The diff panel gets serious about large diffs
Late July 2026 builds collapse large git diffs by default and add a collapse-all toggle — fixing the review step where agentic coding goes wrong.
24–25 July 2026 — two changes to the diff panel:
v0.0.29-nightly.20260724.890— large git diffs collapse by default for readabilityv0.0.29-nightly.20260725.898— a collapse-all toggle for the panel
Recent builds also added diff scope switching, letting you look at one change versus everything in a thread.
Why this is more than a UI tweak#
The failure mode of agentic coding isn't dramatic. Agents rarely produce something obviously broken — that would be easy to catch. They produce plausible code that quietly does the wrong thing: an edge case dropped, an error swallowed, a condition inverted somewhere reasonable-looking.
The only defence is reading the diff. And whether you read the diff is decided almost entirely by how much friction stands in the way.
An agent finishes a refactor. The diff is 1,400 lines: a regenerated lockfile, a snapshot update, formatting churn, and eleven lines of actual logic. If everything is expanded, finding those eleven lines means scrolling past 1,389 you don't care about. Most people, most of the time, will skim and move on.
Collapsing large diffs by default inverts that. You expand what you want to see rather than hunting through everything.
What you get now#
| Change | Effect |
|---|---|
| Large diffs collapsed by default | Generated files don't bury real changes |
| Collapse-all toggle | Reset a sprawling review in one click |
| Diff scope switching | One change, or everything in the thread |
Paired with earlier work — settled PR labels coloured on hover, settled threads staying accessible — the review surface has had a consistent run of attention.
Reviewing agent output well#
Since the tooling now supports it, the habits that matter:
Read the diff before the summary. The agent's description of what it did is a summary of its intent, not its output. Those diverge, and the diff is the only record of what actually happened.
Expand anything touching error handling, auth or data. Agents are good at happy paths. The interesting failures live in the branches nobody exercised.
Skim generated files, read handwritten ones. A regenerated lockfile deserves a glance. A change to your auth middleware deserves every line.
Roll back rather than argue. When a thread has gone somewhere wrong, checkpoints get you back faster than trying to talk the agent out of its approach. Worktrees and checkpoints →
The pattern in the changelog#
This isn't isolated. Read a few weeks of T3 Code releases and the diff and review surface comes up repeatedly — collapsing, scoping, PR label states, keeping settled threads available.
That's what a project looks like when the people building it are also running it on large real changes. The friction they're removing is friction they hit.
FAQ#
Does T3 Code collapse large diffs?#
Yes. Since the 24 July 2026 nightly build, large git diffs collapse by default so generated files don't bury the changes that matter. A collapse-all toggle followed on 25 July.
How do I review agent changes in T3 Code?#
Use the diff panel before committing. Expand the files that matter — error handling, auth, data access — and skim generated output. The panel supports scope switching between a single change and everything in the thread.
Can I undo an agent's changes in T3 Code?#
Yes. T3 Code checkpoints during a thread, so you can roll back to an earlier state, and everything is backed by git so ordinary git recovery also works.
Why does diff review matter so much with coding agents?#
Because agents typically produce plausible code rather than obviously broken code. Subtle problems — a dropped edge case, a swallowed error — only surface when someone reads the changes.