Context & Reliability — Quick Reference
Context Management & Reliability · 15% of the exam. Density over narrative — this is for review, not first-pass learning.
What counts against the context budget
System prompt + tool definitions + retrieved data + conversation history — all of it, not just the visible chat text.
Prompt caching mechanics
Only the stable prefix before the first change is reused. Append new content after the cached prefix to keep hitting cache.
Compaction trade-off
Summarize to free space, but don't over-compact — the 'summarization trap' drops details the model still needs, producing vague later answers.
Escalate vs proceed
Escalate to a human when the decision is high-stakes AND hard to reverse AND confidence is low. Routine, reversible calls don't need escalation.
Error propagation choice
Match the response to criticality: halt on errors that invalidate the task, retry on transient ones, degrade-and-continue on non-critical ones. Never a single blanket policy.
Idempotency
Required for any tool call that might be automatically retried — guard non-idempotent actions (sending a message, charging a card) explicitly.
Codebase exploration strategy
Targeted grep/search before reading whole files. Reflexively reading an entire large repo wastes context budget the exam expects you to protect.