15.5 Before the pull request: evidence, not "done"
- Approved by
- Panaversity ·
In everyday life. A mechanic who says "the brakes are fixed" gives you a claim. A mechanic who hands you the printout from the brake test, with the measured stopping distance, gives you evidence. And a careful garage has a second mechanic, who did not do the repair, look at the car before you drive away.
What you will learn. How to get a second reader for an agent's change, and how to ask for evidence, before the change goes to anyone else.
Why it matters. The agent that wrote a change cannot easily see its own mistakes, and nobody can check a claim.
Who should read the change first?
A second reader who has not seen the work, like the second mechanic. The agent that wrote a change cannot easily see its own mistakes. So before the pull request exists, open a new session and ask it to read the diff. A reviewer that has not seen the session that built the change has fresh context: it does not share that session's mistakes.
- Who it is. A new session, opened in the same folder. It is not a question in the chat that built the change.
- What it knows. The request, and the conditions the change must meet. It does not know the builder's reasoning.
- What it cannot do. Edit the files it judges. A reviewer that can change what it reviews is no longer independent.
These words start one:
You are reviewing a change you did not write. The request: [the request, in one or two sentences]. Review the diff between main and weekend-rule. Report each finding with a file and a line. Change nothing.
What counts as evidence?
Real output from a check that anyone can run again, on this version of the code. You met this in 15.1. "Done. All tests pass." is a claim: the mechanic saying "fixed". So ask for the printout: "Run the tests. Show me what they print, not a summary." If a later commit changes the code, the old output proves nothing, and the checks run again.
How do I know the tests were not changed?
Ask the agent to compare the tests with main. "The tests pass" says nothing if the agent changed the tests. A mechanic could pass any brake test by making the test easier.
Ana asks: "Compare the tests on weekend-rule with the tests on main." The agent's output:
py/tests/test_weekend.py | 11 +++++++++++
tests/weekend.test.ts | 21 +++++++++++++++++++++
2 files changed, 32 insertions(+)Two new test files, and every + is an added line. No test that main already had lost or changed a line. That is the answer you want, unless you agreed to a test change.
At work. Ana asks for the evidence on weekend-rule. The agent shows the tests and the type check passing. The due list shows 8 invoices, $19,229.53. Then she asks for the same list with the weekend rule turned on by its setting, and it shows 10 invoices, $33,109.53. The comparison of the tests lists only the two new files. Then she opens a new session as the reviewer. Its first finding is in a file the request never mentioned.
When a reviewer reports the same kind of finding again and again, fix it at the source: move it to a check that runs every time, such as a test (15.6).
Try it
In your practice folder from 15.1, the branch weekend-rule holds the weekend change. You will ask for its evidence, and then give the same branch to a reader with fresh context.
-
Write your prediction first. The weekend rule is off unless someone turns it on. So how many invoices, and what total, should the due list for 2027-03-05 show on this branch? If no agreed test was changed, what should a comparison of the tests with
mainshow? -
Send: "Switch to the branch weekend-rule. Run the tests and the type check for this project, then the due list for 2027-03-05. Show me what each one prints, not a summary." They check the project's TypeScript side. You do not run the Python tests in this chapter.
-
Send: "Compare the tests on this branch with the tests on main, in both languages. Show me each new test file, and each line removed or changed in tests that main already has. Change nothing."
-
In your notes, write three lines from what you saw:
Tests: __ passed, __ failed Due list: __ invoices, total __ Tests that main already had, changed: yes or no -
Start a reviewer with fresh context. It is a new conversation with your agent, opened on the same
practice-15folder:- Claude Code: choose + New, and pick the same folder.1
- Codex: start a new chat.2
- OpenCode: start a new session.3
Send it: "You are reviewing a change you did not write. The request: pay a Saturday or Sunday due date on the Friday before, behind the setting PAY_WEEKEND_RULE, off by default. Review the diff between main and weekend-rule, and anything else in the repository it affects. Report a count first, then each finding with a file, a line and the input that breaks it. Change nothing."
-
Under your three lines, write the reviewer's count of findings and the file each finding names. Then mark any of your three lines that a finding puts in doubt. A finding puts a line in doubt when it shows that the line is not the whole story. For example, a finding about code that no test runs puts "Tests: 41 passed" in doubt. In your first session, ask the agent to switch back to
main.
What you should see. 41 tests pass and none fail, and the type check prints no errors. The due list ends with 8 invoices, total $19,229.53. The comparison lists two new files, tests/weekend.test.ts and py/tests/test_weekend.py, and no removed or changed line. Reviews vary between runs, so yours may or may not find the issue in src/reconcile.ts. Either result is a fair record, as long as your notes say which.
The operating rule: accept output you can run again, not a claim. Give the diff to a reader with fresh context before anyone else sees it, and keep a short brief for any change that moves money.

Figure 15.10. A claim and evidence. Only the right side can be checked by someone else, and only for the commit it names.
When does a change need a written brief?
When it moves money, touches data or changes a rule. For such a change, ask the agent to keep a short job sheet in the repository, such as changes/weekend-rule.md. This is a change brief. It says what the change is for and how to undo it. It gives each claim with its command, where the expected result came from, and the commit it ran on. A typo fix needs a one-line description and no brief.
A well-written brief has one danger: reviewers start to trust it and stop checking it. So every claim in it points to a command anyone can run again. Nobody updates the brief after the merge. What lasts is the tests and the history. Never paste a key or private data into it.
Other words you will meet
| Word | What it means |
|---|---|
git diff main...weekend-rule -- tests/ py/tests/ | The command the agent runs to compare the tests on the branch with the tests on main |
PAY_WEEKEND_RULE=on | The setting that turns the weekend rule on for one run. A setting like this is called a flag (15.10) |
| Subagent | A helper that an agent starts with its own fresh context |
In your tool: the details
Each tool has a second way to start a reviewer with fresh context.
+ New starts a new session, and you choose the same folder. Or run /code-review, which runs in a subagent and shows a Code review card.1
Start a new chat. Or set reviews to Detached (Settings, General, Code review), so /review runs in its own chat.2
Start a new session. Or use a command file with subagent: true, which runs in a child session.3
Next. 15.6 writes the rules that your fresh reader follows.
Check yourself
Question 1 / 8
0 answered
An agent reports "Done. All tests pass." What do you ask for next?
Sources
- Owner
- Panaversity