Foundations Mock Exam 3 — 60 Scenario Questions, Timed (Q649–Q708)

A full-length CCAR-F practice simulation, built like the official exam: 60 questions in 120 minutes, under 4 scenarios drawn from the bank of 6, 15 questions per scenario. This mock uses Code Generation with Claude Code · Multi-Agent Research System · Developer Productivity with Claude · Claude Code for Continuous Integration. Every question was written for this guide against a task statement of the official Claude Certified Architect – Foundations Exam Guide (v1.0, July 2026), and the domain counts follow its blueprint. 4 questions are multiple-response and say “Select TWO”; the rest say “Select ONE”. This is not an official Anthropic or Pearson VUE exam. It does not reproduce, reconstruct or recall any live exam item, and the guide’s sample questions are used only as a model of style. Other full mocks: Mock 1 · Mock 2 · Mock 4.

How to read your result. The runner reports a practice score: raw questions correct out of 60, a per-domain breakdown, the elapsed time and the unanswered count. Unanswered questions count as incorrect. The live exam reports a scaled score (pass mark 720 on 100–1,000); this page does not convert a raw score into a scaled score, because no public conversion exists. A useful study heuristic: aim for 80% here, with no domain under 65%, on a mock you have not seen before. Untimed practice by domain lives in Practice Questions and Concept Drills.

Foundations mock exam 3 · 60 questions (Q649–Q708)

Scenario 1 · Code Generation with Claude Code

D2 · Claude Code

Q649. SQL migration files live in apps/api/db/migrations/, packages/ledger-core/migrations/ and several jobs/*/migrations/ folders, and new jobs add new folders every month. Every migration must include a reversible down step and must never drop a column in the same release that stops writing it. Which way of giving Claude Code this convention is most maintainable? (Select ONE.) A) A CLAUDE.md inside each existing migrations folder that repeats the convention, plus an onboarding note asking job authors to copy it into new folders. B) A .claude/rules/ file with paths frontmatter such as /migrations//*.sql, so it loads for migrations in any folder. C) A project skill named migration-rules whose description mentions SQL migrations, so Claude can load the convention when it decides the task involves one. D) A section in the root CLAUDE.md that states the convention and asks Claude to apply it when the file being edited appears to be a migration.

D2 · Claude Code

Q650. Tallyforge must replace its hand-written date helpers with the Temporal API across about 60 files in apps/web and apps/api. Two valid approaches exist: wrap Temporal in an adapter module first, or replace call sites directly, and the choice changes how much test churn the team absorbs. How should the engineer run this work in Claude Code? (Select ONE.) A) Start in plan mode to map every helper usage and compare both approaches, approve one plan, then let Claude execute the approved plan with normal edits. B) Use direct execution with a long first prompt that specifies the adapter approach and exact edits for every file, so no exploration step is needed. C) Start in direct execution and switch to plan mode only if Claude reports unexpected dependencies between helpers partway through the edits. D) Use direct execution file by file with a commit after each file, so any bad edit can be reverted and the best approach becomes clear as work proceeds.

D5 · Context & Reliability

Q651. An engineer spends a three-hour Claude Code session mapping the legacy jobs/ folder before a rewrite. Early on, Claude identified specific classes such as SettlementWindow and FxRateCache. By hour three, its answers describe "typical ETL patterns" and contradict findings from the first hour. What should the engineer do to keep later answers grounded? (Select ONE.) A) Switch to a model with a larger context window so the entire three-hour exploration fits without any earlier findings falling away. B) Restart the session every hour and let Claude rediscover the folder from scratch, so each answer is based on fresh reads of the code. C) Have Claude record key findings in a scratchpad file such as docs/jobs-map.md and consult it for later questions. D) Remind Claude in each new prompt to remember carefully what it found earlier and to avoid giving generic answers about common patterns.

D5 · Context & Reliability

Q652. An engineer types: "make the retry policy back off longer." Claude Code's search finds three retry policies, in the HTTP client, the queue consumer and the ledger sync job. It edits the most recently changed one, which was not the one the engineer meant, and the mistake reaches review. What should the team add so this does not repeat? (Select ONE.) A) A CLAUDE.md instruction telling Claude to apply the requested change to every matching candidate, so the intended one is always included. B) A CLAUDE.md instruction telling Claude to choose the candidate with the most references in the codebase, since it is most likely the shared one. C) A CLAUDE.md instruction, with a short example, telling Claude to list the matching candidates and ask which one is meant before editing when a request matches several. D) A PreToolUse hook that blocks every Edit until the engineer approves it manually, so wrong edits are caught before they reach review.

D2 · Claude Code

Q653. Tallyforge split its root CLAUDE.md into .claude/rules/terraform.md, storybook.md and pytest.md. Terraform modules sit under infra/ and several services/*/deploy/ folders. A week later, /context shows all three rules files loaded even in sessions that only touch Python jobs, using about 9,000 tokens. What change makes each file load only when it is relevant? (Select ONE.) A) Convert each rules file into a project skill with a clear description, so Claude pulls in a convention only when it judges that the current task needs it. B) Shorten each rules file to a few summary lines so the three files together cost far fewer tokens in every session, while still loading all of them. C) Add paths frontmatter with glob patterns to each rules file, such as **/*.tf for terraform.md, so each file loads only when matching files are worked on. D) Import the three rules files from the root CLAUDE.md with @import lines instead of keeping them in .claude/rules/, so they are all referenced from one place.

---
paths:
  - "**/*.tf"
  - "**/*.tfvars"
---
# Terraform conventions
Pin every provider version. Never hard-code account identifiers.

D2 · Claude Code

Q654. After a painful refactor, a Tallyforge team lead proposes a rule that every Claude Code task must start in plan mode. Another engineer argues that well-scoped changes waste time in a planning step. The team reviews four tickets from this sprint to decide which ones should go straight to direct execution. Which TWO tickets are best handled with direct execution? (Select TWO.) A) Decide whether ledger events should move to an event-sourced store or stay in the current tables, and then restructure the ledger-core modules to match. B) Fix the TypeError in formatInvoiceDate, where the stack trace points to one line in one file and the failing unit test already reproduces it. C) Add a check in createTransfer that rejects amounts below one cent, matching an existing check pattern used in the same function for maximum amounts. D) Move session handling in apps/web and apps/api from cookies to short-lived tokens, touching auth middleware, clients and about 40 test files.

