Resource Management

1. Responsibility boundary

The resource control plane combines reusable safety primitives in scienceflow/safety/resource/ with the LNR task-local runtime in solver/lnr/resource_runtime/. It observes Bash jobs, enforces resource boundaries, manages GPU ownership and waiters, reviews execution progress, and returns compact operational feedback.

2. Control loop

The default resource_smart_llm profile uses a monitor-first loop. Deterministic code identifies hard boundaries and decides when a review is warranted. The owning-agent advisory and resource arbiter reason over a compact fact packet, and a strict execution gate validates any destructive intervention.

3. User-facing modes

config/resource_modes.py expands the selected mode into low-level switches. Explicit keys in the same LNR mapping take precedence, allowing narrow expert overrides without duplicating a complete profile.

4. Task-local multi-worker control

All workers in one LNR task share a task-local resource runtime, unified event store, GPU lease store, pressure state, history, and waiter queue. No worker is permanently the “main” resource holder. The current lease owner is determined by actual admission and can change as workers explore stronger routes.

5. Startup admission and monitored trials

The default startup policy is trial_first. Admission first checks deterministic hard boundaries and atomic lease availability. Uncertain or historical soft blocks can become OBSERVE_THEN_RUN monitored trials, allowing the runtime to gather current evidence before making a value judgment.

6. Observed execution facts

BashTool and the observer build a time-consistent snapshot rather than treating a growing log file as sufficient evidence of value.

Artifact, recoverability, and nested update ages share the same elapsed-time basis. This avoids decisions where a top-level artifact appears stale while a nested recoverability record still appears fresh.

7. Review state machine

safety/resource/review_signal.py compresses raw observations into ResourceReviewSignal. review_state.py advances one state per job, and review_boundary.py decides when to request a review.

8. Advisory, arbiter, and strict gate

When the owning main agent is blocked on a Bash result, the control plane can ask a narrow no-tools question through a side channel. The advisory reports a preference such as safe_to_stop, timebox_continue, continue, or unknown, plus confidence and status. The raw exchange is audited but not inserted into the live research transcript.

The resource arbiter then evaluates progress, metric history, remaining budget, ETA, recoverability, waiter pressure, and prior review commitments. review_outcome.py normalizes the value-review execution surface to:

• KILL — request termination and replan feedback.
• TIMEBOX — continue for a bounded interval with a concrete clear_on condition.
• NO_ACTION — continue observation, normally without main-agent feedback.

resource_runtime/review/arbiter_gate.py is the execution fuse. It validates action class, target job, lease ownership, advisory/evidence support, confidence, deliverable state, and secondary-only targeting before an actual kill is allowed.

9. Timeboxes and proof windows

A timebox names a measurable clear condition such as metric update, artifact growth, structured progress advance, active-work recovery, or cleared opportunity cost. compute_timebox_sec combines observed cadence, heartbeat, configured minimum/maximum, remaining-budget fraction, and waiter pressure.

During TIMEBOX_ACTIVE, ordinary route-value retriggers are frozen. If the clear condition is met, intervention pressure falls. If a proof window expires without useful progress, its failure is recorded. After resource_review_max_proof_windows failures, the proposal carries resource_budget_escalation.kind=final_resource_review; the resource arbiter makes one final budget decision from structured facts without repeating owning-agent advisory.

10. GPU queue, sharing, and release

11. Feedback and audit hygiene

Complete control-plane events are appended to task_logs/resource/resource_events.jsonl. The main agent receives only actionable wait, share, replan, kill, or compact review facts through the current tool result. ResourceFeedbackMemoryDeduper folds repeated capability states and does not create a continuously changing live-summary slot.

12. Current file ownership

scienceflow/
├── safety/resource/
│   ├── review_signal.py      # Stable observed-fact summary
│   ├── review_state.py       # Per-job state machine and timebox sizing
│   ├── review_boundary.py    # Review trigger types
│   ├── review_outcome.py     # KILL / TIMEBOX / NO_ACTION normalization
│   └── ...                   # Process, completion, artifact, lifecycle helpers
└── solver/lnr/resource_runtime/
    ├── runtime.py            # Task-local coordination
    ├── observer/controller.py# Bash-job observation and event production
    ├── admission.py          # Startup admission arbiter
    ├── gpu_lease_store.py    # Lease persistence
    ├── gpu_sharing.py        # Sharing policy and configuration
    ├── unified_store.py      # Canonical resource event store
    └── review/
        ├── arbiter.py        # Resource-value decision
        ├── arbiter_gate.py   # Pre-execution strict gate
        ├── execution_facts.py
        ├── progress.py
        ├── research_cadence.py
        └── state_generation.py

13. Important defaults