Context management in large-codebase exploration
Manage context effectively in large codebase exploration.
This subtopic (5.4) sits in Context Management & Reliability (D5) on Anthropic's Claude Certified Architect — Foundations (CCA-F) exam. The bank holds 15 practice questions here — 3 easy, 9 medium, and 3 hard — with 5 free to try, answers and explanations included. 3 of the free questions are below; the rest are in the practice stream.
What the exam tests here
- context degradation in extended sessions — inconsistent answers, generic references
- structured state persistence — agent exports state, coordinator loads manifest on resume
- scratchpad files for persisting key findings across context boundaries
- /compact to reduce context during extended exploration sessions
Free practice questions: Context management in large-codebase exploration
Midway through a long billing dispute session, your support agent begins giving inconsistent answers about whether process_refund requires manager approval. No code changes were made. What is the most likely cause?
Show answer & explanation
Correct answer: D. Auto-compaction cleared older turns containing the approval policy context
Auto-compaction removes older conversation turns when context approaches the limit, erasing earlier context about approval rules and causing inconsistent answers. 'The MCP server hosting process_refund' unavailability would produce errors, not inconsistency. 'The system prompt was overwritten' requires an explicit action that didn't happen. 'The tool definition changed' would affect all future calls uniformly, not produce mid-session drift.
A customer support agent built with the Claude Agent SDK completes a billing dispute workflow using get_customer and lookup_order. The session ends before the coordinator finishes writing a summary. On resume, the coordinator has no record of which steps completed. What structured state persistence approach prevents this?
Show answer & explanation
Correct answer: C. Store completed step names and retrieved customer IDs in a manifest file the coordinator reads on resume
Exporting completed steps and key IDs to a manifest file gives the coordinator a structured, persistent record it can load on resume — independent of conversation history. 'Include in the system prompt...' is for static instructions, not dynamic session state. 'Re-run get_customer and lookup_order...' wastes calls and may return different results. 'Conversation history alone...' is unavailable after a session ends and may be truncated.
A Claude Code plan mode session generates a multi-step refactoring plan. Midway through execution the session is interrupted. On restart, the developer asks Claude Code to resume, but it has no record of which steps completed. What prevents this on the next run?
Show answer & explanation
Correct answer: B. After each step, write completed step names and file paths to a manifest file the session reads on resume
Writing completed steps to a manifest file gives Claude Code a structured, persistent record to load on resume, surviving a session interruption. 'Store in CLAUDE.md...' is for stable project instructions, not dynamic execution state, and grows polluted with transient data. 'Custom slash command...' could trigger a save action but does not itself persist state. 'Increase context window...' does not preserve session state across a process restart.
2 more free questions on this subtopic in the practice stream, plus 10 in the full bank. Keep practicing →