D5 · Context & Reliability

Q655. Two hours into exploring packages/ui, an engineer's Claude Code session is about 85% full, mostly with long file listings and search output. The session also holds decisions the engineer needs for the next step: which components to split and the agreed prop names. The next step is implementation in the same session. What should the engineer do now? (Select ONE.) A) Run /clear to empty the context completely, then start implementation in the same session with whatever decisions the engineer can retype from memory. B) Continue straight into implementation and keep working until the context is completely full, then decide at that point what to do about the overflow. C) Start a brand-new session for implementation with no summary, so Claude reads the components again and makes its own fresh decisions. D) Run /compact with an instruction to keep the split decisions and agreed prop names, so verbose discovery output is condensed and the decisions remain.

D2 · Claude Code

Q656. Engineers run the team's dependency-audit skill in the middle of feature work. It reads every lockfile and prints about 3,000 lines of package analysis into the conversation. After it runs, Claude's later coding answers in the same session lose track of earlier decisions. Engineers only need the final list of risky packages. Which frontmatter change should the team make to the skill? (Select ONE.) A) Keep the skill inline and add a final step in its body that tells Claude to run /compact automatically after it prints the full analysis. B) Add context: fork to the SKILL.md frontmatter so the audit runs in a separate subagent context and only its summary returns to the main session. C) Add an argument-hint to the SKILL.md frontmatter that asks for a package name, so engineers audit one dependency at a time instead of all. D) Add allowed-tools listing only Read and Grep to the SKILL.md frontmatter so the audit cannot run noisy shell commands that produce long, verbose output.

---
name: dependency-audit
description: Audit lockfiles and list risky packages.
context: fork
---
Read every lockfile, analyse each package, and return only a table of risky packages with reasons.

D5 · Context & Reliability

Q657. During a long debugging session, Claude Code runs the full Vitest suite after every edit. Each run prints about 4,000 lines, almost all of them passing tests, and after six runs the session is close to its context limit while only two failing tests matter. What change best preserves the session's useful context? (Select ONE.) A) Let the session continue unchanged and rely on automatic compaction to summarize the older test runs once the context window finally fills up completely. B) Run /clear after every test run so the next attempt starts with an empty context and no leftover output from earlier runs. C) Switch to a model with a much larger context window so that many more complete test runs fit in the session before it finally reaches its limit. D) Have Claude run the suite with a reporter or filter that prints only failing tests and their errors, so each run adds a few dozen relevant lines.

D5 · Context & Reliability

Q658. Before a release, an engineer pastes one combined report from twelve checks (lint, type-check, unit, contract and security scans), about 30,000 tokens, and asks Claude Code to fix everything. Claude fixes the lint issues at the top and the security findings at the bottom but misses most contract-test failures from the middle. How should the engineer restructure the input? (Select ONE.) A) Paste the same report again with a sentence asking Claude to read every part of it carefully and fix all failures without skipping any. B) Sort all findings alphabetically by file path so related failures sit next to each other regardless of which check produced them. C) Put a short summary listing every failing check and count at the top, then give each check's details under an explicit section header. D) Switch to a model with a larger context window so the 30,000-token report sits comfortably inside the available space with room left over.

D2 · Claude Code

Q659. Claude Code's refactor of reconcileDay() has three problems. It rounds per line instead of per invoice, which changes totals. The duplicate-detection key is built from those totals. And the retry wrapper now re-runs the step after a duplicate is detected. Fixing any one changes how the others behave. How should the engineer give this feedback to Claude? (Select ONE.) A) Ask Claude to revert the whole refactor and redo it from the beginning, this time with a reminder to be careful about totals and retries. B) Report the rounding problem first, wait for the fix, then report the duplicate key, then the retry wrapper, so each change can be checked alone. C) Describe all three problems in one detailed message, explain how they interact, and ask for one coordinated fix with tests. D) Report only the retry wrapper problem, since it is the most visible symptom, and expect the other two to resolve once retries behave correctly.

D2 · Claude Code

Q660. Tallyforge shares a release-notes skill in .claude/skills/release-notes/. One staff engineer wants a much terser personal version for their own changelog drafts. The team version must keep working unchanged for everyone else, and the engineer still wants to run the team version on some days. Where should the engineer put the personal variant? (Select ONE.) A) In the project CLAUDE.md, adding a line that tells Claude to write release notes tersely whenever this engineer is the one asking for them. B) In ~/.claude/skills/release-notes-terse/SKILL.md, a personal skill with a different name, so both versions stay available and teammates see no change. C) In ~/.claude/skills/release-notes/SKILL.md, using the same name, so the personal copy replaces the team version whenever this engineer runs the skill. D) In the shared .claude/skills/release-notes/SKILL.md on a local branch that is never pushed, so the change stays on the engineer's own machine only.

D5 · Context & Reliability

Q661. In phase one, a Claude Code session mapped how apps/api calls the deprecated ledger v1 client. For phase two, the engineer spawns three subagents, one per service, to plan the migration. Each subagent spends its first twenty minutes rediscovering the same call graph, and two reach different conclusions about a shared wrapper. What should change before phase two starts? (Select ONE.) A) Configure the subagents to inherit the main session's full conversation history automatically, so each one sees all of phase one exactly as it happened. B) Give each subagent a larger turn limit so it has enough time to rediscover the call graph and still finish its migration plan. C) Write a concise summary of the phase-one findings, including the call graph and the wrapper decision, and include it in each subagent's initial prompt. D) Run the three subagents one after another instead of in parallel, so each can read the files the previous subagent already opened.

D2 · Claude Code

Q662. The ledger-core maintainers report that Claude Code keeps applying the UI team's Storybook and CSS rules while they work on money arithmetic, because every convention sits in the 900-line root CLAUDE.md. They want each package to load only the shared standards files that its own maintainers choose, such as docs/standards/money.md. Which change best meets this goal? (Select ONE.) A) Turn each standards file into a project skill in .claude/skills/ whose description names the package, so Claude loads it whenever the work seems related. B) Keep the single root CLAUDE.md but add a heading for each package and a line asking Claude to apply only the section that matches the current package. C) Give each package its own CLAUDE.md with @import lines for only the standards its maintainers select, and trim the root to universal rules. D) Move every section of the root file into separate files under .claude/rules/ without paths frontmatter, so each topic sits in its own file and is easier to edit.

