Question 120
A monorepo's shared instructions are bloated, individual packages need different conventions, and a framework upgrade made some rules obsolete. Which THREE maintenance actions correct these issues? Choose THREE.
- Move package-only conventions into instructions scoped to those package directories
- Revise the outdated framework rules and test them after the upgrade
- Trim the shared CLAUDE.md and link out to details needed only for specific tasks
- Settle the conflicting scope and establish one authoritative test command
Explanation
Correct Answers
A:Selecting package-specific guidance matters in this scenario: directory-local guidance gives specialized expectations without imposing them on unrelated packages; the file's location communicates the intended subtree.
B:This choice addresses maintaining instruction accuracy after upgrades: standing guidance can become stale as dependencies and conventions evolve; keeping it in the code-review process prevents outdated patterns from propagating.
C:High-signal standing instructions are easier to apply and less costly in context. Bulk background should be retrieved as needed rather than always appended to coding sessions.
Incorrect Answers
D:Because overlapping instructions can create inconsistent tool behavior, the measure supports eliminating conflicting project instructions rather than the needed keeping persistent instructions focused.
Question 121
A session is resumed after its source documents were updated. How should the agent avoid making decisions on stale evidence?
- Build a new fork from the last accepted checkpoint for revised analysis
- Capture and compare policy versions before accepting the fork's recommendation
- Revalidate time-sensitive evidence before continuing the resumed workflow
- Order state-changing updates or use optimistic concurrency on the session
Explanation
Correct Answer
C:Session retention preserves past context, not the current truth of external systems. A targeted refresh protects decisions that depend on updated source records.
Incorrect Answers
A:Branching preserves what was previously considered while enabling a corrected trajectory; that may serve revising an analysis without erasing history, but it is no substitute for refreshing stale evidence during resume.
B:A session branch can retain a different configuration snapshot; this may support detecting configuration drift across forks but cannot alone meet the requirement for refreshing stale evidence during resume.
D:Concurrent writes can produce lost updates and inconsistent tool state; a method suited to preventing concurrent session-state conflicts does not meet the need for refreshing stale evidence during resume.
Question 122
A critical security patch has a tested, approved diff ready to apply. Why might a new open-ended plan be unnecessary?
- Apply the reviewed patch within its stated scope and verify the security regression
- Collect performance evidence and define the target metric before code changes
- Keep the task in read-only planning or inspection mode
- State deployment order, checkpoints and rollback triggers before execution
Explanation
Correct Answer
A:Replanning a settled narrow remediation can delay correction and introduce scope drift. The key controls are authorization, targeted execution and regression evidence.
Incorrect Answers
B:Optimization without a baseline makes success impossible to judge and may harm unrelated paths; that relates to baselining performance before optimization, not the necessary avoiding needless replanning of an approved patch.
C:Although an assessment boundary prohibits side effects regardless of how confident the proposed fix appears, the choice addresses honoring a read-only operating constraint and not the requested avoiding needless replanning of an approved patch.
D:Multi-service updates have dependencies and failure domains that cannot be improvised safely; the deciding condition is avoiding needless replanning of an approved patch, whereas this action concerns preparing release sequencing and rollback.
Question 123
A contract extractor copies sample company names into unrelated output records. What redesign reduces example leakage?
- Use varied fictional entities and emphasize that examples illustrate structure, not actual values
- Keep a small representative few-shot set and evaluate the trade-off empirically
- Prioritize making examples conform to a single documented schema and labeling convention
- Display labeled cases where multiple classes are present and document the priority rule
Explanation
Correct Answer
A:Reusing the same names across all demonstrations can invite copying. Distinct examples and explicit source-grounding criteria separate format teaching from data extraction.
Incorrect Answers
B:Large demonstrations consume context and tokens; in this scenario, controlling context cost from examples is secondary to preventing sample values from leaking into outputs.
C:When considering eliminating inconsistent example output formats, contradictory examples undermine explicit instructions; the requirement for preventing sample values from leaking into outputs remains unmet.
D:When addressing teaching category precedence through examples, mixed-category examples teach how conflicts resolve; the present task instead calls for preventing sample values from leaking into outputs.
Question 124
A help-desk agent obtains a successful tool_result containing eligibility and account status. What is the next orchestration step?
- Send the result back into the model loop so Claude can decide the next action
- Implement a step budget and a defined fallback when the loop exceeds it
- Pair bounded retries with an operation idempotency key and outcome lookup
- Produce an explicit failed tool result with diagnostic information rather than fabricate data
- Confirm the side effect with an authoritative follow-up read before promising completion
Explanation
Correct Answer
A:The application does not simply display raw tool output as the final assistant response. Reintroducing observed results lets the model interpret evidence and either answer or request another step.
Incorrect Answers
B:Independent limits on turns, time, and side effects prevent runaway agent cycles; bounding an autonomous agent loop and resuming model reasoning after a result are distinct responsibilities under these constraints.
C:Retries of side-effecting actions can duplicate work when acknowledgment is lost; the action supports safe recovery after an ambiguous tool timeout but falls short of resuming model reasoning after a result in this case.
D:The tool-error signal should flow into the conversation so the agent can retry, escalate, or communicate limits; this can serve surfacing client-tool execution failure, but leaves the requirement for resuming model reasoning after a result unmet.
E:A tool's acknowledgement can be stale or incomplete relative to durable state; that addresses verifying externally visible side effects, whereas this case calls for resuming model reasoning after a result.
Question 125
A PR reviewer should read diffs but must not push commits from the CI job. Which permission strategy fits?
- Demand file path, line range and finding rationale in a validated result schema
- Restrict allowed tools and repository credentials to read-only review operations
- Limit retrieval and review to the changed files and their affected dependencies
- Check the review payload and apply a documented failure policy
Explanation
Correct Answer
B:CI runners often have elevated tokens. A read-only tool surface and minimal permissions prevent a review agent from making unrequested source changes.
