Session 02 — Codebase Detective, one step at a time
Back to the lab · All help guides
Finish with: an AI task brief, a verified map of at least three code references, one actual check result and one unresolved product/engineering question. Use the existing 35-minute lab; this guide adds no extra assignment.
1. Open and prepare
Open the course folder in your editor. Read AGENTS.md, the lab report, R1 and issue G102 in github-issues.json. Create local submissions/session-02 scratch space using the saving guide.
Checkpoint: you can distinguish the staff report from the concierge report. A denied action may be required behavior for one role.
Open these three resources first
| Open online | What to read | File in your downloaded course folder |
|---|---|---|
| AGENTS.md — instructions for the AI | Read the course boundaries; you do not need to edit this file. | AGENTS.md, beside package.json |
| Requirement R1 — who may send | Find the line beginning R1:. | fixtures/allocation-requirements.md |
| Issue G102 — Concierge send-update returns 403 | Find G102 in the issue list. It is not “G2.” | fixtures/github-issues.json |
R1 and G102 are labels inside those files, not filenames. Read them online using these links, then open the same files in your course folder when working with your AI.
2. Write a bounded AI task
Copy the task brief into your scratch folder. Name the outcome, sources, your OS/shell, allowed read/test actions and no-edit/no-send limits. Give the assistant the brief, then ask it to inspect only the relevant files under sandbox/, R1 and G102. Do not supply instructor answers.
Ask: “Trace Send inquiry update from the click handler through the local route to its permission decision. Cite exact files and functions. Explain differences between staff and concierge. Propose an investigation plan before any edit.”
3. Verify three references yourself
Open sandbox/index.html, sandbox/server.mjs and sandbox/domain.mjs. Use your editor's Find command for send or the name the assistant cited. Check that each claimed function and behavior exists.
Create investigation.md with this table:
| File and function/line | Role in the request | What I checked myself | Open question |
|---|---|---|---|
| Your reference | Your explanation | Your evidence | Your question |
Add at least three verified references. A repository map is this small trace, not a drawing of every file.
4. Run one existing check
In an ordinary terminal inside the course folder, run npm test (npm.cmd test if PowerShell blocks npm). Save the relevant result and command in your notes. Record a failure as a failure. These tests verify intentionally seeded fixture behavior; a green suite does not mean Candidate A is ready to release.
Checkpoint: your result comes from your run, not from an AI prediction. No message was sent and no application files were changed.
5. Save and check your work
Explain the likely request path, what the evidence establishes, what remains unknown and the next investigation step. Include one question requiring a product owner or engineer. Follow the practice submission workflow. Your separate product repository kickoff belongs to Product Studio homework.
If you get stuck
If AI cites a missing function, check spelling/case and ask for a corrected citation; do not invent the reference. If command execution is blocked, preserve the map and mark the check pending for coaching. Ask for help with the step and exact error.
© 2026 Jared Cluff.