6-Essentials for standalone use.
• SSH destinations: an SSH client for OpenSSH, or pip install paramiko for the optional backend with password authentication.
• Local destinations: bash, or the Windows scheduler with PowerShell.
• WSL destinations: Windows with a WSL distribution that contains bash.
• Job Server destinations: the optional cryptography>=50.0.2 package. The HTTP client is bundled; see Job Server setup.

Usage

1. Open Extensions → Job Manager → Job Monitor, then Hosts.... Add a host, choose its backend and scheduler, configure its working directory, and press Test Connection.
2. Select New Job..., add inputs, set resources, and choose a command from Template.... Check Script preview before submitting. Save a preset to reuse the settings.
3. Watch the job table or open its log. Completed outputs can be downloaded automatically and opened in the appropriate result analyzer.

Drop an input file onto the Job Monitor to open a prefilled submission. Dropping several files creates one job per file; hold Shift to submit them together. For staged inputs, select Work already on the host and check the existing remote directory. Command-only jobs need no input file.

Inputs must not use the same names as Job Manager's execution wrapper, log or completion files. Conflicts are reported before uploading. Fetching a file that is no longer available on the host reports an error instead of showing an old cached copy. Downloading a running job keeps its status monitoring active.

Command templates and file paths

Templates cover ORCA, Gaussian, CP2K, GAMESS, MOPAC, NWChem, Psi4, PySCF, Quantum ESPRESSO, VASP and xTB. Use the full executable path when needed, especially for ORCA: a bare orca can resolve to the Linux screen reader.

Placeholders accept either {input} or [input]. Available fields include input, stem, basename, output, nodes, ntasks, cpus, memory, walltime and queue. Save custom commands through Template... → Save current command as....

The input table shows File, Folder and Status. Add or paste paths, or drag files onto the dialog. Move up / Move down selects the first input used by {input}. Inputs with the same upload filename must be renamed or submitted separately.

See command placeholders and the standard workflow for examples and program-specific requirements.

Backends

OpenSSH uses batch mode so a background operation cannot wait for an invisible password prompt. For password-only hosts, choose paramiko and enable Ask for a password when connecting. SSH passwords stay in session memory and are never written to disk. Unknown host keys require fingerprint approval.

For WSL, choose the distribution in Hosts... and use a Linux Remote root such as ~/moleditpy_jobs. Inputs are copied into the distribution and results back to Windows. Files within the selected distribution can also be supplied through \\wsl.localhost\Ubuntu\... paths.

See WSL setup, Job Server setup, and the security model.

Verified backends

These configurations have been checked with real jobs through submission, completion tracking, log reading and result download. Local and WSL checks cover both direct background mode and the built-in helper queue.

The Ubuntu checks include Windows input files, \\wsl.localhost paths, existing job directories, failed-command exit codes and jobs that continue after submission returns. Version 3.0.0 fixes a WSL launch race that could leave a job pending without a log. Real WSL tests run when a distribution with bash is available; Windows CI also tests the transport through a shell-backed stand-in.

OpenSSH and paramiko are verified against real hosts; their automated transport tests additionally use mocked connections. A separate, manually dispatched real scheduler workflow provisions SLURM, OpenPBS and Grid Engine daemons in disposable environments and exercises them through real SSH, using synthetic inputs. It covers submission, result transfer, CPU allocation, cancellation, dependencies and scheduled starts. It does not run on routine pushes, PRs or releases.

Verified with real scheduler daemons on 2026-10-25: all cases passed on SLURM 21.08.5, OpenPBS 23.06.06, and Debian Grid Engine 8.1.9. These used synthetic jobs, real SSH and single-node disposable environments. Single-node lab results do not certify every Torque/commercial UGE version, multinode MPI/GPU setup or site-specific policy.

Scheduling and monitoring

On hosts without a cluster scheduler, the helper queue dispatches jobs within physical-core and memory budgets. It runs only while work remains, keeps its queue in readable files on the host, and supports dependencies, pausing and scheduled starts. A chained-lanes mode is also available.

Cluster jobs use the scheduler's own dependency and submission flags. Configure Submit options on the host or preset for billing groups, resource groups or other arguments that belong on the sbatch/qsub command line. Host options come first; arguments are parsed and quoted individually.

Polling uses one queue-status request per host per cycle. The default is 120 seconds, with a minimum of 5 seconds and a warning below 30 seconds. Polling stops when no jobs are active; connection failures trigger backoff. Use a longer per-host interval for shared login nodes and disable or slow host CPU/memory sampling where appropriate. Refresh Now requests an update; View → Reload List (Ctrl+R) rereads the saved job list.

Generated wrappers write an exit-code sentinel atomically when they exit: zero means DONE, a nonzero code means FAILED, and a missing sentinel after the job disappears means LOST. Signal traps record termination codes such as 143 for SIGTERM. Jobs blocked by a failed dependency show BLOCKED.

See the helper queue, job states, and troubleshooting.

Dependencies form a chain, with one predecessor per job. There is no workflow graph or automatic retry; complex cluster dependencies can be supplied through extra scheduler directives. Job Server has its own queue and feature limits, described in its setup guide.

