Bottom line
- Do not stop for normal coding work, schema updates, migrations, config changes, dry runs, testing, logs, or routine deploys.
- Stop for
edgespark loginwhen the user must open the browser login URL. - Stop for secret entry when the user must provide the real secret value through EdgeSpark’s secure flow.
- Stop when enabling a new social OAuth provider requires the user to prepare provider credentials and external app setup first.
- Stop before clearly destructive actions such as dropping schema, deleting bulk data, deleting secrets, or targeting the wrong project.
Stop only for these cases
Never ask the user to paste a secret into chat or terminal output.
How to ask when you must
- Be brief.
- Explain the exact next action.
- Explain the consequence in plain language.
- Ask one thing, not five.
Communicating a decision point
“This step needs your browser. Please open the EdgeSpark login URL and approve it. After that, I can continue.”
“This feature needs a STRIPE_WEBHOOK_SECRET. Please add it through EdgeSpark’s secure secret-entry flow. Do not paste the value here.”
“To enable Google login, please first prepare the OAuth app in Google and the client credentials it requires. After that, I can update the EdgeSpark auth config.”
“The next command will drop existing table data. That is destructive and may be irreversible. If you want me to continue, I’ll run [action] next.”
See also
The harness model
What the platform enforces automatically — decisions you never need to make.
Deploy and test loop
The current deploy and test workflow, including why dry runs matter before live deploys.
Handling errors
How to diagnose and fix errors without stopping for human input.
AGENTS.md reference
How project-specific rules in the instructions file extend these defaults.