# packages/ledger-core/CLAUDE.md
Ledger core: pure arithmetic, no I/O.

@../../docs/standards/money.md
@../../docs/standards/errors.md

D2 · Claude Code

Q663. An engineer asks Claude Code to generate a mapper from legacy CSV headers to camelCase DTO fields. The prose instructions describe the naming rules in detail, yet results stay inconsistent: acct_no_2 became acctNo2 in one run and accountNumber2 in another, and IBAN_CD became ibanCd once and iBANCode once. What is the most effective next step? (Select ONE.) A) Give Claude two or three concrete header-to-field examples, including acct_no_2 and IBAN_CD with the exact expected names, and ask it to regenerate the mapper. B) Ask Claude to interview the engineer about naming preferences before it writes any code, so hidden considerations come out before the next attempt. C) Rewrite the prose instructions with more precise wording about abbreviations, acronyms and trailing digits, and add a sentence asking Claude to apply them strictly. D) Switch the session to plan mode so Claude must describe its full naming approach first and wait for approval before generating the mapper again.

Scenario 2 · Multi-Agent Research System

D4 · Tools & MCP

Q664. The doc_analysis subagent has two tools: lookup_document, which searches the PDFs a client uploaded for this engagement, and search_archive, which searches Marrowfield's past published reports. Both descriptions read "Searches documents for relevant information." Logs show the subagent often queries the archive when the analyst asked about the client's own studies. What is the most effective first fix? (Select ONE.) A) Rewrite both descriptions to state each tool's source, input format, example queries and when to use it instead of the other tool. B) Add a keyword router in front of the subagent that sends any request mentioning the client's name to lookup_document before the model chooses. C) Add a system prompt line telling the subagent to prefer lookup_document in general and use search_archive only when nothing else works. D) Merge both tools into one search tool that queries the client PDFs and the archive together and returns all the matches in one list.

D5 · Context & Reliability

Q665. A client uploaded 11 PDFs. Three are scanned images with no text layer, so doc_analysis cannot read them. It currently returns status success with findings from the other eight and no mention of the three, and the final report states that "client documents show no data on retailer margins," a topic covered in one of the scans. What should doc_analysis return instead? (Select ONE.) A) The eight documents' findings plus, for each unreadable PDF, the failure type, what was attempted, and possible alternatives such as OCR, so the coordinator can decide. B) The same success status and findings, but after first retrying each scanned PDF three more times in case the text becomes readable. C) The eight documents' findings plus an empty findings list marked success for each scanned PDF, so every document has an entry in the output. D) A failure status for the whole subagent call with no findings, so the coordinator knows something went wrong and can rerun the analysis from the start.

D1 · Agentic Architecture

Q666. An analyst finished a 40-minute session in which the coordinator analyzed 120 client documents and web sources for a pet-food market study. The client now wants two alternative report framings, one positioning its new range as premium and one as a value line, and the analyst wants to compare them side by side without redoing the analysis. Which approach fits best? (Select ONE.) A) Resume the analysis session, generate the premium framing, then continue in the same session to generate the value framing from exactly where the first one stopped. B) Start two new sessions, one per framing, and have each one repeat the full document and web analysis before writing its framing. C) Resume the analysis session once and ask the coordinator to produce both framings together in one response so they share every assumption. D) Fork the analysis session twice with fork_session, giving one branch the premium framing and the other the value framing, while the original stays unchanged.

D4 · Tools & MCP

Q667. Each client engagement has hundreds of uploaded documents in nested folders. Traces show the coordinator spends six to eight tool calls per run on list_folder and get_file_info just to learn which documents exist before it can plan any delegation. The document MCP server is maintained by Marrowfield. What change best reduces these exploratory calls? (Select ONE.) A) Add a list_all_documents tool that returns the full catalog in one call, and instruct the coordinator to call it at the start of every run. B) Cache the results of list_folder and get_file_info on the server, so repeated exploratory calls during a run return faster than before. C) Expose the engagement's document catalog as an MCP resource (titles, dates, topics) that the host gives the coordinator at the start of each run. D) Paste a copy of each client's document list into the coordinator's system prompt by hand once a week, so the coordinator never needs to explore.

D1 · Agentic Architecture

Q668. Marrowfield's coordinator loop ends a run as soon as a response contains any text block. Logs show many runs stopping right after the coordinator writes "Next I will send these PDFs to doc_analysis" alongside a tool_use block, so no analysis ever happens and the report is built from search results alone. What change fixes the loop's control flow? (Select ONE.) A) Continue the loop while stop_reason is "tool_use", running the requested tools and appending their results, and end it only when stop_reason is "end_turn". B) Add a system prompt rule telling the coordinator never to write any text in a response that also requests a tool, so the text check works. C) Keep the text check but also cap the loop at 25 iterations, so runs that never produce a final text block are still stopped at a safe point. D) End the loop only when the coordinator's text contains the phrase FINAL REPORT, and instruct the coordinator to write that exact phrase once it is completely done.

while True:
    response = client.messages.create(model=MODEL, max_tokens=4096, tools=TOOLS, messages=messages)
    messages.append({"role": "assistant", "content": response.content})
    if response.stop_reason != "tool_use":
        break
    results = [run_tool(block) for block in response.content if block.type == "tool_use"]
    messages.append({"role": "user", "content": results})

D1 · Agentic Architecture

Q669. Reports often omit findings from client PDFs even though doc_analysis returned them. The coordinator's delegation to synthesis says only "Synthesize the findings gathered so far into a briefing," and the synthesis subagent's output cites web sources alone. Logs show doc_analysis returned 14 findings with page numbers. What should change in how synthesis is invoked? (Select ONE.) A) Give the synthesis subagent the load_document tool so it can re-read the client PDFs itself whenever it needs more details from the uploaded documents. B) Include the full web_search and doc_analysis findings in the synthesis prompt as structured data, with each finding's source URL or document page. C) Raise the synthesis subagent's turn limit so that it has enough time to locate and gather the findings gathered so far on its own. D) Configure the synthesis subagent's definition so that it inherits the coordinator's conversation history and can read the doc_analysis results from there.

{"question": "Demand for refillable cleaning products in Germany", "web_findings": [{"claim": "Refill purchases grew in the first quarter", "source_url": "https://example.org/report", "published": "2024-03"}], "document_findings": [{"claim": "Refill shelf space rose at 40% of audited stores", "document": "client-retail-audit.pdf", "page": 12}]}

D4 · Tools & MCP

Q670. Inside doc_analysis, your application makes direct Messages API calls with three tools: extract_document_metadata, extract_findings and compare_with_sources. Metadata such as publication date must be extracted before anything else, but in 15% of documents the model goes straight to extract_findings and the dates are never captured. How should the first request be configured? (Select ONE.) A) Keep tool_choice on auto and add a system prompt line telling the model that metadata extraction must always come before every other tool. B) Set tool_choice to "any" on the first request, so the model must call one of the three tools instead of replying with plain text. C) Keep tool_choice on auto and add a PostToolUse step that fills in missing dates afterwards by searching each document's text for a year. D) Set tool_choice to {"type": "tool", "name": "extract_document_metadata"} on the first request, then use auto on the follow-up requests for the other tools.

