Workbench loop: connect delegation, integration, and verification

[Hmbown]

Problem

The current Work / Tasks / Agents sidebars show useful pieces, but the loop does not close visually. In the v0.8.43 prep screenshot, the parent had delegated harvest work, background checks were running, agent results were being evaluated, and verification was pending, but the UI made those look like unrelated tool rows.

The user-facing story should be: delegated -> running -> summary ready -> parent reviewed -> integrated -> verified -> closed.

Scope

For v0.8.45, model delegated work as a visible lifecycle across Work, Tasks, Agents, and transcript cards.

• Show whether a child result has been reviewed by the parent.
• Link Work checklist items to child sessions and background jobs where applicable.
• Distinguish durable background jobs from sub-agents. A cargo check should not look like a teammate, and a teammate should not look like a shell process.
• Collapse repeated progress rows into one live card with expandable details.
• Keep final verification as a first-class state, not just another tool row.
• Make timeout/partial-result states obvious and actionable.

Acceptance criteria

• A completed child session can be marked summary ready and then reviewed once the parent consumes it.
• Work checklist items can show linked agent/job status without raw tool plumbing.
• Background jobs and sub-agents have distinct visual language.
• The main transcript no longer needs to show raw delegate done / agent_eval JSON for the common path.
• Tests cover the state transitions and rendering of at least one delegated task through review and verification.

Related tracker: #1879
Related control-plane issue: #1895
Related slash-command issue: #1888
Related child-output issue: #1738

[Hmbown]

Addendum from the Workbench/Agents scout: the Agents panel should expose child-session lifecycle actions directly once #2018 gives rows stable identity. A focused agent row should be able to support actions such as wait/read summary, mark reviewed, open transcript, cancel, copy summary, attach result to a work item, and close when done.

The important design point is that the common path should not require raw agent_eval / tool JSON. The visible lifecycle should be delegated -> running -> summary ready -> reviewed -> integrated -> verified -> closed, with timeout and partial-result states inspectable instead of buried in recent-tool rows.

[Hmbown]

Maintenance follow-up from the redacted session-log analysis: I split three concrete workbench/delegation slices out of this broader loop.

• Agent routing: detect when parent work should delegate to scouts or RLM #2024: detect when parent work should delegate to scouts or RLM.
• GitHub triage scout: delegate issue, PR, and milestone maintenance passes #2025: GitHub triage scout for issue/PR/milestone maintenance.
• RLM/log workbench: route large local logs and structured payloads out of the parent transcript #2023: RLM/log workbench for large local logs and structured payloads.

The main pattern was not "the model forgot how to work" so much as repeated parent-thread reconnaissance that had clear independent scout/log-analysis boundaries. These should feed into the delegation → integration → verification loop here.