get-playground-task - SKILL.md Agent Skill

name: get-playground-task description: Gets detailed status and output of a specific examiner task on a playground. Use to debug why a task is failing. argument-hint:

Get detailed information about examiner task $1 on playground $0.

Inspecting a task's status and last-run output (preferred: tunnel-free)

Use the API path — it needs no SSH tunnel, covers all machines at once (including noSSH ones), and includes each task's status plus the stdout/stderr (and hintcheck/failcheck output) of its last execution:

labctl playground tasks $0 -o yaml

Grep for $1 in that output to find the specific task. This reports the last run only; it does not trigger a fresh run. Prefer this over labctl ssh $0 -- examinerctl task get $1 — the SSH form shows a single machine and burns a tunnel slot, whereas the YAML API form gives the same per-task output for every machine without SSH.

Waiting (blocking) until a task reaches a terminal state (needs SSH)

examinerctl task wait does not trigger, start, or re-run anything — there is no task start/task run, and you cannot manually fire a task. The examiner evaluates the task on its own; task wait just blocks until it reaches a terminal state (pass/fail) or the timeout expires. This is the one task operation the API can't do:

labctl ssh $0 -- examinerctl task wait $1 --timeout 120s

Use it to block after setting up a task's preconditions — e.g. a .solution-NN.sh ends with examinerctl task wait <verify-task> so applying the solution blocks until that task has settled.

Notes

Tasks are edge-activated, not level-triggered, and their status never regresses: the examiner keeps evaluating a task until it first observes a success or failure state, then the task settles into its terminal state (completed 40 or failed 35) and never runs again, even if the world state later changes. The status only ever goes up — a passed task cannot later fail — so a single completed observation is authoritative.
Avoid SSH churn. Each examinerctl ... call opens an SSH tunnel. Use the tunnel-free labctl playground tasks $0 -o yaml for inspection and polling, and never put labctl ssh -- examinerctl task wait inside a backgrounded retry loop — orphaned ssh children exhaust the tunnel pool (see the SSH tunnel discipline notes in run-playground-command).