CCAR-F Practice Questions: Tool Design & MCP Integration Domain

Correct answer (B): Glob efficiently narrows candidate paths, Grep locates symbol occurrences, and Read inspects only the files needed to understand the implementation and callers. The CCAR-F principle is incremental repository exploration before broad content loading or modification. In production, this reduces context use and lowers the risk of changing the wrong code.

Why the other options are wrong:
- Option A: Reading everything first is inefficient and defeats the context-minimization requirement.
- Option C: Writing before locating and understanding the existing implementation risks duplication and breakage.
- Option D: Recursive reading is unnecessarily broad, and editing without inspecting callers can introduce incompatible behavior.

Correct answer (B): Because the known starting point is a content string, Grep should locate candidate files, after which Read can inspect only relevant implementations and dependencies. The CCAR-F principle is incremental codebase exploration using the built-in tool suited to the known information. In production, this limits irrelevant context, reduces latency, and lowers the risk of editing the wrong code.

Why the other options are wrong:
- Option A: Reading broadly before searching for the known symbol is inefficient, and Glob is primarily useful when a path pattern is known.
- Option C: Creating a replacement before locating and understanding the existing implementation risks duplication and incompatible behavior.
- Option D: Reading every file defeats the goal of minimizing irrelevant context before content search.

Correct answer (C): Typed errors and explicit retryability let the orchestrator apply different policies to transient, validation, permission, and business failures. The CCAR-F principle is that MCP failures should combine machine-readable control fields with human-readable, actionable details. In production, this prevents wasteful retries, supports safe correction, and routes unrecoverable conditions appropriately.

Why the other options are wrong:
- Option A: A readable message is useful, but inferring retryability from prose produces less deterministic recovery behavior.
- Option B: Reporting a failed operation as successful content obscures failure semantics and couples callers to backend details.
- Option D: Validation, permission, and business failures normally require correction, escalation, or termination rather than uniform retries.

Correct answer (D): Claude Code can use multiple MCP servers simultaneously, and an MCP resource is suitable for exposing reusable contextual data such as a read-only handbook. The principle is to compose purpose-built servers and resources rather than consolidating them solely because of a false connection limit. In production, this preserves clear ownership while making shared context consistently available.

Why the other options are wrong:
- Option A: Consolidation is unnecessary because Claude Code can use multiple MCP servers simultaneously.
- Option B: Manual prompt injection is repetitive and error-prone and does not provide the requested issue-tracker integration.
- Option C: Separate sessions introduce avoidable state transfer and prevent direct use of both integrations in one project workflow.

Correct answer (B): Renaming the tools by purpose and documenting their contracts directly addresses the selection ambiguity. The relevant CCAR-F principle is that tools need clear, distinguishable responsibilities, including when each tool should and should not be used. In production, explicit boundaries reduce incorrect calls without sacrificing the agent's ability to select the appropriate operation.

Why the other options are wrong:
- Option A: A generic operation parameter moves the ambiguity into a loosely defined argument and weakens the boundaries between operations.
- Option C: Prompt guidance may help temporarily, but it does not repair the ambiguous tool names and contracts causing the problem.
- Option D: An account number may appear in requests concerning orders or case history, so this rule could force the wrong operation.

Correct answer (C): Each agent should receive the minimum capabilities required for its assigned responsibility, with the coordinator routing tasks that require another role. The underlying CCAR-F principle is least-capability tool distribution across specialized agents. In production, this reduces accidental publication, improves tool-selection reliability, preserves role boundaries, and makes privileged actions easier to audit.

Why the other options are wrong:
- Option A: Prompt guidance can reinforce roles, but it does not remove unnecessary privileged capabilities and leaves more tools for the model to select incorrectly.
- Option B: Centralized tool use can control access, but preventing specialists from directly using their necessary tools adds bottlenecks and may discard source detail through summaries.
- Option D: Confirmation parameters can add protection, but agents without publishing duties still retain an unnecessary destructive or externally visible capability.

Correct answer (D): The team-required integration belongs in the repository's project-scoped configuration, while the cross-project personal integration belongs in user scope. Environment-variable expansion supplies deployment-specific values without committing them. The underlying CCAR-F principle is to align MCP configuration scope with sharing requirements. In production, correct scoping improves onboarding, portability, consistency, and secret or environment-value handling.

Why the other options are wrong:
- Option A: Project scope is appropriate for the schema server, but committing personal integrations and developer-specific endpoints creates unwanted team coupling and configuration leakage.
- Option B: User scope works for personal integrations, but storing a team-required server only in personal files prevents the repository from carrying its required shared configuration.
- Option C: This reverses the appropriate scopes: the required repository integration becomes individually managed while the personal integration is imposed on the team.

Correct answer (A): Glob first narrows files by expected path, Grep locates the symbol within those candidates, and Read loads only the relevant content. The underlying CCAR-F principle is incremental codebase exploration using the tool that matches each discovery need. In production, narrowing before reading reduces context consumption, latency, irrelevant data, and the chance of overlooking important matches in excessive output.

Why the other options are wrong:
- Option B: Reading everything could eventually reveal the symbol, but it consumes excessive context and ignores more efficient path and content-search tools.
- Option C: A broad Grep might find the symbol, but searching outside the repository and writing a temporary aggregation introduce unnecessary scope and side effects.
- Option D: Glob is useful for path discovery, but reading all matching files before content search can load thousands of irrelevant files.

Correct answer (C): Search, retrieval, and citation extraction have distinct intents, inputs, outputs, costs, and error conditions, so separate explicit contracts are the clearest design. The underlying principle is to divide tools along meaningful operational boundaries rather than minimizing tool count at all costs. In production, this improves selection, validation, observability, and downstream handling of evidence.

Why the other options are wrong:
- Option A: Prompt documentation could explain combinations, but replacing typed inputs with strings weakens validation and leaves the overloaded interface intact.
- Option B: Inferring intent from fields leaves ambiguous combinations unresolved, while a universal text response makes downstream processing less reliable.
- Option D: Source-specific tools can be useful when sources differ operationally, but this split preserves the ambiguous mixture of search, retrieval, and extraction modes.

Correct answer (A): A successful empty result must mean that the search completed and found no matches, while an authentication failure must be represented as an error. The CCAR-F principle is that MCP results should distinguish valid empty outcomes from failed operations using explicit error signaling and useful details. In production, this prevents false business conclusions and enables appropriate escalation.

Why the other options are wrong:
- Option B: Embedding a status code in an apparent success payload leaves the operation's success state ambiguous.
- Option C: Retries may help transient failures, but an expired token requires authorization recovery and remains indistinguishable from no matches.
- Option D: A placeholder record corrupts the customer data contract by representing an error as customer data.