Menú
Use when monitoring a running Eve experiment and recovering from problems while it runs
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Basado en la clasificación ocupacional SOC
Use when creating new skills, editing existing skills, or verifying skills work before deployment
Use when creating new subagents, editing existing subagents, or verifying subagents work before deployment
Use when choosing, editing, or overriding EvE driver settings for codex_smoke, codex_max, or interactive/debug runs.
Use when an Eve agent is behaving wrong (stuck, gaming the score, bad reasoning) and you need to read its own rollout to find out why
Use when implementing a sanity check subagent for the phase 2 optimization in Eve loop.
Use when implementing a subagent.
| name | supervise-run |
| description | Use when monitoring a running Eve experiment and recovering from problems while it runs |
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.
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):
grep -c "Phase 3: updated Elo" <run_root>/runner.log; confirm the lineage DBs were written recently.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.
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.tmux session). A bare nohup ... & can leave half-initialized workspaces if the launching shell exits and the runner dies mid-iteration.