Two interaction modes

Both produce proposals that need review.

Choose the surface by the work

Use the controls available in the editor.

Sidebar and inline Chat

Sidebar Chat

• keeps a task thread visible;
• supports follow-up questions;
• works well for cross-file reasoning.

Inline Chat

• starts from selected code;
• shows an edit close to the source;
• fits a narrow change.

Quick and terminal context

A lightweight surface is useful when history is unnecessary.

Terminal context is useful for:

• a failed test command;
• an unfamiliar error;
• the next command to check the result.

Read the actual output before acting on the summary.

Context should be explicit

Weak:

Fix this.

Reviewable:

The weighted-average test expects 95 but receives 190.
Inspect the selected method and focused test.
Identify the defect without changing other methods.

Workspace, file, selection, terminal

Include the context the task needs.

Explain before changing

Ask for:

1. the current formula;
2. the role of each guard;
3. the failing example;
4. the exact mismatch.

An explanation can expose a wrong assumption before code changes.

Fix narrowly

Fix only weighted_average so it divides by total weight.
Preserve the method signature and length check.
Add no dependency.

Scope is part of correctness.

Generated tests need review

Check whether each test:

• states accepted behavior;
• would fail for the original defect;
• calculates the expected value correctly;
• avoids invented requirements;
• uses existing tools.

Check the expected behavior too. A generated test can encode the same mistake as
the code.

Refine one concern at a time

Explain
   ↓
Diagnose
   ↓
Fix one behavior
   ↓
Add focused tests
   ↓
Review the complete method

Focused turns make mistakes easier to locate.

Know when to restart

Continue the thread when:

• the task is unchanged;
• the useful context remains accurate;
• the current structure is sound.

Start fresh when stale assumptions keep returning.

Compare against the task

If policy allows two available model choices:

1. use the same synthetic prompt;
2. keep the same context;
3. run the same tests;
4. compare correctness and useful citations.

Judge the results by the checks, not how well they read.

Access policy

The lab requires Chat access allowed by your organization.

If Chat is unavailable, stop the lab and restore access.
Manual work does not replace the Chat and inline-suggestion practice.

If only a second model is unavailable, continue with Chat and compare the
assisted result with the reference solution.

Demo: weighted average

weighted_sum = sum(
    value * weight
    for value, weight in zip(values, weights)
)
return weighted_sum / len(values)

For [80, 100] with weights [1, 3], the denominator must be 4.

Failure to fix

Expected: 95
Actual:   190

The focused test gives Chat a concrete symptom. The reviewed fix changes the
denominator to the total weight.

Lab handoff

You will use one calculator project to:

1. reproduce the defect;
2. explain and diagnose it;
3. apply a narrow fix;
4. review generated tests;
5. use workspace and terminal context;
6. decide zero-total-weight behavior;
7. run the full suite.

Deliverable: corrected code, passing tests, and a Chat decision record.