Design retries, degradation, and recovery

Map failure classes for POST /v1/support/turn: retryable vs fatal vs needs-human. Retries need budgets, jitter, and idempotency rules aligned to tool args bind to session customer_id; refunds never auto-executed. The chapter’s signature failure — model passes attacker order_id and tool executes — must take a deliberate branch, not a generic catch-all.

Implement the degrade/rollback/refuse behavior customer asking where order A12 is without leaking other customers' orders needs when CAP-SUPPORT-TOOL-09 repeats. Prefer scoped controls (one flag, one weight, one tenant, one secret version) over fleet-wide restarts. Preserve evidence; do not delete logs to “clean the demo.”

After containment, check unauthorized_tool_calls == 0 and citation_rate ≥ 0.9 moves in the safe direction and watch for retry amplification. Write the stop condition that ends the incident response for this lab.

Chaos note: inject only one fault class at a time and restore fixtures after. Watch for dual failures — dependency down and retry amplifier — which is how model passes attacker order_id and tool executes becomes an outage. Customer communication templates (even if only for the drill) beat silence. If you queue deferred work, define poison-message handling. Budget documents should state the maximum extra spend allowed during retries. Close the loop by linking the containment action to a dashboard panel for unauthorized_tool_calls == 0 and citation_rate ≥ 0.9.

Draw a state diagram for degrade modes and put it in the repo as ASCII if needed. Cap concurrent retries across the process, not only per request. Ensure cancellation propagates to downstream HTTP clients. When failing closed, choose a user-visible message that does not leak internals. Practice the single command that flips the kill switch or weight to zero. After recovery, drain or inspect deferred work before declaring green. In this chapter the product is support API: versioned policy retrieval + scoped order tool + cited answer or escalate, the human stakeholder is customer asking where order A12 is without leaking other customers' orders, and the incident id you design against is CAP-SUPPORT-TOOL-09. Re-state the oracle in your notes — order lookup returns status+policy cite; cross-customer order_id → deny+escalate — and keep the invariant visible: tool args bind to session customer_id; refunds never auto-executed. Track unauthorized_tool_calls == 0 and citation_rate ≥ 0.9 as the scoreboard. Surface under change control: POST /v1/support/turn. If you only have forty minutes, finish the fixture for model passes attacker order_id and tool executes before polishing UI. Promotion language stays ternary: promote, hold, or roll back based on evidence, not hope.