Create abstract Monitor class and new PrimeMonitor for Prime Intellect API logging

[manveerxyz]

Note

Introduce a pluggable monitoring system with Prime Intellect API support, refactor W&B monitor to a shared base, and wire both into orchestrator and config.

• Monitoring framework:
  • Abstract base: Add Monitor with shared helpers (select_sample_problems, should_log_*) and finalization APIs.
  • Prime monitor: Implement PrimeMonitor to POST metrics/samples/distributions/finalization to Prime Intellect API (env: PRIME_INTELLECT_API_KEY, RUN_ID).
  • Composition: Add MultiMonitor and NoOpMonitor; new setup_monitor/get_monitor in utils/monitor/__init__.py to assemble monitors.
  • W&B refactor: WandbMonitor now extends Monitor, shares selection/interval logic; removes old module-level singleton setup.
• Orchestrator:
  • Initialize unified monitor with wandb and prime configs; use for all log, sample/distribution logging, and finalization.
• Config:
  • Add PrimeMonitorConfig and OrchestratorConfig.prime; generalize LogExtrasConfig descriptions; disable prime.log_extras in bench mode.

^{Written by Cursor Bugbot for commit a80ba56. This will update automatically on new commits. Configure here.}

[cursor]

Bug: Uninitialized attributes accessed in PrimeMonitor methods

The PrimeMonitor.__init__ method has early returns at lines 37, 46, and 57, but attributes like last_log_samples_step, samples, last_log_distributions_step, and distributions are only initialized at lines 60-69 after these early returns. When the monitor is disabled (e.g., missing API key or RUN_ID), these attributes remain uninitialized. Later, methods like log_samples() and log_distributions() attempt to access these attributes, causing AttributeError crashes even when is_master is True.

 

[cursor]

Bug: WandbMonitor accesses uninitialized attributes after refactoring

The refactoring changes log_samples() and log_distributions() to call Monitor.should_log_samples() and Monitor.should_log_distributions(), which access self.last_log_samples_step and self.last_log_distributions_step as parameters before checking the config. The old code checked the config first, then accessed these attributes. If the monitor is disabled during initialization (early return at line 36), these attributes are never initialized, causing AttributeError when the methods are called, even if the config would have prevented logging.

Additional Locations (1)

• src/prime_rl/utils/monitor/wandb.py#L165-L166