[FEATURE] Cross-session "needs your attention" inbox — questions, decisions, and findings aggregated outside the transcripts

Status Open
Maintainer reply None cached
Activity 0 comments · opened Aug 4, 2026

Preflight Checklist

  • [x] I have searched existing requests and this feature hasn't been requested yet
  • [x] This is a single feature request (not multiple features)

Problem Statement

Running several parallel sessions (multiple clones, background agents, long-running QA), the model regularly produces items that need me: a question it asked mid-reply, a decision blocking it, a bug it discovered in passing, a suggested next task. Today each of those exists only as prose inside one session's transcript.

Humans skim. An item buried in the middle of a long reply — in a session I wasn't watching — is an item I never saw. There is no place I can look to answer "across all my sessions, what is waiting on me right now?"; I have to open each session and scan.

Proposed Solution

An attention inbox: a cross-session panel listing items that need the user, each linking back to its session/message. Populated either by a dedicated tool the model calls (like spawn_task, but for questions/decisions/findings rather than new-session work) or by the harness detecting blocked-on-user states. Items clear when answered or dismissed.

The session list could also show a per-session "waiting on you" badge derived from the same data — making "which of my 12 sessions needs me" a glance instead of an audit.

Alternative Solutions

Currently emulated by convention in my instructions: every reply that contains such an item must restate it in a "Needs you" section directly after the TLDR, so skimming is safe within one reply. But it's per-reply and per-session — there's still no cross-session view, and the guarantee only holds as well as the model follows the convention.

Priority

Medium - Would be very helpful

Feature Category

Other (Desktop app)

Use Case Example

  1. Morning: I start three sessions — release verification, a bug investigation, a docs sweep — and go into meetings.
  2. The investigation hits a judgment call ("this deviates from the spec — reconcile or rationalize?") and blocks; the verification finds an unrelated bug worth filing; the docs sweep finishes clean.
  3. Back at my desk, the inbox shows exactly two items with deep links. Today, I'd instead open all three sessions and scan the transcripts to discover the same two things.

View original on GitHub ↗