Watch

Sokosumi Watch

Use this skill after creating a READY coworker task or a direct agent job that is still running. The goal is that the human does not need to manually ask whether the work is done.

Authentication

Use SOKOSUMI_API_KEY if set. Otherwise ask the user to create one at https://app.sokosumi.com/connections.

Workflow

1. Resolve the target id from the user's request or from the most recent Sokosumi task/job id in the conversation.
2. Identify the kind:

sokosumi tasks get task_or_job_id --api-key "$SOKOSUMI_API_KEY" --json
sokosumi jobs get task_or_job_id --details --api-key "$SOKOSUMI_API_KEY" --json

3. If the task/job is already terminal or waiting for user input, report the result now and stop.
4. If your agent runner supports background shell commands, start a background poller. The command must reference $SOKOSUMI_API_KEY; never paste the key value into the command or a file.
5. If background commands are unavailable, set a timer or re-check every few minutes in the main agent loop.

Background Poller

For a task, run a background shell command equivalent to:

for i in $(seq 1 60); do
  status=$(sokosumi tasks get <ID> --api-key "$SOKOSUMI_API_KEY" --json | node -e 'let s="";process.stdin.on("data",d=>s+=d);process.stdin.on("end",()=>{let j=JSON.parse(s||"{}");console.log(j.task?.status||"")})')
  case "$status" in
    COMPLETED|FAILED|CANCELED|CANCELLED|AUTHENTICATION_REQUIRED) echo "SOKOSUMI-WATCH-DONE task <ID> $status"; exit 0;;
  esac
  sleep 60
done
echo "SOKOSUMI-WATCH-TIMEOUT task <ID>"

For a job, use sokosumi jobs get <ID> --details --json and treat completed or failed as terminal.

On Wake

• SOKOSUMI-WATCH-DONE: fetch final details once with sokosumi tasks get <ID> --api-key "$SOKOSUMI_API_KEY" --json or sokosumi jobs get <ID> --details --api-key "$SOKOSUMI_API_KEY" --json, then summarize status, output, files, links, and any requested user action.
• SOKOSUMI-WATCH-TIMEOUT: tell the user the work is still running after about 60 minutes and offer to keep watching.
• Still running on a manual loop: re-arm the next check. Keep each update short.

Do not run two monitors for the same id.