supervise-run

name	supervise-run
description	Use when monitoring a running Eve experiment and recovering from problems while it runs

Supervise a run

Watch a live run, detect problems, and either recover (mechanical) or escalate (judgment). This skill is the monitoring loop; it delegates the actual fixes to the tool skills it references.

run_root layout, the completion marker, and data sources are in docs/using-eve.md.

Monitoring loop

Run detached so you supervise from logs, not the foreground. (codex_max and codex_smoke run headless by default; the open_iterm2: false knob applies only to interactive/tmux-pane drivers and will error if passed to a subprocess driver.) On a sparse cadence (e.g. ~15 min steady-state, tighter right after launch or after a problem):

1. Read progress: completed iterations = grep -c "Phase 3: updated Elo" <run_root>/runner.log; confirm the lineage DBs were written recently.
2. Classify the run: running / stalled / halted / completed.
3. Act per the table.

When to act vs escalate

Principle: handle mechanical problems yourself; escalate scientific / taste judgments to the human.

Signal	Meaning	Action
runner.log tail shows a crash/halt, process gone	mechanical halt	recover: follow ../resume-run/SKILL.md
no new Phase 3 marker for a long time, process alive	stalled	inspect the agent's behavior via ../debug-agent-session/SKILL.md; if truly stuck, pause + resume
persistent synced 0 Phase 2 optimizers without a benign "guidance tree was not modified" reason (see Notes)	silent no-op: optimizer guidance not landing	stop and flag to the human
score plateau or regression over many iterations	possibly a bad direction	flag to the human (do not silently change the experiment)
reached loop.max_iterations / clean exit	completed	report results (../inspect-population/SKILL.md) and stop

Scientific calls — "is the optimizer bad?", "should the prompt be bolder?", "should we restart?" — are the human's. Surface them; act only when the human decides. They may then ask for an adjust (../resume-run/SKILL.md) or a fresh start keeping good solvers (../import-run/SKILL.md). The human can explicitly delegate authority (e.g. "you tune the prompt") — only then adjust autonomously.

Notes

• Count completed iterations only from the Phase 3 marker, never from Iteration X / Y lines (printed before completion).
• synced 0 Phase 2 optimizers just means no optimizer candidate was added that iteration. It is benign when paired with a produce_optimizer_in_phase2 is enabled but the guidance tree was not modified warning (the optimizer agent simply did not edit its guidance, normal in quick smokes). Only treat persistent synced 0 as a real silent-no-op to flag in a longer run where you actually expect the optimizer to be evolving.
• Keep the runner process itself alive for the whole run (e.g. a persistent tmux session). A bare nohup ... & can leave half-initialized workspaces if the launching shell exits and the runner dies mid-iteration.
• Cluster / HPC submission and monitoring is environment-specific; if you run on a cluster, drive submission through your own cluster setup, not this skill.