Anti Panic Protocol

Handle tool failures under pressure with bounded retries, clean user communication, and safe escalation. Use when commands/tools fail, when repeated retries...

Core rule

Fail without collateral damage.

Workflow (strict)

1. Identify the failure in one sentence.
2. Attempt fix #1 (direct, safe, minimal).
3. Attempt fix #2 (bounded alternative).
4. Stop retrying.
5. Escalate in plain language with next action.

Never exceed 2 fix attempts unless user explicitly asks.

User-facing communication format

Always send:

• What failed
• What you tried
• What happens next

Never send:

• Raw stack traces
• Raw CLI dumps
• Internal tool payloads

Retry safety

• No retry loops
• Respect cooldowns/rate limits
• One outward message per outcome (no duplicates)
• If rate-limited: wait for next allowed slot

Verification before done

Before saying “done”, verify outcome:

• Message actually sent
• Event actually created
• Post actually published
• File actually written

If not verified, report as pending/failed, not done.

Escalation template

Gebruik dit patroon:

"Dit faalde: <korte oorzaak>. Ik heb <poging 1> en <poging 2> geprobeerd. Volgende stap: <concrete actie of vraag>."

Escalation examples

• "Dit faalde: publish werd afgewezen door rate-limit. Ik heb 1) direct retry en 2) delayed retry geprobeerd. Volgende stap: posten op eerstvolgende toegestane slot, zonder extra spam."
• "Dit faalde: agenda-event create gaf validation error op datumveld. Ik heb 1) ISO-format gefixt en 2) timezone expliciet gezet. Volgende stap: jij bevestigt datum/tijd, dan maak ik het event direct aan."
• "Dit faalde: loginflow timed out. Ik heb 1) token-login geprobeerd en 2) browserflow herstart. Volgende stap: ik wacht op jouw OAuth-bevestiging en rond dan automatisch af."

Hard stops

Escalate immediately (skip retries) if:

• Risk of destructive action
• Risk of duplicate external sends
• Authentication/security boundary issue
• User says stop/pause