[Feature]: Show the current session's subagents and background tasks, with stop and transcript view #439

Activity

1. ayush5harma

   changed the title [-][Feature]: Open Claude Code's agent view with Left on an empty prompt, and show background agent status in the footer[/-] [+][Feature]: Show the current session's subagents and background tasks, with stop and transcript view[/+] on Oct 10, 2026
2. srothgan commented on Oct 13, 2026

   

   srothgan

   on Oct 13, 2026

   Owner

   More actions

   Following up on your replacement design in #443: let's use this issue for the design discussion and keep a future PR focused on reviewing the agreed implementation.

   The active-session scope and SDK-based direction make sense: include both subagents and background tasks, including background shell jobs. Use the task state we already have, with no separate CLI invocation, private job-file access, global roster, or second task store.

   On your UI questions:

   • Give the active session's subagents and background tasks a separate tab/view, grouped by task type, while reusing the existing configuration display and navigation logic.
   • Use /task to open it, and support access through Left Arrow navigation. Please describe the focus and key behavior, including what happens while editing a draft, so we can settle the interaction here.
   • A footer counter for running subagents and background tasks, including background shell jobs, is acceptable. Derive it from the existing normalized lifecycle state and count each execution once, even when it also appears as a detached tool call. A saved transcript or completed task must not count as running.

   Your proposed SDK operations are the right place to start: the task lifecycle events for live state, Query.stopTask() for requesting cancellation, and listSubagents() / getSubagentMessages() for transcript discovery and inspection. Keep cancellation acknowledgement separate from the eventual stopped state, and keep transcript reads separate from live task status.

   Before implementation, please resolve the task-to-transcript association you raised: do listSubagents() agent IDs match local_agent task IDs in the pinned runtime, and what does the view show when a task has no verified transcript association? Also check which metadata is actually missing: lifecycleMetadata() in tasks.ts already retains progress summary.

   Have your agent read AGENTS.md, docs/src/architecture.md, and the relevant declarations in agent-sdk/node_modules/@anthropic-ai/claude-agent-sdk/sdk.d.ts: SDKTaskStartedMessage, SDKTaskProgressMessage, SDKTaskUpdatedMessage, SDKTaskNotificationMessage, listSubagents(), getSubagentMessages(), and Query.stopTask(). Build on tasks.ts and task_links.ts, and update the Rust and bridge contracts together where needed.

   Please continue the discussion here with the proposed navigation, counter semantics, and runtime ID findings before opening the replacement PR.
3. srothgan

   added

   type: featureNew feature or capabilityNew feature or capability

   status: discussionNeeds community input or design discussionNeeds community input or design discussion

   area: agent-sdkBun bridge and Agent SDK integrationBun bridge and Agent SDK integration

   on Oct 13, 2026
4. ayush5harma commented on Oct 13, 2026

   

   ayush5harma

   on Oct 13, 2026

   Last edited by ayush5harma

   ContributorAuthor

   More actions

   #456 did not follow this issue. I'll close it, and open nothing for this until the points below are agreed. @srothgan please review these

   Runtime ID findings

   Measured on SDK 0.3.288 / runtime 2.1.288 with one subagent and one run_in_background Bash command, in two runs that agreed.

   • A listSubagents() id equals the local_agent task_id from task_started, and it differs from the Agent call's tool_use_id.
   • A running subagent was listed before it completed, but not at the first poll about 2 s after task_started. So a running task can briefly have no transcript.
   • The background shell task (task_type: local_bash) never appears in listSubagents(). background_tasks_changed listed only the shell task, never the foreground subagent.
   • stopTask(task_id) on the shell task produced task_updated (killed), then task_notification (stopped) about 3.8 s later, both with the same task_id.
   • The stopTask promise itself rejected with "Query closed before response received", because the single-prompt query ended first. I'd confirm the acknowledgement on a long-lived query before relying on it.
   • Not covered: nested or backgrounded subagents, resume.

   Missing metadata

   Nothing the view needs is missing. lifecycleMetadata() keeps every field these runs sent: task_type, subagent_type, is_backgrounded, output_file, end_time, the terminal status and summary. Only task_progress's usage (tokens, tool uses, duration) and last_tool_name are dropped; last_tool_name is only logged. I'd add them only if you want them on a row.

   Navigation

   • Opening it: a Tasks tab in the existing fullscreen view, grouped by task type, reusing its tab and list navigation. /task opens it on that tab.
   • Left Arrow: opens it only when the composer is empty and no menu is open. With a draft, Left moves the cursor as today.
   • Keys in the view: Up/Down select, Enter opens a verified transcript, s stops a running task, r re-checks transcripts. Esc returns to the composer with any draft intact.

   Counter

   The footer counts running local_agent and local_bash lifecycle tasks, once per task_id, including a task that is also a detached tool call. Completed, failed and stopped tasks and saved transcripts never count. The counter is hidden at zero.

   Stop and transcripts

   Stop is offered only for a running lifecycle task_id. A row shows Stopped only when the lifecycle messages say so, never on the request. A row gets a transcript only when its local_agent task_id matches a listed id. Until then it shows "no transcript", and nothing is inferred.

   Open questions

   1. Should other task types (local_workflow, MCP tasks) appear and count?
   2. Should finished tasks stay listed for the whole session?
   3. Should the id equality get a contract test against the pinned SDK?
5. srothgan commented on Oct 13, 2026

   

   srothgan

   on Oct 13, 2026

   Last edited by srothgan

   Owner

   More actions

   1. Do we want a subagent/agent/task list?
   2. I want a new full screen view. Not part of the current config/mcp/usage view but a new dedicated full screen, just matching copying some logic. I dont want to see Agents next to MCP in the tabs list

   Maybe only subagents list? Since we already have a dedicated function we can call and use
6. ayush5harma commented on Oct 13, 2026

   

   ayush5harma

   on Oct 13, 2026

   ContributorAuthor

   More actions

   1. Yes, one list for the current session, grouped by task type: subagents (local_agent) and background shells (local_bash). "Agent" and "subagent" are the same rows here, so that is two groups today, and more only if you want other task types. Selecting a subagent shows its transcript in the same screen.
   2. Agreed: a dedicated full-screen Tasks view with its own title, not a tab next to MCP. /task, or Left Arrow from an empty composer, opens it, and Esc returns to the composer with the draft intact. It would reuse the list, selection and scroll helpers the config view uses, without being part of that view.

   The rest of the proposal stays as written. Does this match what you have in mind? I'll wait for your go before opening a PR.
7. srothgan commented on Oct 13, 2026

   

   srothgan

   on Oct 13, 2026

   Owner

   More actions

   Go ahead, but use /agents since we only do subagents now
8. srothgan

   assigned

   ayush5harma

   on Oct 14, 2026