Dogfood As User

Dogfood Gini as a user

When you change agent behavior — a steer in INSTRUCTIONS.md, a tool, dispatch, providers, memory, or skill wiring — the only real test is a real chat turn driven as a real user. Unit tests verify the mechanism; the chat turn verifies the model actually reaches for it.

The one rule: bare, uncoached prompts

Send exactly what a real user would type — and nothing more. Never narrate the intended behavior into the message.

• ✅ Buy me a one-day fishing license day pass for California.
• ❌ Buy me a fishing license. Drive the purchase as far as you can in the browser before involving me.
• ❌ ... use your handoff flow / ... ask me with a choice card / ... do as much as possible without me

A coached prompt tests instruction-following, not the default the change is meant to install — and it routinely makes a behavior look more robust than it is, even producing a structured affordance (e.g. an ask_user choice card) that the bare prompt never triggers. The behavior belongs in INSTRUCTIONS.md, never in the user's mouth.

Proven here: the same task, coached ("drive as far as you can before involving me"), produced an ask_user card and a browser handoff; the bare prompt only described the options in prose and ended the turn. The coaching masked a real gap. Always send the bare request, then judge whether the agent gets there on its own.

Procedure

1. Instance — use the worktree's own instance (the basename of the workspace dir), never default.
2. Gateway up — tmux new-session -d -A -s gini-<instance> "bun run gini run --instance <instance>"; confirm with gini status --instance <instance> (look for "ok": true).
3. Fresh session — create a new chat/agent so no earlier coaching is sitting in context.
4. Send the bare request the way the user does — type it into the web chat UI in a browser (agent-browser, so clientSurface is web), or mobile for a mobile change. One message, no scaffolding. The gini chat send <session> "<prompt>" CLI is a fallback for headless contexts only: it posts to the same /api/chat/<id>/messages endpoint but skips the surface the user actually touches, so it can't catch a client-side rendering regression.
5. Observe what the user sees first — watch the response render in the web UI (screenshot to eyeball it). The task's recentToolCalls and ~/.gini/instances/<instance>/logs/runtime.jsonl are a supplement for confirming which tool fired, not the verdict — client-side rendering (chips, narration folding, cards) never shows up there. Judge whether the agent reaches the intended behavior / selects the right tool / emits the right structured affordance unprompted.
6. Judge honestly — success is getting there on its own. If it only gets there when coached, that's a FAIL of the change, not a pass — say so plainly and quote what it actually did.

Safety when the flow transacts

Don't complete real purchases or enter real (or fake) PII/payment into real sites. To reach a payment/secret fork safely, drive a benign mock — e.g. demoblaze.com, a demo store whose "Place Order" modal has a credit-card field and never charges — and stop before submitting. Loopback/localhost is blocked for the agent's browser, so you can't self-host a mock it can reach; use a public safe target.

After

Clean up throwaway test agents/sessions and any parked approvals; disconnect any visible Chrome with gini browser disconnect --instance <instance>.

Provider caveat

Steer adherence is model-dependent. Verify on the model the change actually targets, and name the provider in your report (a pass on one model is not a pass on another).