Surface background work in the app so a user can watch it and take it over

[yh928]

Summary

Background AI runs — trigger triage, escalated sub-agents, flow runs, scheduled jobs — are invisible in the app. Give them a surface where a user can see what is running, open one to read what it is doing, and attach to it: take over, steer it, or stop it.

Problem

The agent already works on its own: a Composio webhook is triaged and can escalate into an orchestrator run without any user turn (agent::triage::escalation::dispatch_target_agent). None of that is visible. A user sees the effects — a draft, a memory, a message — with no way to see the work that produced them, and no way to intervene while it is happening.

What exists today and why it is not enough:

• Progress events — AgentProgress is emitted during runs, but it is scoped to a chat turn the user is watching. A background run has no chat surface subscribed to it.
• The run ledger — threads::todos::runs records autonomous runs, and tinyagents_store/journal holds per-run event JSONL. Both are reachable from code, neither is rendered.
• The task board — todo_* tools and openhuman.todos_* RPC exist, and cards carry status, assigned_agent, blocker, evidence. It is thread-scoped, so background work started from a webhook does not land on any board a user would open.

The gap is not data collection so much as a place to look. And once there is a place to look, "attach" is the part that changes what the agent can be trusted with: a run a user can interrupt is a run that can be allowed to do more.

Solution (optional)

Scope is app + core. Sketch:

• View — a list of active and recent background runs: what triggered it, which agent, current phase, elapsed, and terminal status. Reading from the run ledger and journal rather than a new store.
• Open — one run's timeline: the progress events, tool calls, and their results, in order. The journal already holds this.
• Attach — bind the run's thread into the chat surface so the user can send a turn into it, and stop it. HarnessRunStatus already models Cancelled, and the startup sweep already writes it (agent::tinyagents::reaper), so cancellation has a precedent to follow.
• Approvals — a background run that parks on the approval gate currently has nowhere to prompt (security::approval::gate notes this explicitly for TrustedAutomation: "There is no chat thread to route the prompt to for a background/triggered flow run yet"). This surface is where that prompt belongs.

Depends on the progress record from the trigger-side issue for anything triggered by a webhook; the run ledger covers the rest and could ship first.

Acceptance criteria

• Runs are listable — active and recently-finished background runs, with trigger, agent, phase, and status.
• A run is openable — its ordered event/tool timeline is readable without going to logs.
• A run is stoppable — the user can cancel from the UI and the run reaches a terminal state.
• A run is attachable — the user can send a turn into a live background run.
• Parked approvals are reachable — a background run awaiting approval prompts here instead of TTL-denying unseen.
• Diff coverage ≥ 80% — the implementing PR meets the changed-lines coverage gate.

Related

• src/openhuman/threads/todos/runs.rs — the autonomous-run ledger.
• src/openhuman/agent/tinyagents/journal.rs + tinyagents_store/journal/ — per-run event JSONL.
• src/openhuman/agent/tinyagents/reaper.rs — how a run is moved to Cancelled.
• src/openhuman/security/approval/gate.rs — the TrustedAutomation arm that documents the missing review surface.
• Record what a trigger's verdict actually led to #5408 — recording what a trigger's verdict led to; this surface renders it.