How do I know what my coding agent actually changed before I merge?
Put what the agent says it did next to the list of files that actually changed. Read first the files nobody mentioned, and any file that touches sign-in, payments, permissions, secrets, migrations or deploy config. You can do that by hand with Git. Or paste the pull request link into BuildGist and get the same list in about 30 seconds, with no sign-in and no install.
A ready-made example opens on the same page if you just want to look.
The usual way: scroll the transcript, then read the raw diff
Most people scroll back through the chat to remember what they asked for, then read the diff top to bottom.
Both cost time.
- The transcript is long. It mixes your requests, the agent's plans, tool output and dead ends. When a session compacts or restarts, the part you need may be gone from the chat.
- The agent's summary is the agent's own account. It is usually right. It can also leave out a file it touched on the way, or say tests were added when none are in the diff.
- The raw diff has every line, but in file order, not risk order. A one-line change to an auth check sits between a renamed variable and a reformatted test.
What you want to know before you merge is narrower: what did the agent touch that it did not tell me about, and which of those files can hurt me?
Do the check by hand
- Get the file list. On a branch,
git diff --stat main...HEADshows every file changed since the branch leftmain, with its size. On GitHub, it is the Files changed tab. - Get the claims. Copy the PR description, the commit messages, or the agent's last message in the session.
- Cross off every changed file the claims name. What is left is the work nobody described.
- Mark the risky ones: paths with auth, login, session, token, payment, billing, permission,
migration, schema,
.env, workflow or deploy in the name. - Check each test claim. If the summary says tests were added or pass, look for a test file in the diff or a test run in the session.
- Read those files first, line by line. Then read the rest.
This works for Claude Code, Codex, Cursor or any other agent, because it only needs Git and the words the agent wrote. A fuller checklist is in how to review AI-generated code before you merge.
What BuildGist does with the same two inputs
BuildGist runs that cross-check for you on /try.
- Paste a public GitHub pull request or commit link. BuildGist reads the PR title, description and commit messages, and compares them with the files in the diff.
- No PR yet? Paste a Claude Code, Codex or Cursor session, or a git diff. The check uses your requests and the agent's last words as the claims.
The result leads with three lists:
- Files or tests the claims name that are not in the diff.
- Files the diff changes that the claims never mention, risky ones first.
- Test claims with no test file in the diff and no test run in the session.
That check is plain matching. No AI writes it. Under it, BuildGist adds a plain-English explanation, written by AI, of what changed and what to check. For a pasted session it also builds a short handoff you can copy into your next agent, so a fresh session starts with the task instead of a blank page. The handoff is built with no AI call.
What you paste is masked for keys and tokens, used once, and dropped. BuildGist keeps neither the paste nor what it wrote. Privacy has the details.
Side by side
| Question before merge | Transcript and raw diff | BuildGist |
|---|---|---|
| What did I ask for? | Scroll back, if the chat still has it | Your requests, from the pasted session |
| Which files changed? | git diff --stat | The same list |
| What changed that nobody mentioned? | Cross off by hand | Listed, risky first |
| Did the tests it claims exist? | Search the diff | Flagged when missing |
| Is the code correct? | Read it and run the tests | Read it and run the tests |
What it does not do
BuildGist does not prove the code is correct. It does not run your tests or your app. The check matches file names and test claims, so a summary that names a file can still describe the change wrong, and that file will not be flagged. The explanation is written by AI and can be wrong too. The diff is still the thing you are agreeing to ship. For auth, deletion, payments, dependencies and config, the lines win. PR links work for public repositories only.
If the change is a few lines in files you know, and you watched the agent do it, the diff alone is enough. When each one is enough is in BuildGist or the git diff. If your session already compacted and the request is gone, see what survives compaction. Already use an AI code reviewer? See BuildGist vs an AI PR reviewer.
Questions
Does BuildGist prove my agent's code is correct?
No. It shows what changed, what nobody mentioned, and which files are risky, so you know where to look. Correctness still comes from reading the code and running the tests.
Does it work with Claude Code, Codex and Cursor?
Yes. A pasted session can come from Claude Code, Codex or Cursor. A PR link works for any public GitHub pull request, whichever agent or person opened it.
Do I need to install anything?
No. The check runs in your browser at buildgist.com/try. BuildGist for Mac does the same for each meaningful run on your own machine, without pasting.
Sources
- Git: git diff (the
--statoption and theA...Bform), checked November 13, 2026. - BuildGist's try page and privacy page (what the check reads and what is kept).
What you need
- A Mac with Apple silicon and macOS 12 or later.
- Claude Code, Cursor or Codex.
- A Git project. No commit, push, remote or GitHub repository is needed.
- A GitHub account, to sign in. BuildGist does not access your GitHub repositories.
Codex needs you to approve BuildGist once, in Codex. BuildGist shows the steps. Setup is in Getting started.
To explain a run, your computer sends BuildGist what that run needs, and BuildGist passes it to Anthropic. BuildGist does not keep this material. It does not send your whole repository, files the run did not change, or files named like secrets, such as .env files and keys. It keeps the explanation and short cited excerpts of changed code, at most 12 lines each. Secret masking is best-effort. Privacy has the full list.
Every account gets 7 days of BuildGist Pro free, with no card. The 7 days start at your first explanation or Continue handoff. Nobody is charged when the trial ends. After that, BuildGist Pro is $12 per month, plus any applicable tax shown at checkout. See pricing.
Try BuildGist Download for Mac
No Mac handy? Try it in your browser: paste a session or a git diff and see the explanation.
Apple silicon, macOS 12 or later. United States, 18 or older.
Related
- What did Claude Code change?
- Switch from Claude Code to Codex
- Understand code an agent wrote
- Hand off Claude Code to Cursor
- Review what Codex changed
- You’ve hit your session limit in Claude Code
- You’ve hit your usage limit in Codex
- Keep context across agents
- Review AI code before merging
- BuildGist or the git diff
- Claude Code session history
- Undo Claude Code changes
- BuildGist vs CLAUDE.md, AGENTS.md and HANDOFF.md
- Keep context from Claude Code to Cursor
- Resume an agent session in another tool
- BuildGist vs Conductor
- Codex context window full
- Undo Codex changes
- Cursor usage limit reached
- Review Cursor agent changes
- Codex session history location
- Your conversation is too long in Cursor
- One default coding agent, or two?
- Tools that summarize agent sessions
- Switch from Codex to Claude Code
- Move from Cursor to Claude Code
- Switch from GitHub Copilot to Claude Code
- Move from Windsurf to Cursor
- Switch from Gemini CLI to Claude Code
- Coding agent crashed mid-task
- Claude Code lost context after compacting
- What commands did the agent run?
- Turn an agent session into a PR description
- Which coding agents read AGENTS.md?
- Switch from Cursor to Codex
- Export a Claude Code conversation
- Keep the dead ends when you switch agents
- Write a handoff note for a fresh session
- Usage limit hit in the middle of a task
- Mac apps that explain agent runs
- Explain a Claude Code session
- BuildGist vs an AI PR reviewer
- Review what Claude Code changed
- Review what Copilot changed
- Review what Windsurf changed
- Review what Gemini CLI changed
- Review what Devin changed
- Review what Cline changed
- Review what Aider changed
Questions: kayvanandre@gmail.com.