@skydiveai/pi-extensions 0.1.0-beta.3067 → 0.1.0-beta.3069
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/index.mjs +1 -1
- package/package.json +1 -1
package/dist/index.mjs
CHANGED
|
@@ -4293,7 +4293,7 @@ function buildTool(messageId) {
|
|
|
4293
4293
|
"Use it to parallelize independent work, to keep a large or noisy subtask out of your own context, or to run a task under a specialized persona.",
|
|
4294
4294
|
"Fire-and-forget: this returns immediately after queueing. It does NOT wait for results. Each subagent runs on its own and, when it finishes, sends you its result on this thread — so queue the work, then keep going or end your turn. To chain, re-delegate after a result lands.",
|
|
4295
4295
|
"To spawn, pass { task, title, persona?, model?, timeoutMinutes? }. To keep waiting after a timeout, pass { taskId, timeoutMinutes }; this schedules another wake and leaves the same child running. title is the short human-visible name for new work. persona and model apply only when spawning.",
|
|
4296
|
-
"Peering: each queued task comes back with its own conversation id. A subagent is a real linked conversation, so to see what one is doing RIGHT NOW while it runs — its reasoning, the tools it has called and their results, its progress — read that conversation with `platform conversations show <conversationId>` (you are already authorized; it is your own delegated run). Check in that way instead of waiting blind for the final result. The read reflects the child's persisted state, which lags a few seconds behind live (tool results land as they complete; in-progress reasoning can be up to ~5s stale), so
|
|
4296
|
+
"Peering: each queued task comes back with its own conversation id. A subagent is a real linked conversation, so to see what one is doing RIGHT NOW while it runs — its reasoning, the tools it has called and their results, its progress — read that conversation with `platform conversations show <conversationId>` (you are already authorized; it is your own delegated run). Check in that way instead of waiting blind for the final result; never sleep before a check-in, that blocks your turn pointlessly. If the child needs more time, re-arm it with tasks: [{ taskId, timeoutMinutes }] and end your turn; the timeout wake brings you back to react. The read reflects the child's persisted state, which lags a few seconds behind live (tool results land as they complete; in-progress reasoning can be up to ~5s stale), so it is a progress view for checkpoints, not something to poll in a tight loop.",
|
|
4297
4297
|
"Steering: to add context, correct course, or answer a question a subagent needs mid-run, post to its conversation with `platform conversations post <conversationId> --message \"...\"`. If the subagent is still running, your message lands as a live steer picked up in that same turn; if it has gone idle, it queues as its next turn. This is the same primitive as any conversation message — there is no separate steer channel."
|
|
4298
4298
|
].join(" "),
|
|
4299
4299
|
promptSnippet: "subagent — delegate tasks to isolated subagent runs; each rewakes you with its result when done",
|