{"model": "MODEL_ID", "max_tokens": 1024, "tools": [{"name": "extract_document_metadata", "description": "Extract title, author and publication date.", "input_schema": {"type": "object", "properties": {"title": {"type": "string"}, "published": {"type": ["string", "null"]}}, "required": ["title", "published"]}}], "tool_choice": {"type": "tool", "name": "extract_document_metadata"}, "messages": [{"role": "user", "content": "DOCUMENT TEXT"}]}

D5 · Context & Reliability

Q671. For one briefing, web_search found a 2022 trade association study reporting that refill-station sales fell 8% after a deposit scheme began, and doc_analysis found a 2024 retailer panel survey reporting a 3% rise. Synthesis kept only the 2024 figure and cited it as the effect of the scheme. How should synthesis handle these two findings? (Select ONE.) A) Report both figures with their sources, methods and collection dates, and note that the difference may reflect timing rather than a direct contradiction. B) Average the two figures into a single net estimate of about a 2.5% decline, and cite both sources together for that combined number. C) Keep only the 2022 trade association study, since an industry-wide body is more authoritative than one retailer panel, and drop the retailer panel survey. D) Keep only the more recent 2024 figure, since newer data usually reflects current market conditions better, and cite the retailer panel survey as the source.

D1 · Agentic Architecture

Q672. Analysts reviewing a refill-products market report find that the synthesis covers consumer attitudes and pricing in depth but says almost nothing about packaging regulation, which the client's question named explicitly. The current pipeline makes one search pass, one analysis pass and one synthesis pass, then writes the report. What design change best addresses this? (Select ONE.) A) Have the coordinator check synthesis against the question's required topics, send targeted queries for gaps, and re-run synthesis. B) Instruct the synthesis subagent to fill any gap from its own general knowledge when the gathered sources do not cover a topic the client asked about. C) Ask web_search to return three times as many results on its single pass, so the first round is more likely to include sources on every topic. D) Give the synthesis subagent its own search_web tool so it can look up missing topics by itself while it writes, without involving the coordinator.

D1 · Agentic Architecture

Q673. Marrowfield's coordinator always runs all four subagents in order. For a client request that reads "summarize what our own uploaded consumer panel studies conclude, no outside sources", it still runs web_search for six minutes, and 30% of runs like this exceed the 30-minute budget. What change best fixes this? (Select ONE.) A) Raise the run budget to 45 minutes for requests that mention client documents, so the full four-stage pipeline can complete without timing out. B) Have the coordinator analyze each request's requirements and invoke only the subagents it needs, skipping web_search when the client limits sources to its own documents. C) Keep the fixed pipeline but tell report_writer to discard any web sources whenever the request says the client wants no outside sources. D) Run all four subagents in parallel instead of in sequence, so the web search time overlaps with document analysis and the total run time drops.

D5 · Context & Reliability

Q674. Analysts review a random 10% of reports, and measured citation accuracy is 96% overall. A client then found misattributed figures in a report built mostly from scanned retailer sales reports. Leadership wants to reduce review for claims the system marks as high-confidence. Which TWO changes should Marrowfield make before reducing review? (Select TWO.) A) Calibrate the claim-level confidence thresholds against a labeled validation set of checked claims before using those scores to decide what skips review. B) Measure citation accuracy separately by source type and claim type, using stratified random samples that include high-confidence claims, before cutting review anywhere. C) Ask the synthesis model to rate its own confidence in each claim and skip human review for any claim it rates above 0.9. D) Raise the uniform random review sample from 10% to 20% of reports, so the overall accuracy estimate becomes more precise for leadership.

D1 · Agentic Architecture

Q675. For broad topics, Marrowfield's coordinator splits research into four independent subtopics and delegates each to web_search. Traces show it requests one subagent per turn and waits for each result before requesting the next, so the search phase takes about 24 minutes of the 30-minute budget. The subtopics do not depend on each other. What change best reduces the search phase time? (Select ONE.) A) Merge the four subtopics into one larger delegation to a single web_search subagent, so the coordinator waits for only one result. B) Send the four subtopic searches through the Message Batches API as one batch, so they are processed together and collected when the batch completes. C) Have the coordinator request all four subagents in a single response, with one Task (Agent) tool call per subtopic, so they run in parallel. D) Instruct each web_search subagent to return shorter results, so each sequential delegation finishes sooner and the four calls fit the budget.

D1 · Agentic Architecture

Q676. Besides open-ended briefings, Marrowfield runs a weekly market digest: fetch the same six regulator and trade-association sources, extract new items, summarize each, then format one digest. The team runs it through the dynamic coordinator, which sometimes skips a source or reorders steps, so digests vary week to week and a missed source went unnoticed for a month. Which design fits the digest best? (Select ONE.) A) The same dynamic coordinator with a stronger prompt that lists the six sources and asks it to always include every one of them each week. B) A dynamic coordinator that first maps which sources look most newsworthy this week and then plans an adaptive investigation around those sources. C) A fixed sequential pipeline that runs fetch, extract, summarize and format for all six sources every week, with a check after each step. D) Four parallel subagents that each pick sources they consider relevant, followed by a synthesis step that merges whatever the four return.

D4 · Tools & MCP

Q677. The load_document MCP tool returns the text "Operation failed" for three different problems: a password-protected PDF, a file ID that does not exist, and a storage service timeout. The doc_analysis subagent retries every failure five times, wasting about four minutes per run, and never tells the coordinator which documents need the client's help. What should the tool return instead? (Select ONE.) A) The same "Operation failed" text, with the subagent's prompt changed to retry each failed document only twice before it moves on to the next document. B) An isError result with an errorCategory such as permission, validation or transient, an isRetryable flag, and a short human-readable description of the cause. C) A raised exception that stops the whole research run, so an engineer can inspect the logs and fix the document problem before any retry. D) An empty successful result for any failed document, so the subagent continues with the remaining documents instead of spending its limited time on retries.