Incorrect Answers
A:Location-aware structured findings allow a bot to attach actionable comments; that serves generating location-specific PR findings, whereas the case requires enforcing read-only CI review permissions.
C:A diff-aware prompt and bounded context reduce cost while preserving relevant evidence; its benefits for constraining CI review to changed scope do not resolve the need for enforcing read-only CI review permissions.
D:CI must distinguish tooling errors from confirmed code defects; supporting handling invalid AI-review output in a pipeline cannot replace the required enforcing read-only CI review permissions.
Question 126
An MCP server gets deployed with a new name and existing agent tools stop resolving. What migration discipline helps?
- Load credentials through environment or a secret manager, then rotate the leaked token
- Begin the MCP server as a local stdio process
- Coordinate server identifier and tool-name changes with client configuration
- Employ supported authenticated transport and credential renewal rather than embedding a fixed secret
Explanation
Correct Answer
C:Clients refer to discovered capabilities using identifiers. Uncoordinated renames can break allowlists, prompts and tool routing even if backend functionality remains intact.
Incorrect Answers
A:Configuration may be shared and reviewed; raw secrets do not belong in it; the stated problem requires migrating MCP tool identifiers safely rather than removing leaked secrets from MCP settings.
B:Stdio transport connects the client to a child process using local streams; a method suited to identifying local stdio MCP transport does not meet the need for migrating MCP tool identifiers safely.
D:Remote MCP services need a secure authentication lifecycle; that may serve maintaining authenticated MCP transport, but it is no substitute for migrating MCP tool identifiers safely.
Question 127
An incident investigation needs different security and operations specialties while keeping customer data compartmentalized. Which orchestration decision is appropriate?
- Introduce a targeted citation repair pass before further broad delegation
- Direct cross-specialty requests through an orchestrator with an explicit task contract
- Pass each worker only the minimum task-specific context and authorized tools
- Divide independent expert analyses and reserve integration for the orchestrator
Explanation
Correct Answer
C:Specialization should come with least-necessary information and access. Broadcasting the entire customer history across all workers increases exposure without improving their assigned analyses.
Incorrect Answers
A:Although missing source attribution is a bounded post-processing defect, the choice addresses targeted repair instead of repeated broad delegation and not the requested compartmentalizing delegated context and privileges.
B:Hub-and-spoke mediation controls ownership, context, and failure propagation; the deciding condition is compartmentalizing delegated context and privileges, whereas this action concerns minimizing unpredictable peer-to-peer dependencies.
D:Focused workers can receive targeted instructions and smaller context windows; this addresses matching specialist agents to distinct competencies, while the question is about compartmentalizing delegated context and privileges.
Question 128
A first code patch passes one unit test but fails the integration suite. What is the correct next step?
- Validate behavioral acceptance cases rather than treating lint success as completion
- Use failing integration output to refine the patch and rerun relevant tests
- Inspect and checkpoint cohesive changes with their test evidence
- Launch contract or compatibility tests for existing consumers
Explanation
Correct Answer
B:Iterative development uses tool feedback as evidence. A local unit test cannot override a failing broader acceptance condition.
Incorrect Answers
A:When considering distinguishing static hygiene from functional success, linting checks style and some static issues; it does not prove runtime correctness; the requirement for refining code based on integration-test feedback remains unmet.
C:Small validated increments provide rollback boundaries and clearer blame history; checkpointing incremental implementation progress and refining code based on integration-test feedback require different controls in this situation.
D:Code that compiles can still break callers relying on prior inputs or outputs; the unsatisfied requirement is refining code based on integration-test feedback, not testing API compatibility after code generation.
Question 129
A research agent cannot fit all full-text sources into the current context window. Which retrieval strategy is preferable?
- Retrieve relevant passages on demand while retaining stable citations
- Log policy version and recheck rules that govern pending decisions
- Keep claim-to-source mappings alongside the condensed findings
- Preserve a concise referenced policy finding and its source version
Explanation
Correct Answer
A:Selective retrieval focuses tokens on the current question and leaves large source archives outside the prompt. Source handles permit later verification without full duplication.
Incorrect Answers
B:Summaries are snapshots rather than proofs of current policy; this can serve detecting stale policy within condensed state, but leaves the requirement for using targeted retrieval to manage context limits unmet.
C:Compression should preserve provenance links for verifiable conclusions; that addresses maintaining citations during context compression, whereas this case calls for using targeted retrieval to manage context limits.
D:Cached validated findings reduce repetitive reads while retaining provenance; this concerns avoiding repeated retrieval of unchanged evidence, leaving the requirement for using targeted retrieval to manage context limits outstanding.
Question 130
A code-audit skill requires helper scripts and reference examples that would distract every normal session. How should it be packaged?
- Split dependency installation behind explicit approval and documented prerequisites
- Specify required headings, evidence fields and a completion checklist
- Store supporting resources within the skill and load them when needed
- Explain the skill as plan analysis and deny deployment or mutation tools
Explanation
Correct Answer
C:Skills can bundle reference material and executable helpers for a focused task. On-demand loading keeps general conversation context lighter.
Incorrect Answers
A:Skills are reusable instructions, not a blanket permission grant for arbitrary software installation; its benefits for requiring approval for skill-driven dependency changes do not resolve the need for packaging on-demand skill resources.
B:A bounded output contract makes a reusable workflow easier to verify; supporting standardizing a command's output shape cannot replace the required packaging on-demand skill resources.
D:Skill language and access control should agree; that serves separating review skills from deployment skills, whereas the case requires packaging on-demand skill resources.