protocol-info-batch-operator - SKILL.md Agent Skill

name: protocol-info-batch-operator description: Focused sub-skill selected by protocol-info-router for long protocol-info batches, manual queues, background crawls, ScheduleWakeup/task-notification issues, stuck crawls, killed batches, and throughput diagnosis. Use after the router chooses the batch-operations path. Do not use for one-off crawls or record-field edits. user-invocable: false

Protocol-info batch operator

This is a focused sub-skill. If you arrived here directly and the request is a normal one-off crawl or existing-record edit, switch to the protocol-info-router router skill first.

Use this skill when the user is coordinating many protocol-info crawls and the problem is orchestration, queueing, background tasks, or throughput. The normal tool surface is still /protocol-info:protocol-info; this skill decides how to pace and observe the work.

Preferred execution

Prefer one slash-command invocation with built-in batching when the user gives a known protocol list:

/protocol-info:protocol-info --parallel 4 --i18n none \
  --batch --display-name "Pendle" \
  --batch --display-name "Morpho"

Use --parallel min(N_providers, 4) unless the user gives a different parallelism. Avoid hand-rolling per-protocol shell loops when the built-in batch mode is enough.

Manual background queue

If you are manually feeding one protocol at a time in background commands, rely on the background task completion notification to start the next protocol immediately. Do not pace the queue with fixed ScheduleWakeup delays such as 40 minutes; that can add large idle gaps after runs that already finished.

Scheduled wakeups are only a coarse fallback watchdog for missing task notifications. They should not be the normal queue driver.

Diagnosing slow throughput

Before changing timeouts or blaming crawl quality, separate actual crawl runtime from orchestration idle time:

compare adjacent out/.runs/<run-id>/ directory timestamps to estimate actual per-protocol runtime;
check whether a protocol-info process is still alive before waiting longer;
inspect out/<slug>/_debug/r1/r1-status.json only when a current process is still running or the run failed;
remember that stale *.envelope.json files are not a reliable "current run completed" signal unless their mtime belongs to the active run.

Stuck runs

R1 writes live subtask telemetry to out/<slug>/_debug/r1/r1-status.json: state, pid, elapsed time, timeout, and error kind. Claude calls have a default wall-clock watchdog (CLAUDE_TIMEOUT_MS, 15 minutes; R1_CLAUDE_TIMEOUT_MS overrides R1), so a stalled subtask should end as error_kind=timeout and allow partial R1 continuation.

If the user asks to kill a current run, first verify the active protocol-info processes. Kill only the relevant process group and report the exit status.