{"isError": true, "content": [{"type": "text", "text": "{\"errorCategory\": \"permission\", \"isRetryable\": false, \"document_id\": \"doc-118\", \"description\": \"PDF is password-protected; ask the client for the password.\"}"}]}

D4 · Tools & MCP

Q678. To keep things flexible, the report_writer subagent was given every tool in the system, 14 in total, including search_web and fetch_page. Reviews show report_writer now runs its own searches while formatting and inserts new statistics that never went through doc_analysis or synthesis, and some have no citation. What change best fixes this? (Select ONE.) A) Restrict report_writer to the formatting and citation tools its role needs, and route new research needs through the coordinator. B) Give report_writer a larger model so it can tell better when outside research would improve the report and when it is not necessary. C) Keep all 14 tools but add a system prompt line telling report_writer not to search the web or fetch pages while it formats the report. D) Keep all 14 tools and add a final check that deletes any statistic in the report that does not appear anywhere in the synthesis output.

Scenario 3 · Developer Productivity with Claude

D1 · Agentic Architecture

Q679. Yesterday an engineer used Atlas in a named session, billing-map, to analyze 40 files of the billing service and draft a refactoring plan. Overnight a teammate merged a pull request that changed 3 of those 40 files. Today the engineer wants to continue the plan without paying for a full re-analysis. What should the engineer do? (Select ONE.) A) Start a new session and have Atlas re-read all 40 files from scratch, so every part of the plan is based on fresh reads of the code. B) Resume billing-map and continue planning directly, since 37 of the 40 analyzed files are still unchanged and most of the earlier analysis is still valid. C) Fork billing-map into a new branch and continue planning there, so the original session stays untouched in case the changes cause problems. D) Resume billing-map and tell Atlas which three files changed in the merged pull request, asking it to re-read only those files and update the plan.

D1 · Agentic Architecture

Q680. Atlas's scheduled bot must never force-push and must never edit files under infra/prod/; such changes need a human-approved pull request. The system prompt states both rules, yet last week one run force-pushed a branch. The team must guarantee these rules in the Agent SDK configuration. Which change does that? (Select ONE.) A) List only safe commands such as git commit and normal git push in allowed_tools, so a force push and edits under infra/prod/ are no longer available to the bot. B) Move both rules to the top of the system prompt, mark them as critical, and add an example of a refused force push, so the model treats them as top priority. C) Add a PostToolUse hook that detects a force push or an infra/prod/ edit after it happens and immediately notifies the team channel with the details. D) Register a PreToolUse hook that denies force-push Bash calls and Edit or Write calls under infra/prod/, telling Atlas to open a pull request instead.

from claude_agent_sdk import ClaudeAgentOptions, HookMatcher

async def block_unsafe(input_data, tool_use_id, context):
    tool_input = input_data.get("tool_input", {})
    command = tool_input.get("command", "")
    path = tool_input.get("file_path", "")
    if "push" in command and ("--force" in command or " -f" in command) or path.startswith("infra/prod/"):
        return {"hookSpecificOutput": {"hookEventName": "PreToolUse", "permissionDecision": "deny", "permissionDecisionReason": "Not allowed for the bot. Open a pull request for human approval."}}
    return {}

options = ClaudeAgentOptions(hooks={"PreToolUse": [HookMatcher(matcher="Bash|Edit|Write", hooks=[block_unsafe])]})

D4 · Tools & MCP

Q681. Atlas's system prompt says, "When the user asks about a service, always check the service catalog first." Tool descriptions for catalog_lookup, Grep and Read are clear and specific. Yet questions like "what does this service class's retry method do" trigger catalog_lookup instead of reading the code. What should the team change first? (Select ONE.) A) Add few-shot examples of code questions answered with Grep and Read, while keeping the existing catalog instruction in the system prompt unchanged. B) Rename catalog_lookup to service_metadata_lookup and extend its description further, so its purpose is even clearer to the model than it is today. C) Add a routing classifier in front of Atlas that detects code questions and removes catalog_lookup from the tool list for those particular requests. D) Reword the instruction: use the catalog for ownership, deployment and dependency metadata, and read the code for implementation questions.

D4 · Tools & MCP

Q682. Harborlight runs a code_search MCP server that indexes symbols across all 300 repositories and returns call sites with file and line. Its tool description says only "Searches code." Traces show Atlas using Grep in the current checkout for cross-repository questions and missing callers in other repositories about 70% of the time. What should the team do first? (Select ONE.) A) Replace the code_search server with a custom MCP server built by the platform team that wraps Grep across a local copy of every repository. B) Remove Grep from Atlas's tool list entirely, so code_search becomes the only search tool available and Atlas must use it for every code search in any repository. C) Add a system prompt line telling Atlas to always use code_search for every code search it runs, including quick searches within the current checkout. D) Expand code_search's description: all 300 repositories, symbol-aware call sites with file and line, preferred over Grep for cross-repository caller questions.

D1 · Agentic Architecture

Q683. When Atlas cannot finish an unattended framework migration, it opens a ticket for the on-call engineer containing only its last message, such as "I was unable to complete the migration." On-call engineers cannot see Atlas's session transcript and spend about 40 minutes rediscovering what happened. What should the ticket contain? (Select ONE.) A) The same short message, but with Atlas instructed to retry the migration three more times before it opens the ticket for the engineer. B) The complete session transcript exported as an attachment, so the on-call engineer can read every step Atlas took from the beginning of the run. C) A structured handoff: repository and branch, what Atlas attempted, the root cause it found, the failing command and an excerpt of its output, and a recommended next action. D) A severity label and the name of the failing service, so the on-call engineer can prioritize the ticket quickly and then investigate from scratch.

D1 · Agentic Architecture

Q684. Atlas's coordinator delegates to three subagents: code_explorer, catalog_lookup and boilerplate_gen. To save a round trip, the team let code_explorer call boilerplate_gen directly. Since then, failed generations never reach the coordinator, traces show work the coordinator cannot explain, and engineers sometimes get boilerplate for the wrong service. What change best restores control? (Select ONE.) A) Route all communication between subagents through the coordinator, so it decides every delegation, sees every result and handles every error consistently. B) Add a shared message bus that all three subagents can publish to and read from, so each one sees what the other two are doing in real time as work proceeds. C) Give every subagent the tools of the other two, so no subagent ever needs to call another one and round trips disappear completely. D) Add detailed logging inside each subagent, so engineers can reconstruct direct calls between subagents from the logs after a failure.

