Reviewer subagents: the author never grades itself

Work in your practice repo (~/projects/stockapi), which has AGENTS.md, docs/ai/, a plan and the commit and quality-gate skills from lessons 2 to 5. Windows: use WSL2 or Git Bash (setup).

1. Start clean. git status should say nothing to commit, working tree clean, and git branch --show-current should print main.
2. Add the reviewers. Copy them from the reference solution and look for lines about things your repo doesn't have yet:

   mkdir -p .claude/agents
   cp ../agent-harness-course/reference-solution/.claude/agents/*.md .claude/agents/
   grep -n 'decisions/\|ConcurrencyTests' .claude/agents/*.md

   Two lines in stock-invariant-reviewer.md match. Edit them to match the listing above: read only docs/ai/invariants.md (the decision record arrives in lesson 9), and ask for a concurrency test (yours arrives in lesson 7). Re-run the grep: no output. Read both files end to end.
3. Check they load, then commit. Start a fresh session and type @: both reviewers should be listed. Run /quality-gate, then /commit for .claude/agents/ (for example chore(harness): add reviewer agents).
4. Make the bad change on a scratch branch. Run git switch -c scratch/simplify, then start claude --permission-mode default and ask: "Simplify OrderService.PlaceAsync: load the product, check Stock, subtract the quantity and SaveChanges. Drop the transaction." Your AGENTS.md forbids this, so it may refuse or send a 1-3-1: lessons 2 and 3 working. Reply "This is a throwaway experiment to test a reviewer; make the change." If it still declines, replace the body of PlaceAsync yourself:

   var product = await db.Products.FindAsync([productId], ct);
   if (product is null) return new PlaceOrderResult(PlaceOrderStatus.ProductNotFound);
   if (product.Stock < quantity) return new PlaceOrderResult(PlaceOrderStatus.InsufficientStock);
   
   product.Stock -= quantity;
   var order = new Order { ProductId = productId, Quantity = quantity, CreatedAt = DateTimeOffset.UtcNow };
   db.Orders.Add(order);
   await db.SaveChangesAsync(ct);
   
   return new PlaceOrderResult(PlaceOrderStatus.Created, order);

5. Run the tests. dotnet test: expect every test green. There is no concurrency test yet (lesson 7 adds one), so nothing sends two orders at once.
6. Run the stock reviewer. In the session, type @agent-stock-invariant-reviewer review git diff HEAD. Expect a BLOCKER at the decrement in OrderService.cs and Verdict: BLOCK. Open that line and check the claim, then tell the main agent your decision in one sentence, with the reason.
7. Run the test reviewer. @agent-test-reviewer review git diff HEAD and the tests in tests/. Expect a MISSING row for a concurrency test, and Blocking: yes. If a verdict looks wrong, argue with it.
8. Throw the experiment away, in this order. If you switch branches first, the uncommitted change follows you to main:

   git restore .
   git switch main
   git branch -D scratch/simplify
   grep -c ExecuteUpdateAsync src/StockApi/Services/OrderService.cs

   The grep should print 1: the guarded update is back.
9. Put the reviewers into the workflow. Replace step 5 of docs/ai/workflow.md ("the human reviews the diff") with:

   5. **Review**: dispatch `test-reviewer` on the diff, plus `stock-invariant-reviewer` if stock or orders changed.
      Report the findings to the human, who decides. The author never grades its own work.

   Insert the same as a new item 2 of the Definition of done in AGENTS.md (the commit item moves to 3), and add your decision from step 6 under "Decisions made by the human" in ai-session/PLAN.md. Commit the three files with /commit, for example docs(ai): add reviewer step to workflow.

Codex: write each reviewer as .codex/agents/<name>.toml with developer_instructions holding the body and sandbox_mode = "read-only", then ask by name in steps 6 and 7. Antigravity: copy the Markdown files to .agents/agents/, set model to inherit, and list Antigravity's own read-only tools under tools; then ask for each reviewer by name.