Standalone mode

Launch the full Job Manager without opening MoleditPy, or without installing it at all — see Job Manager without MoleditPy:

pip install moleditpy-job-manager
moleditpy-job-manager
moleditpy-job-manager --host-monitor

From the plugin folder or the release zip, python -m job_manager does the same.

Alternatively, run python __main__.py inside the plugin folder. On Windows, use launch_job_manager.bat or launch_host_monitor.bat; create a desktop shortcut to either file for one-click access. They use pythonw to avoid a persistent console window.

Standalone uses the same hosts, job list, downloads and notifications as the plugin. Opening results in the molecular editor requires MoleditPy.

A second standalone launch brings the existing Job Manager to the front instead of starting another tracker. In Settings..., Keep the Job Manager running after MoleditPy closes can leave it in the tray to poll, notify and download. This option is off by default.

Integrations

AI agents and automation

An AI agent running on the same machine as Job Manager can use the HTTP API, Python client or command-line client to submit calculations without opening the submission dialog. It can list configured hosts and presets, check host load and queue status, submit jobs, read logs, cancel jobs and download results.

For multiple agents targeting one workstation, use the same host profile with Queueing set to Queue them with a helper on the host. Configure its Cores available and Memory available budgets, and have each agent declare its job's CPU and memory requests. The shared queue starts jobs only when those requests fit the remaining budgets; other jobs wait. For example, with an 8-core, 32 GB budget, jobs requesting 4 cores and 16 GB each run two at a time, even when several agents submit work at once.

These budgets account for declared requests, rather than enforcing process CPU or memory limits. Set requests to match the calculation's actual usage, leave capacity for other applications, and avoid Force run, which bypasses the budgets. This coordination applies to jobs in the same helper queue; SLURM, PBS and SGE destinations use their scheduler's resource management. See the helper queue guide.

Enable Local API... under Extensions → Job Manager and keep Job Manager running. The API is off by default, requires a token and binds only to 127.0.0.1. The bundled client discovers the active endpoint and token.

Input generators and scripts

ORCA and Gaussian Input Generator Pro offer Submit to Cluster... beside Save when Job Manager is installed. Other plugins can call job_manager.submit_file(paths, name="") through the host's plugin list.

External programs and AI agents can use the bundled command-line or Python client:

python -m job_manager.api_client submit h2o.inp --host mycluster --command "/opt/orca/orca {input} > {stem}.out" --cpus 8 --memory 16GB
python -m job_manager.api_client wait JOB_ID --download

from job_manager.api_client import JobManagerClient

client = JobManagerClient()
job = client.submit(
    host="mycluster",
    files=["h2o.inp"],
    command="/opt/orca/orca {input} > {stem}.out",
    cpus_per_task=4,
    memory="16GB",
    fetch_globs=["*.out", "*.xyz"],
)
final = client.wait(job["id"])
if final["state"] != "DONE":
    raise RuntimeError(f"Job {final['id']}: {final['state']}")
paths = client.download(final["id"], wait=True)["files"]

The client discovers the running Job Manager and uses the same submission service as the GUI. See API documentation for authentication, endpoints and client examples.

Web monitor and notifications

Enable Serve the Host Monitor to a browser in Settings... → Web Monitor for a read-only browser dashboard. It binds to 127.0.0.1; Links and Tailscale... can publish it to your tailnet after HTTPS certificates are enabled. The monitor works from the tray process too, and samples hosts only while someone is viewing the page.

See the web monitor guide for setup, access controls, refresh intervals and data usage. For desktop alerts and Slack/Discord/Teams webhooks, see notifications.

Where your data lives

Persistent state is stored in ~/.moleditpy/job_manager/, outside the plugin folder so updates preserve it.

New lists use SQLite for both internal storage and portable saved files. File → Save As... offers .pmejdb and legacy JSON .pmejbs. Both extensions are registered for File Import and drag-and-drop. Existing jobs.pmejbs or jobs.json stays in its original format until File → Convert to Job Database... is chosen. Conversion verifies the DB, archives the original bytes, then switches tracking to the DB. Close other windows using the legacy list before converting. Reading a file never converts or saves it.

Use Save As to share a consistent snapshot; no result files or host credentials are embedded. SQLite uses rollback journaling, so a completed export is one file. Concurrent editing is intended for local processes with working filesystem locks, not a database opened from several computers through a sync folder.

Job lists can also be exported to CSV. Archived lists open read-only. Result-folder imports also produce read-only lists. See job-list management and the security model for storage and trust boundaries.

Documentation

Development

python -m pytest tests/ -q -n auto
ruff format job_manager tests
ruff check job_manager tests

Most protocol and transport tests use fakes or mocks. Execution tests run actual bash or PowerShell scripts; WSL tests use a stand-in or a real installed distribution. GUI tests require PySide6. Tests with the real MoleditPy PluginContext also require a main-app checkout and rdkit.

CI covers Python 3.11–3.13, real Qt, native Windows and integration with MoleditPy. Optional test tiers skip when their prerequisites are absent; CI enforces the tiers assigned to each job. See CLAUDE.md for the development and release workflow.

License

GNU Lesser General Public License v3.0. See LICENSE.

Licensing scope

First-party code in this migration is LGPL-3.0-or-later. See LICENSE, COPYING and dependency/provenance terms. Third-party components and historical releases retain their own terms.