Question 210
A support session's policy version changed during a long interaction. How should a compact summary handle this?
- Carry forward approval identity, timestamp and supporting policy evidence
- Carry forward a concise referenced policy finding and its source version
- Record policy version and recheck rules that govern pending decisions
- Keep pending approvals, blockers and next required actions explicitly
Explanation
Correct Answer
C:Summaries are snapshots rather than proofs of current policy. Version-aware revalidation avoids applying a superseded refund limit to a later action.
Incorrect Answers
A:High-impact decisions require an audit trail, not just a terse conclusion; this contributes to preserving authorization evidence through summarization without meeting the need for detecting stale policy within condensed state.
B:Cached validated findings reduce repetitive reads while retaining provenance; that serves avoiding repeated retrieval of unchanged evidence, whereas the case requires detecting stale policy within condensed state.
D:Pending workflow state is essential for safe resumption; its benefits for keeping unresolved control gates in condensed context do not resolve the need for detecting stale policy within condensed state.
Question 211
A monorepo search produces hundreds of matches for a feature name. The agent needs to investigate one suspected package without exhausting its context window. Which exploration strategy is most focused?
- Narrow search by package path and inspect only the highest-signal matches before expanding scope.
- Read every matching source file in full before deciding which package owns the feature.
- Trust an old package summary without checking whether the current commit has changed the files.
- Edit all matching packages into a consistent interface before tracing which module is responsible.
- Assign several workers the identical full-repository read and combine their overlapping transcripts.
Explanation
Correct Answer
A:Filtering by likely subtree and reading the few most relevant definitions preserves context for analysis. The agent can expand the search if evidence warrants it, rather than flooding the conversation with unrelated modules.
Incorrect Answers
B:Loading hundreds of files creates token pressure and makes important relationships harder to identify. Start with path constraints and targeted code references, then broaden only where the dependency chain requires it.
C:Unverified stale notes can omit renamed symbols and changed interfaces. A narrow inspection should be grounded in the present working tree rather than treating an earlier summary as fresh evidence.
D:Mutating unrelated packages is not an exploration technique. The task is to localize the actual owner and inspect its call chain, not introduce broad changes while the source of the behavior remains uncertain.
E:Repeating an unbounded read across agents multiplies cost without adding useful independence. Task-specific partitioning and a compact index of relevant files are more reliable ways to manage large repositories.
Question 212
The team needs a check whenever Claude Code finishes a task rather than after each file operation. What hook scope is appropriate?
- Review result payload semantics in a post-tool validation hook
- Focus on attaching the validation to the agent-turn or stop lifecycle event
- Carry out a PostToolUse handler after successful file-edit tools
- Choose a tool lifecycle hook to emit structured audit events
Explanation
Correct Answer
B:End-of-turn hooks can trigger aggregate checks that would be wasteful after every individual edit. A tool-specific callback runs too frequently for the stated completion check.
Incorrect Answers
A:Transport success does not necessarily mean business success; that relates to detecting business failure despite transport success, not the necessary validating at the end of agent activity.
C:A post-tool callback can inspect the completed edit and trigger formatting or verification in response; the deciding condition is validating at the end of agent activity, whereas this action concerns reacting to a completed editing tool.
D:Hooks provide a programmatic observation point around matching tool invocations; the option would serve recording tool activity independently of model narration, not validating at the end of agent activity as specified in the stem.
Question 213
A reusable skill breaks on other developers' machines and goes stale when a dependency changes. Which TWO skill maintenance practices address portability and compatibility? Choose TWO.
- Use portable project-relative resources and document the skill's required environment variables
- Maintain versioned skill instructions and compatibility tests as its dependent API evolves
- Keep the skill at project scope rather than globally for the user
- Formalize required headings, evidence fields and a completion checklist
Explanation
Correct Answers
A:Portable skills avoid machine-specific assumptions and resolve resources from a predictable project context. This reduces brittle failures across teammates and CI.
B:Alongside removing machine-specific assumptions from a skill, the workflow needs maintaining version compatibility of a reusable skill: reusable automation is subject to dependency drift; treating skills like maintained code prevents stale tool assumptions from recurring in every invocation.
Incorrect Answers
C:When addressing limiting a sensitive skill to its repository, scope limits where reusable workflows are discovered and available; the present task instead calls for removing machine-specific assumptions from a skill.
D:A bounded output contract makes a reusable workflow easier to verify; standardizing a command's output shape and removing machine-specific assumptions from a skill require different controls in this situation.
Question 214
A document batch returns records out of order, and failed items must still be traceable to their inputs. Which TWO practices establish correlation at submission and error reconciliation? Choose TWO.
- Persist each batch entry's custom_id together with its returned error or status
- Assume a batch's result order always matches its input submission order
- Identify failed batch entries by their position in the output file instead of by request identifiers
- Assign a unique custom_id at submission and use it to map results back to inputs
Explanation
Correct Answers
A:Alongside correlating out-of-order batch results, the workflow needs linking each batch error to its source record: stable item identifiers connect failures to inputs and retry history; without them, a parallel batch becomes difficult to audit or repair precisely.
D:Batch completion order is not guaranteed. Correlation IDs let the consumer attach each output or failure to the correct source document.
Incorrect Answers
B:Batch outputs can be delivered out of order. Correlate each result with its request using a stable custom_id rather than array position.
C:Result position does not establish the corresponding input after asynchronous processing. The unique custom_id and error classification are needed for robust reconciliation.
Question 215
A service recovers from an outage, but old failed tool results are being reused as current status. What recovery practice helps?
