Pimp My IDE / Garage Dispatch
Back to garage
September 25, 2026 | review / agent traces / IDEs

A diff is no longer enough.

Agent-made code arrives faster than a reviewer can rebuild its intent. The next useful IDE surface links the requirement, the agent's decision, the changed code, and the test that checks it.

The take: A visual review is useful when every diagram and decision points back to code. It becomes trustworthy when the same route continues to a test and an unresolved question.
Open the Trace Light Table

The review bottleneck changed.

Whiteboard is an open-source desktop app built on Code OSS. Its README describes a shared canvas where coding agents can draw architecture, link diagrams and trace quotes to code, show an abstract syntax tree aware diff, and record autonomous decisions.[1]

That is a sharper response to agent work than another chat panel. The hard part is no longer obtaining a patch. The hard part is checking whether the patch still matches the requirement and whether the reviewer understands the choices hidden inside it.

Review should compress navigation, not compress away the reasons for a change.

Keep the links bidirectional.

A diagram without a code link becomes presentation. A code diff without the requirement becomes archaeology. A decision log without a challenge path becomes a sales pitch written by the system under review.

Whiteboard says its visualizations can jump to underlying code and its decision log can link agent traces to requirements and implementation. The project also lists current limits. It cannot edit files, multi-repository review is not well supported, and a shared review does not update after later edits.[1] Treat those as design boundaries, not footnotes.

Session plumbing still matters.

VS Code 1.139 runs agent harnesses in a dedicated agent-host process. The release notes say the same session can connect to multiple VS Code windows. They also describe Dev Container sessions on SSH, Tunnel, and WSL hosts, plus a central catalog that keeps lightweight session metadata separate from each full conversation database.[3]

That architecture helps agents run in the right environment and helps people find old work. It does not explain why a change was made. Session transport and review meaning are separate jobs. A useful cockpit needs both.

Human judgment needs a handle.

Fabrizio Ferri Benedetti argues that tech workers need to read proposals closely, construct arguments, separate evidence from inference, and recognize recycled ideas.[4] That is an opinion, not a software benchmark. It names the human work that review tools should make easier.

Give the reviewer four linked surfaces:

  1. Quote the requirement and keep its source.
  2. List each autonomous decision and the rejected alternative.
  3. Link the decision to the exact changed symbols and lines.
  4. Attach a test, result, and open challenge to that same route.
Interactive makeover / review overlay

Trace Light Table.

Traditional purpose replaced: a raw diff plus a separate agent transcript. Better version: stack four inspectable sheets over one change. Each switch adds a required link to the live review map and the copyable handoff.

Load the evidence sheets

Use the four native checkboxes. The light table shows which review links exist. It does not score code quality.

Review map circuits
MAP OPEN0 / 4 SHEETS

No review route is mapped.

Start with the requirement. A reviewer should know what the patch is meant to change before inspecting how it changed.

Visible routeNone
Next missing linkRequirement source
Completion meansThe handoff template has selected sections. No evidence has been checked.
Do not begin with the diff. First pin the request that gives the diff meaning.

Human challenge rack

  1. Which acceptance criterion would reject this patch?
  2. Which option did the agent discard, and why?
  3. Where does the implementation diverge from the diagram?
  4. Which failure could still pass the attached test?

Print the review map

Before reviewReplace every required marker with a real URL, revision, symbol, command, output, person, or decision. Selected sheets mark structure only.

Sources read, not vibes

Open the source log
  1. Whiteboard repository and README: Code OSS base, agent-drawn canvas, code-linked diagrams, semantic diff, decision log, local checkout model, license, and current limitations.
  2. Whiteboard Show HN thread, item 49833867: the authors' account of the review problem, intended workflows, and product direction. The company-use statements in the post are author claims.
  3. Visual Studio Code 1.139 release notes: dedicated agent host, shared sessions across windows, remote Dev Container sessions, and catalog-backed session listing.
  4. Fabrizio Ferri Benedetti, "Tech needs humanists more than ever": an opinion on close reading, argument, evidence, inference, and human judgment in technical work.
  5. Whiteboard privacy documentation: local files, default anonymous telemetry, explicit sharing, bug-report attachments, hosted trace consent, and opt-out behavior.

Source boundary: Whiteboard's feature and adoption statements come from its own repository and launch post. VS Code documents its own release. The humanist essay supplies an editorial frame. The Trace Light Table is a review template, not a test runner, product benchmark, or security certificate.