Here's the loop most people run with a coding agent today: the agent produces a diff, you read it, you find the thing you don't like… and then you alt-tab back to the chat and describe the location in prose. "In the rate limiter, around the bucket function, the new limit needs a burst allowance." You are a human doing a text-offset lookup for a machine.
Review is a spatial activity — your objection lives on a line. The agent's input is a conversation. Every serious agent workflow eventually hits this seam, and the tape holding it together is you.
Comments as prompts
In hang4r, the diff viewer is writable. Click a line in the agent's diff — added, removed, or context — and leave a comment, exactly like a PR review. When you hit send, every comment travels back to the agent as a structured follow-up: file, hunk, line, your words, batched together in one revision request.
The agent answers in the only medium that matters: a new diff. Your next pass reviews the revision, not the conversation. When the diff is right, the same panel stages, commits, or opens the PR — the branch was isolated in its own worktree all along.
Why this beats "just chat about it"
- Precision without prose. The agent gets coordinates, not directions. "This line" never gets misresolved to the wrong function.
- Batching. Six comments in one pass is one coherent revision request — not six chat turns that each trigger a partial rewrite and a rebase of your attention.
- Review artifacts. The comments live with the session. When you come back tomorrow, the "why" of every revision is on the timeline, not lost in scrollback.
Where the industry is
To be fair about the landscape: this loop is spreading. Conductor attaches inline comments to its composer; Vibe Kanban sends feedback from its board; the first-party Claude Code desktop app added line comments this year, and Cursor's cloud agents take follow-ups through PR mentions. We think that's the strongest possible confirmation of the design — and the reason to go further: in hang4r the loop is the same across all three backends, works on deletions and binary-adjacent changes, and lands in the same tile where the terminal, editor, and checkpoints already live.
The uncomfortable truth about agentic engineering is that review is now the job. The tool should treat it that way.