- Reexecute time-sensitive reads after recovery and replace stale failure state
- Classify separately a legitimate not-found result from a tool execution error
- Flag the outcome as uncertain and request reconciliation before replay
- Harmonize structured error codes and retry hints across tools
Explanation
Correct Answer
A:Cached failure artifacts are not live evidence. A targeted fresh read can verify system restoration instead of letting the agent propagate old negative results.
Incorrect Answers
B:A successful search with zero matches is not necessarily an infrastructure failure; its benefits for separating no-match outcomes from failures do not resolve the need for refreshing state after a transient service failure.
C:A timeout does not prove nonexecution for mutating operations; supporting handling an ambiguous post-request timeout cannot replace the required refreshing state after a transient service failure.
D:Consistent machine-readable categories make orchestration decisions repeatable; this contributes to normalizing tool error classifications without meeting the need for refreshing state after a transient service failure.
Question 216
An extractor produces a contract clause that cannot be found on any source page. What should validation do?
- Require a matching source span or flag the clause as unsupported
- Confirm that every cited claim has an inspected supporting source
- Track retrieval time and a source version or approved snapshot reference
- Name confirmed observations separately from estimates and unknowns
Explanation
Correct Answer
A:Evidence-grounded extraction ties fields to document locations. Unsupported text must not be laundered into a legal summary merely because it sounds plausible.
Incorrect Answers
B:For checking citations against actually inspected evidence, a plausible citation string is not evidence of source support; however, this requirement concerns tracing extracted clauses to document spans.
C:Live pages can change or disappear; that may serve preserving time-sensitive research provenance, but it is no substitute for tracing extracted clauses to document spans.
D:Provenance distinguishes system-of-record data from inference; this may support separating verified status from estimates but cannot alone meet the requirement for tracing extracted clauses to document spans.
Question 217
A support tool finds that the requested refund exceeds the agent's approved threshold. What escalation rule is appropriate?
- Send the case to a human approver with amount and policy evidence
- Identify the missing signature and request document correction or human review
- Offer a concise structured handoff with issue, evidence, attempted steps and decision needed
- Hand off to the security or fraud workflow and suspend sensitive actions
Explanation
Correct Answer
A:Threshold exceptions require explicit authority. A model should not override a financial approval limit simply because the customer sounds urgent.
Incorrect Answers
B:Critical missing evidence cannot be filled from a language model's guess; the option would serve escalating missing legally required evidence, not routing above-limit refunds to a human as specified in the stem.
C:Structured escalations let the human act on the actual unresolved choice; this addresses preparing an actionable human escalation, while the question is about routing above-limit refunds to a human.
D:Fraud indicators change the permitted handling of customer accounts; that relates to routing suspected fraud out of standard support, not the necessary routing above-limit refunds to a human.
Question 218
A distributed order workflow has captured card payment, but an inventory reservation subsequently fails permanently. The two external services cannot be rolled back atomically. What orchestration response handles the partial transaction?
- Start a compensating payment reversal and track its confirmed outcome before closing the order.
- Repeat inventory reservations forever and report a completed order while attempting recovery.
- Discard the original payment receipt and replay the whole checkout without an idempotency key.
- Close the order as successful because the payment provider returned a confirmed capture.
Explanation
Correct Answer
A:The transaction has already crossed an irreversible-looking service boundary, so a distributed saga needs a compensating business action. Reversal and explicit reconciliation prevent an inventory failure from leaving an unnoticed customer charge.
Incorrect Answers
B:A permanent stock failure is not solved by endless retries. Marking completion before inventory and payment are reconciled hides the failed prerequisite and leaves the customer charged without fulfillment.
C:Blind replay may capture payment twice because the first transaction already succeeded. Preserving the original operation ID and executing an explicit compensation is safer than restarting the entire workflow.
D:A successful payment is only one branch of the workflow. The missing inventory reservation means the overall business transaction failed and requires compensation or authorized exception handling.
Question 219
A classifier confuses user quotations with instructions to the assistant. What sample pattern would clarify trust boundaries?
- Pick examples representing distinct layouts and difficult ambiguity cases
- Include examples where quoted user content is treated as data, not a control instruction
- Record example-set versions and compare outcomes on a fixed test suite
- Maintain a small representative few-shot set and evaluate the trade-off empirically
Explanation
Correct Answer
B:Adversarial examples can illustrate the boundary between source text and system requirements. The model should classify the quote instead of following its embedded commands.
Incorrect Answers
A:Example diversity matters more than sheer count; this can serve choosing representative examples rather than redundant ones, but leaves the requirement for teaching separation of quoted data and instructions unmet.
C:Versioned few-shot experiments reveal the impact of changes; this concerns measuring few-shot changes with controlled evaluations, leaving the requirement for teaching separation of quoted data and instructions outstanding.
D:Large demonstrations consume context and tokens; controlling context cost from examples and teaching separation of quoted data and instructions are distinct responsibilities under these constraints.
Question 220
A team accidentally committed an API token inside a project MCP configuration file. What change should they make?
- Reference credentials through environment or a secret manager, then rotate the leaked token
- Re-establish or reload the affected server and verify its active configuration
- Review tool exposure, permission filters and server-side authorization
- Select supported authenticated transport and credential renewal rather than embedding a fixed secret
Explanation
Correct Answer
A:Configuration may be shared and reviewed; raw secrets do not belong in it. External secret injection reduces exposure, while rotation addresses the already-compromised credential.
Incorrect Answers
B:Running sessions may retain existing transport state; this contributes to applying modified MCP server settings without meeting the need for removing leaked secrets from MCP settings.
C:Connectivity alone does not grant every capability; that serves diagnosing a connected but unavailable MCP tool, whereas the case requires removing leaked secrets from MCP settings.
D:Because remote MCP services need a secure authentication lifecycle, the measure supports maintaining authenticated MCP transport rather than the needed removing leaked secrets from MCP settings.