Practice — Domain 5: Context Management & Reliability (15%)
18 scenario-based MCQs. Answer key + explanations at the bottom.
Q1
The "CALM" framework for context management stands for:
A. Cache, Adapt, Lower, Maximize B. Cache, Align, Limit, Monitor C. Compose, Allocate, Loop, Merge D. Cancel, Audit, Log, Measure
Q2
A team caches a 50KB reference doc with cache_control: {type: "ephemeral"} but their cache hit rate is near zero. Every request has slightly different metadata (a timestamp + a per-request user_id) inserted in the system prompt before the cached doc. Most likely cause?
A. The TTL is too short B. The volatile content sits before the cache breakpoint, invalidating the cached prefix C. Caching doesn't apply to PDFs D. They need a larger model
Q3
For a long-running customer support workflow (cart → payment → confirmation), the best context-management strategy is:
A. Replay full conversation history every turn B. Maintain a compact JSON state object capturing canonical state and prune older turns C. Sliding window of last 3 turns only D. Retrieve relevant past turns from a vector store every turn
Q4
A non-idempotent tool (charge_card) timed out. The agent doesn't know whether the charge processed. The correct next step is:
A. Retry the charge with exponential backoff
B. Call a read-only check_payment_status to disambiguate before any retry
C. Tell the user the request failed
D. Cancel and start over
Q5
Cache reads cost approximately:
A. Same as regular input B. ~10% of regular input C. ~50% of regular input D. ~125% of regular input
Q6
Cache writes cost approximately:
A. ~10% of regular input B. ~50% of regular input C. ~125% of regular input D. Free
Q7
The default ephemeral cache TTL is:
A. 30 seconds B. 5 minutes (with a 1-hour tier available where supported) C. 24 hours D. Indefinite until invalidated
Q8
How many cache breakpoints can a single request use?
A. 1 B. Up to 4 C. Unlimited D. 8
Q9
A 200-turn conversation is hitting context-window limits. The simplest reliable mitigation that preserves the most user-relevant state is:
A. Switch to a smaller model B. Progressive summarization — replace oldest turns with a concise summary, keep recent verbatim C. Drop the system prompt D. Restart the conversation
Q10
When passing thinking blocks back across turns (for tool use continuation), you must:
A. Strip them — they're internal
B. Include them unchanged, including their signature field
C. Convert to plain text
D. Cache them separately
Q11
A tool_result with is_error: true and type: "permanent" (e.g., 404 not found) should trigger which agent behavior?
A. Retry with backoff B. Adapt (try a different query, ask user, or escalate) — don't retry the same call C. Crash the session D. Re-prompt the system
Q12
Which of the following is the strongest reason to not put live timestamps in a cached system prompt?
A. They take too many tokens B. They invalidate the cache on every request, causing constant cache misses C. They confuse the model D. The API rejects them
Q13
You're agentically retrieving an order status. The user's last 20 turns include three different fetched balances. What's a reliability hazard?
A. The model might quote a stale balance from an earlier turn B. The conversation costs too much C. The model will refuse to answer D. Tool use will be disabled
Q14
"Refresh-on-need" is preferable to "remember-forever" for:
A. Persona / style instructions B. Volatile facts like balances, statuses, prices, inventory C. Tool descriptions D. System prompt text
Q15
The cache breakpoint should be placed:
A. After all volatile content B. Right before any volatile content, with stable content above C. In the middle of a conversation, randomly D. Inside each user turn
Q16
You see cache_creation_input_tokens consistently high request after request. Most likely:
A. Caching is working perfectly B. Your cacheable prefix is changing each request — the cache is being rewritten instead of read C. The model isn't supported D. Cache tier is wrong
Q17
Escalating to a human is appropriate when:
A. Confidence is low, required tool unavailable, iteration cap hit, or policy rules trigger (e.g., refund > threshold) B. The user has a follow-up question C. The first tool call succeeds D. The cache TTL expires
Q18
You're monitoring a Claude-powered agent in production. Which signal is most valuable for catching context-management regressions?
A. Number of API calls B. Cache hit rate (cache_read vs cache_creation tokens) and per-conversation token budget C. CPU utilization D. Wall-clock latency only
Answers
Q1: B. Cache stable prefixes, Align prefix structure across requests, Limit history, Monitor cache hits & tokens.
Q2: B. Volatile content (timestamps, user_ids) before the cacheable prefix breaks alignment and invalidates the cache. Fix: put volatile data after the breakpoint or remove it from the prefix entirely.
Q3: B. State-object pattern is the canonical compaction strategy for transactional / workflow agents. (A) inflates cost; (C) loses context; (D) is overkill for short transactional flows.
Q4: B. Uncertain-state pattern: don't retry a non-idempotent op without first reading state to disambiguate. Retrying (A) risks double-charging.
Q5: B. Cache reads are roughly 10% of regular input cost — the savings lever.
Q6: C. Cache writes cost about 125% of regular input — there's a one-time premium, which is why caching pays off after ~2 reuses.
Q7: B. Default ephemeral TTL is 5 minutes; a 1-hour tier exists where supported. Every access bumps the TTL.
Q8: B. Up to 4 cache breakpoints per request.
Q9: B. Progressive summarization — replace oldest verbose turns with a compact summary, keep recent turns. This is the canonical long-conversation strategy.
Q10: B. Thinking blocks must be passed back unchanged, including signature, for the model to maintain reasoning continuity across tool use turns.
Q11: B. Permanent errors mean adapt, don't retry. (A) is wasteful and may hammer a definitely-failing call.
Q12: B. Timestamps in a cached prefix invalidate the cache every request — you pay write cost forever and never get read economics.
Q13: A. Volatile facts (balances) fetched far back may have changed. Refresh-on-need is the discipline, not stash-and-reuse.
Q14: B. Volatile facts go in the refresh-on-need bucket. Persona, tool descriptions, and system prompts are stable and should be cached.
Q15: B. Stable content goes above the breakpoint, volatile content below. This maximizes cache reuse.
Q16: B. High cache_creation request after request means you're rewriting the cache constantly — likely because the prefix is changing. Investigate prefix stability.
Q17: A. Escalation triggers are: low confidence, missing tool, iteration cap hit, policy triggers. (B) is a normal continuation, not escalation.
Q18: B. Cache-hit ratio and token budgets are the leading indicators of context-management health. Latency (D) is a downstream symptom; (C) is unrelated.