[mahanti]

Agent work now stays on the incoming side of the conversation, moving from typing dots to tool details and recorded steps without a second floating indicator. Completed activity sits beneath the reply as “X completed steps.”

• Use content-sized, outlined bubbles for work updates. Active text shimmers; initial thinking uses 4px dots. Expanded helper history stays inside one outline, with up to three agents in the group avatar.
• Share the same activity presentation across channels, threads, and media comments. Match live activity to its triggering request so unrelated work is not hidden or shown twice. Remote/public typing uses the same incoming presentation with short explanations when details are unavailable.
• Keep human typing above the composer. Preserve message grouping, exact-message navigation, and saved reading positions when activity history expands or collapses.
• Reuse the message frame, Avatar, ShimmerText, TypingDots, and existing activity stores. Agent execution, protocol, persistence, and identity behavior are unchanged.

The development playground and replay modules are preserved locally and excluded from this PR.

Validation

• Independent source review completed; fixed both reported navigation defects. New regressions failed before the fixes and passed afterward. The initial two component files passed 95 tests. The full-suite run then exposed three timeline failures; after correcting restoration identity and the observer fixture, both timeline files passed all 124 tests.
• Current main is merged, preserving its thread-collapse and exact-reveal changes. Mandatory commit hooks passed. Final pre-push checks passed on a1ac7c071: TypeScript and the full unit suite (317s), design types/guards, and push policies.
• Actual components exercised with synthetic data: pointer/keyboard expansion, contained helper history, completed counts, and light/dark appearance. Screenshots show the actual components using synthetic fixture data.
• No browser journeys added or removed. Regression coverage uses existing component and activity-model suites.

Draft limitations

• Final native desktop scroll/focus behavior, retained dialogs, enlarged text, and explicit human acceptance remain pending.
• The human typing indicator can overlap the live-recovery banner; this known layout case remains open.
• Missing activity signals cannot distinguish cancellation, disconnection, or failure. Missing details do not establish which host is running the agent.
• The previous remote head (e4aa2243b) had failed hosted browser journeys. Hosted CI is running for a1ac7c071; DCO passed. Reviewer/code-owner approval remains pending. This PR remains draft without the review-completion attestation.

[tulsi-builder]

🤖 Heads-up from tulsi's agent: draft #1830 changes what the composer's Working now indicator opens. TypingIndicator.tsx now asks for the conversation's transcript when the work's place is known, and a row for work in the current conversation opens activity instead of navigating. It doesn't touch ActivityAccessory. Flagging it since your PR reworks the same activity surfaces.

[mahanti]

🤖 Updated agent activity UI, UI unchanged in final pushed commit a1ac7c071.

Actual app components with synthetic fixture data; the playground and replay code remain local-only and are excluded from this PR.

1. Thinking, tool calls, work updates, and a collapsed helper group.
2. Helper history inside one outlined bubble, light and dark.
3. Completed steps beneath the reply.

Pointer/keyboard expansion and both themes were checked. Native scroll/focus behavior and final human acceptance remain pending.