D4 · Tools & MCP

Q685. Harborlight engineers also use Claude Code directly in their repositories. The platform team wants every engineer who clones the payments repo to get the service catalog MCP server automatically, each authenticating with their own token. Separately, one engineer wants to trial an experimental profiler MCP server for themselves without affecting teammates. Which TWO configurations should they use? (Select TWO.) A) Define the service catalog server in the repository's .mcp.json, committed to git, with its Authorization header set to Bearer ${CATALOG_TOKEN} from each engineer's environment. B) Define the experimental profiler server in the repository's .mcp.json and ask teammates to ignore it until the engineer finishes the trial. C) Define the service catalog server in the repository's .mcp.json with a shared team token written directly into the header, since the repository is private. D) Define the experimental profiler server in that engineer's user-level configuration in ~/.claude.json, so it is available only on their machine.

{"mcpServers": {"service-catalog": {"type": "http", "url": "https://catalog.internal.example/mcp", "headers": {"Authorization": "Bearer ${CATALOG_TOKEN}"}}}}

D1 · Agentic Architecture

Q686. Atlas began as a hand-written loop over the Messages API. A developer trimmed history between iterations to save tokens, so every call carried just the engineer's question and the newest tool output. In traces, the model greps for PaymentRetryPolicy four times in a row and never opens the files its first grep found. Which loop change solves this? (Select ONE.) A) Keep sending only the latest result, and add a cache that returns the earlier output instantly whenever the model repeats an identical Grep. B) Keep sending only the latest result, but add a system prompt line that tells the model never to repeat a search it has already run. C) Keep sending only the latest result, and stop the loop after three tool calls so a repeated search cannot continue for very long. D) Append each assistant tool_use response and its matching tool_result to messages, and resend the full history on every call.

D1 · Agentic Architecture

Q687. In unattended mode, Atlas bumps dependencies and then calls the deployment MCP server's create_preview tool. The system prompt says to run the test suite and confirm it passes before creating a preview. In 4% of runs Atlas creates a preview from a failing commit, and a downstream team then tests broken builds. What change best prevents this? (Select ONE.) A) Run a nightly audit that finds previews created from failing commits, deletes them, and posts a summary of all the deleted previews to the team's channel. B) Add a programmatic gate, such as a PreToolUse hook, that blocks create_preview unless run_tests has returned a pass for the current commit in this session. C) Add three few-shot examples to the system prompt that show Atlas running the tests, reading a pass, and only then calling create_preview. D) Rewrite the system prompt instruction in capital letters and repeat it at the end of the prompt, so the requirement to test first stands out much more clearly.

async def require_green_tests(input_data, tool_use_id, context):
    commit = input_data.get("tool_input", {}).get("commit")
    if TEST_RESULTS.get(commit) != "pass":
        return {"hookSpecificOutput": {"hookEventName": "PreToolUse", "permissionDecision": "deny", "permissionDecisionReason": "Run run_tests on this commit and get a pass before create_preview."}}
    return {}

D4 · Tools & MCP

Q688. An engineer asks Atlas where the retry budget for billing exports is calculated. Atlas starts by using Glob to list all 900 files under billing/ and then Reads them one by one, and it runs low on context before answering. Engineers expect an answer within a few tool calls. What approach should Atlas use for this kind of question? (Select ONE.) A) Read all 900 files in parallel subagents, each returning a short summary, and then search those summaries for the retry budget. B) Switch Atlas to a model with a larger context window, so reading all 900 files fits before it has to answer the engineer. C) Grep for identifiers such as retryBudget to find entry points, then Read those files and follow their imports. D) Glob for **/Retry.java across billing/ and Read every matching file in full before answering, so no relevant file is missed.

D4 · Tools & MCP

Q689. Atlas must change timeout_seconds from 30 to 45 in the payments block of a generated config.yaml. The file has four service blocks that are byte-for-byte identical for 30 lines, differing only in a section header 40 lines above each. Edit fails because the target text is not unique. What should Atlas do to make the change reliably? (Select ONE.) A) Use Bash to run sed on the first line that matches timeout_seconds: 30, since the payments block is listed first in most config files. B) Run Edit with replace_all enabled, so every matching timeout_seconds line in the file is changed from 30 to 45 in one call. C) Retry the same Edit call several times, since an Edit that fails on a large file may succeed on a later attempt. D) Read the full file, change only the payments block's timeout in the loaded content, and Write the complete updated file back.

D1 · Agentic Architecture

Q690. An engineering manager asks Atlas to "add meaningful tests to billing-reconciler," a 180-file legacy service with no tests. Atlas currently starts at the first file alphabetically and writes tests file by file. After two days, most new tests cover logging helpers and config parsing, while the money-moving code is still untested. How should Atlas decompose this task? (Select ONE.) A) Map the structure, find high-impact areas such as money paths, then follow a prioritized plan that adapts as dependencies appear. B) Set a target of 80% line coverage for the service and have Atlas keep writing tests for whichever files raise the coverage number fastest. C) Use a fixed prompt chain that runs the same generate-tests step on every file in alphabetical order, with a review step after each file. D) Spawn 180 parallel subagents, one per file, each writing tests for its own file, and merge all of their test files together at the end.

D1 · Agentic Architecture

Q691. When explaining incidents, Atlas combines events from three MCP servers. The issue tracker returns Unix epoch seconds, the deployment server returns ISO 8601 with local offsets, and the service catalog returns relative strings such as "3 days ago." Atlas sometimes orders events wrongly in its timelines. Two of the servers belong to other teams. What is the most reliable fix? (Select ONE.) A) Add a system prompt section that explains each server's timestamp format and asks the model to convert all of them to UTC before ordering events. B) Ask the other two teams to change their MCP servers to return ISO 8601 UTC, and wait for their releases before changing anything in Atlas. C) Add a convert_timestamp tool that the model can call whenever it is unsure how to compare two timestamps from different servers. D) Add a PostToolUse hook in Atlas that converts every timestamp in these tools' results to ISO 8601 UTC before the model reads them.

D4 · Tools & MCP

Q692. The issue tracker MCP tool search_issues returns an empty list both when a project has no matching issues and when Atlas's token lacks access to that project. Engineers asking about restricted security projects are told "there are no open bugs," which is false. The issue tracker server is maintained by Harborlight. What should search_issues return for the access case? (Select ONE.) A) An isError result with a permission error category, isRetryable set to false, and a description naming the project the token cannot read. B) The same empty list, plus a system prompt line telling Atlas to mention that results may be incomplete whenever a search returns nothing at all. C) A successful result containing one placeholder issue titled "access denied," so the agent sees something instead of an empty result list. D) The same empty list, with Atlas configured to retry each empty search three times in case access problems are only temporary.

