When I wrote code by hand, a change took me hours. By the time it was done, I knew it by heart: what called what, which path a request took, which test covered which piece. I had built that picture while typing, so reviewing it was easy.
My coworkers' code was easy to keep up with too. They moved at the same pace I did, so pull requests came in slowly enough that I had time to understand each one.
Now a coding agent writes most of the code. A change that took an afternoon lands in minutes, then the next one, then the next. The code changes so often that I no longer build a mental map of it the way I used to.
So reviewing became the hard part. I'd open the changed files one by one and rebuild in my head the picture I used to get for free: what calls what, which path a request now takes, which test covers which piece. The diff showed me lines. The shape of the change was left for me to reconstruct, over and over, for code I never lived with.
Then it hit me: the agent already knows how its change fits together. What if it just wrote that down?
What I built
task-view is a small CLI. When an agent finishes a task, it writes a review file describing its change: the functions, components and tests it touched, why, and how they connect. task-view checks every file, line and commit in it against Git, then turns it into one offline HTML page:
- a graph of the changed pieces and how they relate,
- the flows through the change, step by step, with their tests attached,
- a diff map that flags any changed line the agent didn't describe,
- and an inspector: click any piece to see its diff next to the agent's reason for it.
Now I open one page, see the whole change at a glance, and spend my attention on the question that matters: is this code right?
It works with any language and any agent that can follow written instructions.
Set it up in two commands
1. Install the binary (macOS / Linux, no sudo, only needs git):
curl -fsSL https://raw.githubusercontent.com/gtindo/task-view/main/scripts/install.sh | sh
On Windows (PowerShell):
irm https://raw.githubusercontent.com/gtindo/task-view/main/scripts/install.ps1 | iex
2. Add the skill so the agent writes a review after every task. For Claude Code, install it once for all your repos:
mkdir -p ~/.claude/skills/task-review
curl -fsSL https://raw.githubusercontent.com/gtindo/task-view/main/skills/task-review/SKILL.md -o ~/.claude/skills/task-review/SKILL.md
Any other agent that reads skills takes the same SKILL.md.
That's it. Give the agent a task as usual; when it's done, it hands you a command like:
task-view open .review/change-<task>.json
Top comments (0)