4.9 KiB
status, claimed-by, claimed-at, resolved-at, parent, blocked-by, tags
| status | claimed-by | claimed-at | resolved-at | parent | blocked-by | tags | |||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| resolved | 019fbff0-e725-79bd-861c-f1928d963392 | 2026-08-01T21:13:36-04:00 | 2026-08-01T21:51:14-04:00 | 083-pi-subagent-result-delivery-map |
|
|
Pi subagent result delivery product semantics
Question
What product semantics should govern reliable non-blocking subagent result delivery and the improved subagent UI, including inbox versus notification versus event stream, read and unread state, failure visibility, result retention, visible summaries of what each child is working on, and whether any controlled parent-session injection is allowed?
Working decisions
The improvement is not better polling tools for the human. It should make subagent completion part of the parent agent's working obligations. The human should not have to remember to check whether subagents finished. During HITL work, completion should be visible without derailing the active exchange. After HITL work resolves, or when the human explicitly pauses it, the parent agent should deal with completed subagent work. The primary success story is Wayfinder dispatching research workers, continuing a HITL grill, showing completion awareness during the grill, then reconciling the research when the grill pauses or ends. The design must also handle multiple parallel workers completing at different times without losing track of which results still need reconciliation. Completion awareness during HITL should combine brief visible notices with persistent compact state, so completions are announced once and then remain visible as deferred work. Completion notices should identify what finished and whether it succeeded, failed, timed out, or was cancelled, with batching when several workers complete together. Child work labels should use a hybrid with override: accept an explicit caller-provided label when available, infer from known structure such as a Wayfinder ticket title when possible, and otherwise generate or truncate from the prompt. The generic subagent tool should not prescribe a workflow protocol for what handling a result means. It should make completed background work reliably visible and retrievable to the parent, while workflow-specific skills such as Wayfinder decide what to do with the result. The generic layer should provide reliable completion state, reliable terminal result retrieval, completion awareness, and work labels, without workflow-specific reconciliation rules. Terminal result persistence across parent session restart, Pi restart, or extension reload is not required for this effort. Current-session reliability is enough. Within the current session, completed results should remain visible and retrievable until explicitly cleared. Improved UI should have a compact aggregate view and an expanded per-subagent work-item view. The compact view should show counts by agent status rather than one row per child. The expanded view should show one row per subagent with label, state, elapsed time, and last activity or result marker, with truncation when needed. Selecting a subagent for inspection should attach a read-only live view of that child to the parent UI, so the human can watch what is happening without sending input to the child. The attached view should look like a normal Pi session view, but without the text prompt. The attached child view should be an overlay or temporary replacement, not a split pane. The attached child view only needs detach plus normal scroll up and down controls. The attached child view is part of this result-delivery improvement, not a hypothetical first implementation. Feasibility still needs research before implementation slicing because it may touch Pi TUI internals rather than only the existing extension surface.
Result
The product semantics are settled for this phase. The generic subagent layer should provide current-session reliable completion state, terminal result retrieval, completion awareness, and work labels without prescribing workflow-specific reconciliation behavior. Completion should be visible during HITL work but should not derail it. Workflow-specific skills decide what to do with completed results once their own safe boundary is reached. The UI direction is compact aggregate counts by status, expandable per-subagent rows with label, state, elapsed time, and last activity or result marker, plus a read-only Pi-like attached child view without a text prompt. The attached view should use an overlay or temporary replacement rather than split panes and only needs detach plus normal scroll controls.