D1 · Agentic Architecture

Q693. When an engineer asks how a legacy billing flow works, Atlas's coordinator sends code_explorer a 14-step script: Glob *.java in billing/, Read each file, then list classes ending in Service, and so on. It works for Java services but fails on the Go services and on flows that cross into the ledger repository. How should the coordinator's delegation prompt change? (Select ONE.) A) Add a second 14-step script for Go services and a third for cross-repository flows, and have the coordinator choose the script by language. B) State the question to answer, the evidence required such as file paths and call chains, and the expected output, and let code_explorer choose its own search steps. C) Keep the Java script but add a final step telling code_explorer to adapt the earlier steps if the code turns out to be in another language. D) Replace the script with a single line asking code_explorer to explain the billing flow, and leave every other detail to the subagent.

Scenario 4 · Claude Code for Continuous Integration

D3 · Prompting & Structured Output

Q694. A validator checks each structured review finding before posting, and failures are retried with the validation error appended. Last week's failures fell into four groups. The team wants to keep retry-with-feedback only where it can work and route the other groups elsewhere. For which TWO groups is a retry with the specific validation error likely to fix the finding? (Select TWO.) A) Findings whose line number points outside the changed hunks of the diff, although the issue described is in a changed function nearby. B) Findings that must cite the approved design document, which lives on an internal wiki page that is not part of the review input. C) Findings whose suggested_fix uses a variable name from before a rename that is visible in the same diff the model was given. D) Findings whose required ticket_id is missing because the linked tickets exist only in the issue tracker, which the job never provides.

D3 · Prompting & Structured Output

Q695. Brightmoor wants to keep posting high-value inline comments while sending doubtful findings somewhere cheaper for developers to check. Reviewers have a labeled set of 1,200 past findings marked accepted or dismissed. Inline comments should be limited to findings likely to be accepted, and the rest should go to one collapsible summary comment. Which design fits best? (Select ONE.) A) Add a verification pass that attaches a confidence score to each finding, set the inline threshold from the labeled set, and post lower-confidence findings in the summary. B) Post every finding inline as today, and add a reaction button so developers can mark each comment as useful or not useful for later analysis. C) Ask the review model to report its own confidence with each finding, and post inline any finding with a stated confidence of 0.8 or higher. D) Run the review twice in the same session and post inline only the findings that appear in both runs, sending the rest to the summary.

D3 · Prompting & Structured Output

Q696. The CI job generates missing unit tests and then, in the same Claude Code session, asks Claude to "review these tests critically before posting." Several posted tests assert the current buggy behaviour, for example expecting a refund of zero for partial cancellations, and the self-review approved them. What change will most improve the review of generated tests? (Select ONE.) A) Keep the review in the same session and rewrite the review instruction to demand that Claude find at least three real problems in every set of generated tests before posting. B) Review the generated tests in a separate, independent Claude Code invocation that receives the code, the tests and the requirements but not the generator's reasoning. C) Keep the review in the same session and enable extended thinking, so Claude reasons more deeply about each test before approving it. D) Keep the review in the same session and run it twice in a row, posting only the tests that both review passes approve.

D3 · Prompting & Structured Output

Q697. In the legacy adapters/ package, broad except Exception blocks are an accepted pattern when they log the error and re-raise a domain error at an integration boundary. The review flags all of them, and developers dismiss those findings. But last month a broad except that silently swallowed a purchase error caused an incident. How should the team teach the review the difference? (Select ONE.) A) Turn off the error-handling category entirely, since its findings in adapters/ are dismissed so often that developers no longer trust any of them. B) Post-filter findings with a regular expression that removes any comment about except Exception, so the pattern is never flagged anywhere. C) Add prompt examples contrasting a boundary handler that logs and re-raises with one that swallows the error, explaining the reasoning for each. D) Add every file path in adapters/ to a skip list in the prompt, so broad except blocks in those files are never reported by the review.

D3 · Prompting & Structured Output

Q698. The nightly test-generation job runs an agentic loop: Claude writes a test, requests a run_pytest tool call, reads the failure, and revises, usually over six to ten tool rounds per module. A manager proposes moving this job to the Message Batches API to lower cost, since results are not needed until morning. What is the main problem with the proposal as described? (Select ONE.) A) Batch results arrive in random order, so the script would be unable to tell which generated tests belong to which module once the whole batch completes. B) The Message Batches API does not accept tool definitions at all, so the run_pytest tool could not even be offered to the model inside a batch request. C) A batch request cannot run your run_pytest tool mid-request and continue, so each tool round would need a new batch submission with its own processing wait. D) Batch processing is only suitable for blocking workflows such as pre-merge checks, so an overnight job is the wrong kind of workload for it.

D2 · Claude Code

Q699. When the tech lead runs Claude Code locally, reviews follow Brightmoor's review standards, such as always flagging unparameterized SQL. The CI job on GitHub Actions ignores those standards. Running /memory locally shows the standards are loaded from ~/.claude/CLAUDE.md, while the repository's CLAUDE.md contains only build commands. What should the team change? (Select ONE.) A) Copy the tech lead's ~/.claude/CLAUDE.md into the CI runner's home directory during each job, so the runner loads the same personal file as the lead. B) Ask each developer to copy the standards into their own ~/.claude/CLAUDE.md, so every local session and review follows the same rules. C) Add the review standards to the repository's .claude/settings.json, so Claude Code applies them as configuration in every environment. D) Move the review standards into the repository, in the project CLAUDE.md or a .claude/rules/ file, so every checkout, including CI, loads them.

D3 · Prompting & Structured Output

Q700. To cut the 31% dismissal rate, the team added "Be conservative and only report issues you are highly confident about" to the review prompt. Two weeks later dismissals are at 29%, and developers still dismiss naming nits, style preferences and comments about patterns the codebase uses everywhere. What change will most improve precision? (Select ONE.) A) Replace the instruction with explicit criteria: report behaviour-changing bugs, security issues and data-loss risks, and skip style, naming and established local patterns. B) Keep the instruction and post only the three findings with the highest model-stated confidence on each pull request, dropping all the others. C) Strengthen the instruction to "only report issues you are at least 95% confident about" and ask the model to state its confidence for each finding. D) Move the review to a larger, more capable model while keeping the current prompt, since better judgment should reduce noisy findings.

