Agent runs currently show their work as folded groups of raw tool calls in the chat timeline. While a run is active, a folded group makes the channel feel like nothing is happening, and it is hard to audit what the agent did at a glance. Expanding the group does not fix this: a wall of raw commands is hard to follow.

Show a clear, live, plain-language account of the run instead: what the agent is doing now, the meaningful outcomes so far, and anything blocking it. Keep the full tool calls and results available as the audit trail beneath that account.

## Acceptance

- While a run is active, the chat shows what the agent is doing now in plain language, without anyone having to expand a tool group. - Meaningful outcomes, such as files changed, checks passed or failed, issues/PRs filed, and decisions made, read as short statements rather than as the commands that produced them. - Blockers and failures stand out, including when the agent is waiting on a participant, a failing check, or missing access. - Each statement links to the tool calls behind it. Raw commands, output, truncation, and exit/error state stay inspectable and are never replaced by the summary. - The account survives reconnecting and reopening the channel, and it matches the durable run history. - Dogfood: follow a real multi-step run without expanding any groups, then audit one of its claims from its linked tool calls.

## Notes

- Build on live tool progress (#54) rather than a second progress store. #54 covers raw live output; this issue covers the participant-facing account of the run. - Open question: where does the account come from? The agent could narrate it, a separate model could derive it from tool events, or both. Whatever produces it must not change what was actually run or hide anything. - Related: #12 (attention for finished, failed, or waiting runs), #23 (rolling channel summaries).

Part of #98. Dogfooding core/QOL: #5. Product feedback from Nate.