- Fix Process.current() to store psutil cmdline instead of sys.argv for accurate validation
- Fix worker process_type detection: explicitly set to WORKER after registration
- Fix ArchiveResultWorker.start() to use Process.TypeChoices.WORKER consistently
- Fix migration to be explicitly irreversible (SQLite doesn't support DROP COLUMN)
- Fix get_running_workers() to return process_id instead of incorrectly named worker_id
- Fix safe_kill_process() to wait for termination and escalate to SIGKILL if needed
- Fix migration to include all indexes in state_operations (parent_id, process_type)
- Fix documentation to use Machine.current() scoping and StatusChoices constants
Co-authored-by: Nick Sweeting <pirate@users.noreply.github.com>
- Add pwd validation in Process.launch() to prevent crashes
- Fix psutil returncode handling (use wait() return value, not returncode attr)
- Add None check for proc.pid in cleanup_stale_running()
- Add stale process cleanup in Orchestrator.is_running()
- Ensure orchestrator process_type is correctly set to ORCHESTRATOR
- Fix KeyboardInterrupt handling (exit code 0 for graceful shutdown)
- Throttle cleanup_stale_running() to once per 30 seconds for performance
- Fix worker process_type to use TypeChoices.WORKER consistently
- Fix get_running_workers() API to return list of dicts (not Process objects)
- Only delete PID files after successful kill or confirmed stale
- Fix migration index names to match between SQL and Django state
- Remove db_index=True from process_type (index created manually)
- Update documentation to reflect actual implementation
- Add explanatory comments to empty except blocks
- Fix exit codes to use Unix convention (128 + signal number)
Co-authored-by: Nick Sweeting <pirate@users.noreply.github.com>
This consolidates scattered subprocess management logic into the Process model:
- terminate(): Graceful SIGTERM → wait → SIGKILL (replaces stop_worker, etc.)
- kill_tree(): Kill process and all OS children (replaces os.killpg logic)
- kill_children_db(): Kill DB-tracked child processes
- get_running(): Query running processes by type (replaces get_all_worker_pids)
- get_running_count(): Count running processes (replaces get_running_worker_count)
- stop_all(): Stop all processes of a type
- get_next_worker_id(): Get next worker ID for spawning
Added Phase 8 to TODO documenting ~390 lines that can be deleted after
consolidation, including workers/pid_utils.py which becomes obsolete.
Also includes migration 0002 for parent FK and process_type fields.
New section 1.5 adds @property proc that returns psutil.Process ONLY if:
- PID exists in OS
- OS start time matches our started_at (within tolerance)
- We're on the same machine
Safety features:
- Validates start time via psutil.Process.create_time()
- Optional command validation (binary name matches)
- Returns None instead of wrong process on PID reuse
Also adds convenience methods:
- is_running: Check via validated psutil
- get_memory_info(): RSS/VMS if running
- get_cpu_percent(): CPU usage if running
- get_children_pids(): Child PIDs from OS
Updated kill() to use self.proc for safe killing - never kills
a recycled PID since we validate start time first.
Phase 3.3 now includes:
- Module-level _supervisord_db_process variable
- start_new_supervisord_process(): Create Process record after Popen
- stop_existing_supervisord_process(): Update Process status on shutdown
- Process hierarchy diagram showing CLI → supervisord → workers chain
Key insight: PPID-based linking works because workers call Process.current()
in on_startup(), which finds supervisord's Process via PPID lookup.
Phase 2 now includes line-by-line mapping of:
- run_hook(): Create Process record, use Process.launch(), parse
JSONL for child binary Process records
- process_is_alive(): Accept Path or Process, use Process.is_alive()
- kill_process(): Accept Path or Process, use Process.kill()
- ArchiveResult.run(): Pass self.process as parent_process to run_hook()
- ArchiveResult.update_from_output(): Read from Process.stdout/stderr
- Snapshot.cleanup(): Kill via Process model, fallback to PID files
- Snapshot.has_running_background_hooks(): Check via Process model
Hook JSONL contract updated to support {"type": "Process"} records
for tracking binary executions within hooks.
PIDs are recycled by OS, so all Process queries now:
- Filter by machine=Machine.current() (PIDs unique per machine)
- Filter by started_at within PID_REUSE_WINDOW (24h)
- Validate start time matches OS via psutil.Process.create_time()
Added:
- ProcessManager.get_by_pid() for safe PID lookups
- Process.cleanup_stale_running() to mark orphaned RUNNING as EXITED
- START_TIME_TOLERANCE (5s) for start time comparison
- Uses psutil.Process.create_time() for accurate started_at
Key addition: Process.current() class method (like Machine.current())
that auto-creates/retrieves the Process record for the current OS process.
Benefits:
- Uses PPID lookup to find parent Process automatically
- Detects process_type from sys.argv
- Cached with validation (like Machine.current())
- Eliminates need for thread-local context management
Simplified Phase 3 (workers) and Phase 4 (CLI) to just call
Process.current() instead of manual Process creation.
Documents 7-phase refactoring to use machine.Process as the core data
model for all subprocess management:
- Phase 1: Add parent FK and process_type to Process model
- Phase 2: Add lifecycle methods (launch, kill, poll, wait)
- Phase 3: Update hook system to create Process records
- Phase 4-5: Track workers/orchestrator/supervisord as Process
- Phase 6: Create root Process on CLI invocation
- Phase 7: Admin UI with tree visualization
Enables full process hierarchy tracking from CLI → binary execution.
