[wip][harness/env][2/n][buzz] visualize inline agent work streams

[yuchenhe-hai]

Summary

This adds an observable model work stream directly to Buzz channel timelines so users can follow an agent turn where it is happening instead of leaving the conversation to inspect a separate debug surface.

• Place active and completed agent turns inline in the channel timeline.
• Anchor a completed turn immediately before the exact message produced by send_message.
• Add a structured Flow view organized into Input, Explore, Decide, Act, and Deliver phases.
• Retain a detailed Trace view for the chronological observer event stream.
• Render ACP tool names, summarized arguments, and returned results for calls such as sample_model, get_messages, get_thread, and send_message.
• Preserve current channel virtualization, day-divider behavior, Huddle transcript rendering, and identity/access-badge behavior.

Motivation

Buzz already receives observer events while an assistant is working, but the channel UI previously exposed only coarse activity. That made multi-round agent behavior difficult to understand: users could see a final message without seeing which context was loaded, what tools were used, or how the turn moved toward delivery.

The new timeline presentation makes that work legible without exposing private chain-of-thought. It surfaces operational facts available in the observer stream: prompts and context metadata, tool activity, safe summaries, returned results, phase progress, and the relationship between a trace and its delivered message.

Behavior

Live turns

While an agent is working, its work stream appears at the live tail of the channel. The Flow view shows:

• current focus and working/completed state
• phase progress across Input, Explore, Decide, Act, and Deliver
• counts for inputs, findings, and actions
• safe prompt/context metadata
• tool calls and their observable results

Completed turns

When a send_message result can be associated with a rendered channel message, the completed work stream moves from the live tail and is inserted directly before that message. This keeps the explanation at the point of action and avoids duplicating the trace.

Flow and Trace

• Flow groups events into a readable account of what information entered the turn, what the model explored, which actions it took, and what it delivered.
• Trace preserves the complete chronological observer-event sequence for debugging and deeper inspection.

Implementation

The channel timeline hook combines observer-derived agent activity with the existing message model and returns:

• normal virtualized message entries
• per-message leading content for anchored completed traces
• trailing content for active traces
• the shared set of working-agent pubkeys used by composer activity

The virtualized timeline supports both leading content and trailing activity while retaining existing estimates and day-divider controls. Agent-session links continue to open the corresponding detailed session.

Related issue

None found.

Testing

• pnpm --config.verify-deps-before-run=false run check
• pnpm --config.verify-deps-before-run=false run typecheck
• pnpm --config.verify-deps-before-run=false run test (4,485 tests passed across 64 suites)
• pnpm --config.verify-deps-before-run=false run build:e2e

The branch is rebased onto the latest main, and both commits include DCO sign-off.