feat(voice,host,panel): calls are on agents

Merged
DA
5d ago
+2924-6146

PR 5b of the Unified agents rollout (STAG-322). It builds on 5a (#4246) and 2c (#4260), both merged.

This is the PR that changes what the developer sees: a call is about an agent, not a task. Any agent's tab can take a call, and "+ → New voice chat" starts a blank agent and calls it. The fixed Call tab and the client's task-keyed call path are gone. There is no backwards compatibility: kept tab ids, Call pins, and /call links from before are dropped.

What changed

Voice

  • LiveVoiceOrchestrator keys its one call by { taskId, agentId } (VoiceCallAbout).
  • talkAboutTask becomes talkTo(about). Calling another agent hangs up the call that stands first.
  • VoiceView.callTaskId becomes VoiceView.call, and isVoiceView reads it.
  • The renderer's createSession passes the agent on. The session holder and source put agentId on the service's session.create frame.

Host and main

  • voice.createLiveSession now requires agentId and forwards it.
  • Before creating a session, the host reads the agent once. It refuses an agent on another task, or one that is archived or out of reach. This adds one read to call-open latency, and the service checks again.
  • TASKS_TALK and PUSH.TASKS_TALK carry the agent.
  • New codingAgents.calls act, which reads 5a's calls(agentId).
  • call.showTask becomes call.showAgent.

The panel

  • Call button. An agent's composer has a Call button beside Dictate, on desktop only (HOST_CAPABILITY.VOICE).
  • During a call. The tab's foot draws the avatar video, the captions, and the CallBar, and the tab's glyph wears the live mark.
  • Calls in the transcript. Each call's divider and end line sit around that call's spoken rows (callPlaces, from 5a's calls read). Lines heard on the live call that aren't on record yet grow at the bottom.
  • "+" menu. It offers New voice chat, New agent, then each agent. Start's menu offers New voice chat at its foot.
  • No Call tab. Deleted: MAIN_TAB.CALL, TaskCall, LINK_PLACE.CALL, TAB_WORDS.CALL, Call pins, transcriptStreaming, and the Call revision sources.
  • A task with no agent tab opens on its plan. The side panel opens on Plan, and the column shows an empty pane offering New voice chat and New agent.
  • Header. Start and the task's ⋯ stand in the task's header on every tab.
  • Following an ended agent (the #4246 review point). Main's pushed-transcript follow (codingAgents.transcript) already keeps an ended agent followed while its tab is shown, so a call's spoken rows reach it. A tab shown again reads from the start. A test now covers this.
  • Settings. Settings › Coding agents is now Settings › Agents.

Decisions from Dean, passed on by the orchestrator

  • The global keys are removed. The talk key (⌥Space) and the stop key (⌥S) are deleted, along with:

    • their registrar and the mac-talk-key helper;
    • the Settings › Voice "Keys" section and their stored settings (the retired ids are kept);
    • the hotkey pushes and the app-state slice;
    • setShortcutCapturing, LUKE_STOP, and the orchestrator's beginTalk/endTalk.

    The in-call mute and speaker chords and Escape are unchanged. RunMode.registersGlobalKeys gives way to takesFocus for the voice window. Video calls lose the avatar barge-in flush, which only the held talk key fed. Re-keying it on mic activity is a possible follow-up.

  • Plan mode is removed from New task. Every start sends planning: false, and starting a task never opens a call.

Docs: docs/TASKS.md (calls, agent tabs, transcript, call bar, New task), docs/DESIGN.md (keys section removed), PRIVACY.md, README.md, and CONTRIBUTING.md.

Not in this PR

  • Off-call show_code in the Code tab. An agent's show_code reaches the Code tab only during a call, through the call's task.code frames. Drawing it from the agent's record, coloured by the host's highlighter, will be its own PR. The Code feed now heads each code by the turn that showed it, rather than grouping code under calls.
  • Typing into a live call is PR 6.
  • Deleting the per-task Luke and the service's task call path is PR 7.

Tests

  • Orchestrator: with two agents, calling the second hangs up the first, and the view names the agent.
  • Host: an offer into another task's agent, or into an archived agent, is refused and opens no session. The frame names the agent.
  • Tab model: there is no Call tab, and a task opens on its plan. Closing the last tab goes back to the plan, and a kept "call" tab is dropped.
  • New voice chat: starts a blank agent, opens its tab, and calls it. The web offers none.
  • Second call: pressing Call while another agent's call stands asks first.
  • Call button: it stands beside Dictate on desktop and is absent on the web. Mid-call, the call bar and live mark show.
  • Transcript follow: an ended agent stays followed while its tab is shown, and spoken rows pushed after it ended join the transcript.
  • callPlaces: places the dividers and the heard tail.

Locally:

  • tsc is clean for voice, host, ui, panel, settings, features, gateway, service, desktop, and web.
  • biome is clean at error level, and the discarded-effect check passes.
  • vitest passes for voice, host, ui, settings, features, gateway, desktop, and panel (543 files, 4786 tests).
  • knip reports only ChatGptSave and ChatGptSignInRequest in packages/credentials, which come from main's #4243 and aren't touched here.

Evidence

Screenshots follow in a comment. They come from a throwaway harness that draws DesktopTasks over the panel's test fixtures in Chromium. The red "No act port" line in the side panel is the harness's own.

Left as is: the changelog-screenshots skill's staging shot for the old Call tab release (.agents/skills/changelog-screenshots/staging/shots/call.ts), which is applied only to that release's own build.

Made with Luke

Devin Review