D3 · Prompting & Structured Output

Q701. The nightly security audit calls the Messages API directly. The prompt asks Claude to "respond only with a JSON array of findings," and about 3% of responses fail to parse because of trailing commentary, missing brackets or unescaped quotes. The parser then drops the whole file's findings. What is the most reliable way to get schema-compliant output? (Select ONE.) A) Define a record_findings tool whose input_schema describes the findings, force it with tool_choice, and read the tool_use input. B) Keep the prompt and prefill the assistant's reply with an opening bracket, so the model is pushed to start its answer as a JSON array. C) Keep the prompt and add a retry that resends the same request whenever the response fails to parse, up to three times for each file. D) Keep the prompt and lower the temperature to zero, so responses become more predictable and the JSON is less likely to contain mistakes.

{"tools": [{"name": "record_findings", "description": "Record security findings for one file.", "input_schema": {"type": "object", "properties": {"findings": {"type": "array", "items": {"type": "object", "properties": {"line": {"type": "integer"}, "severity": {"type": "string", "enum": ["low", "medium", "high", "critical"]}, "description": {"type": "string"}}, "required": ["line", "severity", "description"]}}}, "required": ["findings"]}}], "tool_choice": {"type": "tool", "name": "record_findings"}}

D2 · Claude Code

Q702. The CI job's test proposals are often rejected. Developers say about half duplicate scenarios that the existing test files already cover, and many others test trivial getters while ignoring the team's fixtures for fake platform stores. The job currently gives Claude only the changed source files. What change will most improve the test proposals? (Select ONE.) A) Add a post-processing step that removes any proposed test whose name matches the name of an existing test function in the repository. B) Move test generation to a larger model, keeping the same inputs, so it can infer the team's fixtures and coverage from the changed files alone. C) Ask the job to generate twice as many test proposals per pull request, and let developers choose the most useful ones from the larger list. D) Include each changed module's existing test files in context, and document valuable-test criteria and available fixtures in CLAUDE.md.

D3 · Prompting & Structured Output

Q703. The audit's finding schema makes cwe_id a required string. Security reviewers find that many findings for business-logic flaws, which have no clear CWE mapping, carry CWE numbers that do not fit, such as CWE-79 on a rounding error. The dashboard groups findings by cwe_id. What schema change best addresses this? (Select ONE.) A) Keep cwe_id required and add a validation step that rejects any CWE number outside a list of the twenty most common web weaknesses. B) Make cwe_id nullable so it can be null when no CWE applies, and add a category enum that includes "other" with a detail string for unmapped flaws. C) Remove cwe_id from the schema entirely and ask reviewers to assign CWE numbers by hand to every finding on the dashboard each morning instead. D) Keep cwe_id required and add an instruction telling the model to choose the closest matching CWE number whenever no exact mapping exists for a finding.

D3 · Prompting & Structured Output

Q704. Brightmoor records when developers dismiss a finding, but the record holds only the comment text. When the team tries to analyze last month's 31% dismissal rate, it cannot tell which code constructs keep producing dismissed findings, because the same issue is worded many different ways. What change best enables this analysis? (Select ONE.) A) Track the overall dismissal rate per week on a dashboard, so the team can see whether its prompt changes are reducing dismissals. B) Add a confidence field to each finding and compare the average confidence of dismissed findings with the average of accepted ones. C) Ask developers to type a short free-text reason every time they dismiss a finding, and store those reasons next to the original comment text in the log. D) Add a detected_pattern field to each structured finding that names the code construct that triggered it, and record that field with every dismissal.

D2 · Claude Code

Q705. When a developer pushes new commits to an open pull request, the CI job re-runs claude -p on the full diff. Developers complain that it re-posts comments they already fixed and repeats open findings in new wording, so one pull request collected 19 near-duplicate comments. What change should the team make to the re-run? (Select ONE.) A) Delete all of the bot's earlier comments on the pull request before each re-run, then post the complete new set of findings. B) Pass the earlier findings into the re-run's context and ask Claude to report only new or still-unaddressed issues. C) Review only the newest commit's diff on re-runs, so Claude sees just the latest changes and cannot comment on older code at all. D) Have the posting script skip any new comment whose text exactly matches an earlier comment, so identical duplicates are never posted twice.

D3 · Prompting & Structured Output

Q706. The review job already uses --json-schema with fields for file, line, severity, issue and suggested_fix, and every response parses. But the text inside those fields varies widely: some issues read "consider refactoring this," some suggested fixes are empty strings, and others are full paragraphs. Detailed prose instructions about writing style are already in the prompt. What change will most improve consistency? (Select ONE.) A) Add minLength constraints to the issue and suggested_fix fields in the JSON schema, so short or empty values are rejected during validation. B) Add two to four example findings to the prompt showing a specific issue, its impact and a concrete fix in the desired style. C) Raise the model's temperature setting slightly, so the wording of findings becomes more varied and less repetitive across pull requests. D) Add more detailed prose instructions describing how long each field should be, which words to avoid, and what a concrete suggested fix means.

{"file": "store/refund.py", "line": 88, "severity": "high", "issue": "refund_amount is compared to the original charge in dollars, but charges are stored in cents, so refunds up to 100 times the charge pass the check.", "suggested_fix": "Compare cents to cents: if refund_cents > charge.amount_cents: raise RefundTooLarge()"}

D3 · Prompting & Structured Output

Q707. Dismissal data by category shows that security findings are accepted 88% of the time, while documentation-accuracy findings are dismissed 74% of the time. Developers now skim past every bot comment, including security ones. The team needs time to rewrite the documentation criteria. What should it do right now? (Select ONE.) A) Keep every category on and add a note to each documentation-accuracy comment saying that this category is still experimental and may safely be ignored. B) Turn off the documentation-accuracy category for now, keep security and other trusted categories on, and re-enable it once its criteria are fixed. C) Turn off all bot comments until the documentation criteria are rewritten, so developers are not exposed to any noisy findings during the change. D) Keep every category on and lower the number of documentation findings to one per pull request, so their noise has much less effect on developer trust.

D3 · Prompting & Structured Output

Q708. The nightly security audit now submits about 2,000 per-file requests through the Message Batches API. Last night 37 requests failed: 29 exceeded the context limit because they were large generated files, and 8 hit transient server errors. The current script resubmits the whole batch whenever any request fails. What should the script do instead? (Select ONE.) A) Move the 37 failed files to synchronous Messages API calls with the same prompts, so they complete immediately without another batch. B) Use each result's custom_id to resubmit only the 37 failed requests, splitting the 29 oversized files into chunks and resending the 8 transient failures unchanged. C) Resubmit the whole batch of about 2,000 requests again, so the results from both runs can be compared and the most complete one kept. D) Resubmit only the 37 failed requests exactly as they were, since retrying them unchanged usually succeeds on a second batch run.