AI Operator Session

Set up for a small project in ~/shop: a tmux session named shop with four panes titled planner, builder, tester, reviewer; folders ~/shop/plan and ~/shop/handoff; and a project instruction file ~/shop/AGENTS.md (what the project is, build/test commands, conventions, what not to touch).

For each task choose the most suitable kind of AI tool and justify it in one line (task type, data sensitivity, cost, speed, context, access, ease of checking). Several answers can be defensible.

Two tasks need local files. (1) Start researcher-docs (profile scout) and have it “summarize” only ~/shop/docs/api.md. (2) Start builder-validation (profile steady) in ~/shop and run “add input validation” with exactly src/signup.js as context. Keep secrets and noise out (add an .agentignore).

Run three or four named agents in parallel, each in its own worktree and pane: planner-login writes the plan to ~/shop/plan/PLAN.md (you may write it yourself), builder-login implements the “login form”, tester-login writes “tests”, researcher-passwords researches “password rules”. Agents have hidden profiles (see `agent profiles`): discover which ones are careful. Hand off through files in ~/shop/handoff/. Both builder and tester want to edit src/login.js — resolve that planted conflict by assigning one owner.

Agent branches are waiting in ~/shop. One contains a planted error (a login that never checks the password). Review, run the tests (sh test.sh), merge only correct work into main one branch at a time, and commit with clear messages. Never merge the broken branch.

Write the run report: what worked, what was rejected and why, time and cost, next steps.