[wingleeio]

This is a beta update that ships together with the stable engine release. It adds a sidebar view called Activity, plus three engine features it depends on: Waiting, synced Sleep, and Woke.

Activity (zgx)

• Turning it on: the sidebar's … menu, or Settings → Activity.
• The list: one list of every chat, newest return at the top, with Pinned above it.
• Shelves at the bottom of the sidebar, folded until opened:
  • Working: running, or waiting on background work that will wake the agent.
  • Asleep: the chat's corner counts down until it wakes.
  • Archived.
• Options set aside while it's on: Organize, Sort, Active when collapsed, Empty projects, Show archived (always on here), and Create section are disabled. "Move to" leaves the chat menu.
• Sleep (right-click a chat):
  • Preset times: in 1 hour, in 3 hours, this evening, tomorrow, next week. Each shows when it wakes, in local wall-clock time.
  • Custom… opens a new ds.DatePicker in a dialog.
  • Wake wakes it.
• Woke: a woken chat shows a new alarm-clock glyph (status-woke, yellow) until it's opened.
• Stop waiting appears on a chat that is waiting on work its agent marked. Use it when a dev server was marked by mistake.
• Settings → Activity: the view itself, the Working shelf, how Sleep works, and Archive idle chats (Never by default, or 1/3/7/14/30 days).
  • Auto-archive skips pinned, open, asleep, woken, under-way and waiting-on-you chats.
  • A chat you unarchive isn't archived again for being old.
• New ds.DatePicker: a month grid plus a 15-minute time select, keyboard support, days before min disabled, cells changed in place. It has a gallery page and 10 tests.

Waiting (engine)

A turn that ends while its agent waits on background work that will wake it now reads Waiting, not finished. "Finished" is announced once, when the agent is actually done.

• What counts as waiting:

  • A runtime that brings its agent back on its own declares it.
    • Claude Code: verified live on 2.1.319. The CLI started a new turn 0.1 s after a background shell ended.
    • Muse: starts its own wake turns.
    • Codex: verified live on 0.167. It doesn't wake its agent for a finished background terminal, so it doesn't declare it.
  • On such a runtime, background agents always count.
  • A background process counts once the agent says so with the Zeron MCP tool wait_on_background. That call is pinned to the agent's own chat by the agent gate.
  • The tool is offered only on runtimes that wake their agent (Harness::wakes_on_background: Claude Code and Muse). On Codex, Pi or a registry agent, a turn that ends means the agent is done, so the tool isn't listed and can't be called.
  • A dev server left running never counts.

• Holding the completion back: the engine withholds the completion marker while the chat waits. That one change fixes the desktop banner, the chimes, the phone push (the edge already keys off the marker) and MCP wait_for_turn (which also now waits through a wait).

• Every way a wait ends:

  Event	Result
  The wake turn starts	Working from that moment. Claude's start message is now passed on to the engine instead of being dropped.
  The work ends and no wake turn starts within 15 s	The held completion is recorded.
  The person writes	The new turn takes over.
  The person presses Stop waiting (StopWaiting, forwarded to the chat's host)	The marks are released.
  The run ends: stopped, crashed, reaped or replaced	The held completion is still recorded, unless the person stopped the run.
  The engine dies	Its row goes stale, and every client reads the turn as finished and unseen.

• Instructions: wherever the tool is offered, the rule for when to use it is in the tool's own description and the Zeron MCP server's initialize instructions. A Claude Code chat run also gets it in its system prompt, as a short --append-system-prompt note. Live probes on 2.1.319 showed that matters: with the MCP text alone, Sonnet never called the tool, and with the note, Sonnet and Opus both did. Muse tracks its background agents itself, so it needs no note.

• Look: in zgx, Waiting is drawn exactly as Working: the same ring, in the same hue. The row title is strong, the chat counts toward a tab's corner, and a folded place keeps it showing. Its tooltip still reads "Waiting".

Sleep (engine, synced)

• Chat.sleep {until, at, wokeAt} is stored field by field on the chat row, so it syncs to every device. It's set by Mutate setChatSleep, which agents can't call.
• The chat's host wakes it early the moment it starts asking for the person or fails. A question that was already open when it was put to sleep doesn't wake it.
• Woke shows until the chat is seen after it woke.

One status rule

• The zgx host's own copy of the status rule is replaced by proto's (display_status_from, which takes a chat's seen marks so raw rows aren't cloned).
• The old desktop app has no Waiting of its own: it reads Waiting as Working, so it draws nothing new.
• Mobile shows Waiting as Working until the apps draw it themselves.

What behaves differently

• With Activity off, in the zgx app:
  • A run that went quiet while working reads idle rather than done, as on every other client.
  • A terminal's own activity now counts toward status.
  • Archive and the updater's "running chats" count also cover Waiting.
• Everywhere:
  • A turn that ends into a wait no longer pings "finished" until the agent is done.
  • wait_for_turn returns at that point too.

Tests

• proto 117, including stale waits reading finished and sleep states.
• doc 169, including sleep set, wake and clear, cross-device last-write-wins, and a lenient waiting.
• harness: background 14, a Claude fake-CLI scenario replaying the live 2.1.319 frame order, and a test that only a chat's run gets the note.
• engine: 7 wait-lifecycle tests on a scripted runtime, plus an agent-gate test.
• mcp: includes tests that the tool and its instructions appear only on a runtime that wakes its agent.
• zgx host: session tests.
• zgx UI: Activity (11) and DatePicker (10). The sidebar set has the same 5 failures as the untouched base.
• lint: zgx lint ui reports 0 errors.
• Unrelated failures: 2 Claude skills harness tests fail identically on the base.

🤖 Generated with Claude Code