CCAR-F questions and answers — Domain 5: Context & Reliability

9 practice questions on Context Management & Reliability, the domain weighted at 15% of the CCAR-F exam. Every answer explains why the distractors fail — that is the part that transfers to questions you have not seen before.

Question 1

An agent's answer quality degrades as its conversation grows past 150k tokens. What is the correct architectural response?

1. Move to a model with a larger context window
2. Isolate voluminous work in subagents, clear stale tool results, and offload detail to storage
3. Increase max_tokens on each request
4. Lower the temperature to keep the model focused

Answer: B. Attention dilutes across long contexts; a bigger window lets the problem grow rather than solving it. The architectural moves are isolation (subagents return conclusions, not transcripts), clearing (drop tool results that no longer inform anything), and offloading (files or a memory tool, read back on demand).

Question 2

In clear_tool_uses_20250919, what does clear_at_least do and why does it exist?

1. Sets the minimum number of tool uses kept, to preserve recency
2. Sets a minimum amount cleared per activation, so the cache invalidation the edit causes is worth paying for
3. Delays clearing until a token threshold is reached
4. Excludes specific tools from clearing

Answer: B. Clearing invalidates cached prefixes at the point of the edit. clear_at_least ensures each activation frees enough to justify that cost, instead of repeatedly paying cache-write prices for trivial savings. keep preserves recency, trigger sets the threshold, and exclude_tools protects specific tools.

Question 3

After /compact, which of these is NOT restored?

1. Project CLAUDE.md, re-read from disk
2. The system prompt
3. The skill listing
4. MCP tool availability

Answer: C. Compaction replaces the conversation with a structured summary, then reloads the system prompt, project CLAUDE.md, auto memory and MCP tools, plus the recently modified files and the bodies of skills you actually invoked. The skill listing is the documented exception — it does not reload.

Question 4

A team wants to cut costs on a customer-facing chat agent. Which is appropriate?

1. Route all requests through the Batches API for the 50% discount
2. Prompt caching on the static system prompt and tool definitions
3. Disable tool use to reduce token overhead
4. Truncate every conversation to the last two turns

Answer: B. Prompt caching cuts cost and latency on the repeated prefix — cache reads are 0.1× base. Batches is asynchronous with a window of up to 24 hours, which is disqualifying for an interactive path. Removing tools or aggressively truncating trades away the product to save tokens.

Question 5

Where should a cache_control breakpoint be placed?

1. On the last content block that stays identical across requests
2. On the final user message, which is always the largest
3. On every block, to maximise cache coverage
4. On the first block only, since caching is prefix-based

Answer: A. Caching is prefix-based, so the breakpoint marks how far the identical prefix extends — put it on the last block that does not change. Marking content that changes every request (a timestamp, per-request context) forces a cache write each time. You get four breakpoints, ordered tools → system → messages.

Question 6

Which change invalidates the cached prefix for tools, system AND messages?

1. Adding an image to the latest user message
2. Changing tool_choice
3. Changing the tool definitions
4. Enabling streaming

Answer: C. Invalidation cascades downward from the level that changed. Tool definitions sit at the top of the hierarchy, so changing them invalidates tools, system and messages. Changing tool_choice or adding an image invalidates messages only. Streaming does not affect caching.

Question 7

A tool call returns HTTP 429. What is the correct handling?

1. Fail the request immediately and surface the error to the user
2. Retry with exponential backoff, honouring the retry-after header
3. Retry immediately in a tight loop until it succeeds
4. Switch to a different model and retry

Answer: B. 429 is a rate limit — transient and retryable, with the server telling you how long to wait. Tight-loop retries make the rate limiting worse. Contrast with 400, 401, 403, 404 and 413, which are permanent for that request and should be surfaced rather than retried.

Question 8

You must process 40,000 documents overnight for a compliance audit, with results needed by morning. Which approach?

1. Standard Messages API with high concurrency
2. Message Batches API, matching results on custom_id
3. Streaming requests in a loop
4. Batches API with stream: true for progress visibility

Answer: B. Bulk, delay-tolerant, overnight — the Batches case exactly: 50% off, up to 100,000 requests or 256 MB, most batches finishing within an hour and all within 24. Results can return in any order, so match on custom_id. Streaming is not supported inside a batch.

Question 9

A multi-day refactor spans several Claude Code sessions. What is the reliable system of record for decisions already made?

1. The conversation transcript, resumed with --resume
2. CLAUDE.md and committed artefacts such as a plan file or issue, with sessions resumed for continuity
3. Auto memory alone
4. Re-explaining the context at the start of each session

Answer: B. Session resumption is genuinely useful, but transcripts are eroded by compaction and session boundaries. Durable decisions belong in files that survive both: CLAUDE.md for standing conventions, a committed plan or issue for the state of the work. Auto memory is machine-local and not shared across machines.

← Previous: Questions & answers — Domain 4  |  Next: Facts to memorise →