OpenClaw Main Is Turning the Control UI Into an Intervention Queue

Three changes merged into OpenClaw’s default branch on 20 July point in the same direction: the Control UI is becoming a place to triage, intervene and keep useful agent interfaces alive, rather than merely watch chat transcripts.

The changes let agents raise temporary attention states in the session list, make runs say when they are waiting for an exec approval, and keep pinned MCP apps interactive on session dashboards. Together, they materially alter the supervision workflow for operators running several sessions.

They do not create an upgrade instruction today. All three merged after 2026.7.2-beta.3 was published, and npm still identified that build as the beta channel when this analysis was verified.

Operator decision

• Packaged stable and beta users: do not expect these controls yet. Wait for release notes that explicitly include them.
• Source-build testers: evaluate whether attention clears reliably, approval status follows the correct run, and dashboard tool grants remain appropriately narrow.
• Multi-session operators: start designing around an exception queue—blocked, approval-required and expiring work—rather than repeatedly opening every session.
• Security-conscious operators: treat interactive pinned MCP apps as delegated capability surfaces. Review each requested tool grant instead of treating a dashboard pin as harmless decoration.

Agents can now declare “I need you”

PR #111583, merged at 04:01 UTC on 20 July, extends the existing sessions tool patch action with three model-facing fields:

• statusNote: a sanitised sidebar message capped at 120 characters;
• attention: one of a curated set of icons—hand, key, alert, flag, lock or hourglass—or clear;
• ttlMinutes: a lifetime from one to 120 minutes, defaulting to 30.

The status is deliberately temporary. The merged implementation says it clears when its TTL expires, when the user reads the session, when the next human message arrives, or when the agent clears it. Attention also requires an active status note, which gives the icon a reason rather than leaving an unexplained alarm in the sidebar.

That is more consequential than a cosmetic badge. In a fleet of concurrent sessions, the scarce resource is operator attention. A session that can say “waiting for a password” or “review required” can move itself into an intervention queue without the operator polling every transcript.

There is an important limitation in the pull request’s own evidence: its author reports schema, gateway and UI tests, but explicitly says a live end-to-end model prompt against a development Gateway was not completed. Operators testing main should therefore verify the complete model-to-sidebar path rather than assuming unit coverage proves the workflow.

Approval waits stop looking like generic thinking

PR #111445, merged at 02:44 UTC, changes the chat working indicator to “Waiting for approval…” when an inline exec approval parks a run.

Previously, that state looked like ordinary thinking even though the Gateway could hold the tool for up to 30 minutes. The UI now consumes waiting-approval and approval-resolved lifecycle phases and correlates the state by approval and run identifiers.

The correlation detail matters. Heartbeat and cron approvals can share a main session with foreground work. The pull request says broad session-level inference was rejected because it could label an unrelated foreground run as blocked. Reload behaviour also remains bounded: the approval card is visible, but authoritative cross-mount run-state hydration is deferred to a future Gateway contract.

The operator benefit is immediate once this reaches a package: a spinner becomes an actionable state. The limitation is equally practical—after a reload, do not assume every status label can reconstruct the exact parked run yet.

Pinned MCP apps become live operating surfaces

PR #111524, merged at 02:51 UTC, completes a larger dashboard step. Pinned MCP apps can render as live board widgets, renew expired view leases and retain operator-granted access to their declared server tools.

The security boundary is part of the feature, not a footnote. The pull request says board cells host the existing sandboxed MCP app view rather than a raw iframe. Tool calls are rechecked against the current board revision, grant and app generation. Ungranted, rejected or pending widgets remain read-only, and transcript-reconstructed chat views do not inherit board interactivity.

For operators, the dashboard can therefore become a durable action surface instead of a collection of dead snapshots. The cost is governance: a pinned app with granted tools can do more than display information. Pinning, approving tools and trusting the app should remain three separate decisions.

The workflow taking shape

These are not three random UI improvements. They form a supervision loop:

1. an agent declares a short-lived reason it needs attention;
2. a parked exec run identifies approval—not “thinking”—as the blocker;
3. a useful MCP interface can remain available on the session dashboard for follow-up action.

That shifts the Control UI from transcript navigation toward exception handling. The best operator interface for many agents is not a wall of chat windows; it is a queue that says what needs intervention, why, and for how long.

The design also contains sensible anti-staleness choices: agent statuses expire, reading or replying clears them, approval labels are run-correlated, and dashboard capabilities are revision-bound. Those safeguards deserve live testing because stale attention and stale authority are exactly what make multi-agent consoles noisy or dangerous.

What is available now?

At verification time, npm’s channels were:

• latest: 2026.7.1-2
• beta: 2026.7.2-beta.3
• alpha: 2026.5.19-alpha.1

The current beta GitHub release was published on 18 July, before these pull requests merged on 20 July. A merged pull request is evidence about main, not evidence about an installed package.

Bottom line

OpenClaw’s Control UI is developing the right primitives for supervising concurrent agents: explicit attention, visible approval waits and durable interactive work surfaces. The cluster crosses the editorial threshold because it changes how an operator can find and resolve blocked work.

But it remains unreleased main-branch work. Watch the next package boundary, test the lifecycle edges if you build from source, and keep MCP dashboard grants narrow.

Verification note: this is source analysis of merged pull requests, changed source contracts and official GitHub/npm release metadata. OpenClaw Academy did not run an unreleased source build or reproduce these behaviours in a live Gateway.