Level 11 — Working a Real Codebase

12 questions, drawn from a bank of 34 across 27 topics. Pass at 80%. Every wrong option carries a note explaining why it is tempting — those notes are the point of the quiz, so read them whether you pass or not.

attempt 1

1 / 12 · Choose one
Checkout is failing and you are fairly sure — not certain — that the currency-rounding helper is involved. You are also mindful that asserting a cause makes everyone build on your guess. What do you do with it?
Checkout is failing and you are fairly sure — not certain — that the currency-rounding helper is involved. You are also mindful that asserting a cause makes everyone build on your guess. What do you do with it?
Report the symptom, then offer the helper as the likely location rather than the cause — a guess you only half believe still bounds the search, and a wrong one is corrected in a turnState it as the cause — you are the one who saw the failure, and the session can correct you if it disagreesAsk Claude to name the likely cause first, then say whether it matches yoursWithhold it and report only what you observed. An unproven guess is exactly the kind of thing a session builds on, and a few extra turns of searching costs less than an investigation anchored to a hunch you could not support
2 / 12 · Choose one
You spent yesterday in a --worktree session and chose to keep the worktree. Today you resume that session from the main checkout. Where does it pick up?
You spent yesterday in a --worktree session and chose to keep the worktree. Today you resume that session from the main checkout. Where does it pick up?
In the directory you launched from; the worktree binding is a per-run thing that does not survive an exitIn a fresh worktree branched from the same base, since the old one may have driftedIn the main checkout, with the worktree's branch checked out there: resuming brings the session's branch to you rather than moving you into a directory that may have been removed since you last used itBack inside its worktree — Claude Code returns a resumed session to the worktree it was in, for interactive resumes, for --continue and --resume under -p, and for the Agent SDK
3 / 12 · Choose one
A refactor moves a daemon's config loader. The behaviour you have to protect — that the daemon comes up with its config intact — is not something the suite observes, and the daemon needs a database and an env file before it will start at all. How do you get a check for it?
A refactor moves a daemon's config loader. The behaviour you have to protect — that the daemon comes up with its config intact — is not something the suite observes, and the daemon needs a database and an env file before it will start at all. How do you get a check for it?
Accept that startup is not the kind of behaviour a check can observe, lean on the pre-existing suite and the clean diagnostics, and note the uncovered path in the pull requestAdd a startup case to the characterization suite with the database and the env file mocked out, so it runs anywhereRun /verify and take the result at face value. It builds and drives the real app, which is the strongest evidence available for a behaviour no assertion covers, and it works without setup on any project type it recognises — a daemon with a package.json and a start command documented in the README includedBuild the check anyway by running the app: /verify builds and runs it rather than falling back to tests or type checks — and because a database and an env file are exactly where its launch inference stops being reliable, record the recipe once with /run-skill-generator first
4 / 12 · Choose all that apply
You add typescript-lsp@claude-plugins-official to the project's enabledPlugins and push. A teammate pulls, opens Claude Code, and has no LSP tool. Which explanations are consistent with the documentation?
You add typescript-lsp@claude-plugins-official to the project's enabledPlugins and push. A teammate pulls, opens Claude Code, and has no LSP tool. Which explanations are consistent with the documentation?
enabledPlugins only works at user scope, so a project setting for it is ignoredThe repository's Read deny rules on generated code switched the tool off for everyoneThey have not installed typescript-language-server; the plugin does not install it for them, and the /plugin **Errors** tab would read Executable not found in $PATHThey are working in a cloud session, where Claude Code does not start plugin language servers
5 / 12 · Choose all that apply
You work from the root of a nine-package monorepo with a CLAUDE.md per package, so each package's conventions stay out of the window until Claude reads a file there. Which of these are true of @ references in that repository?
You work from the root of a nine-package monorepo with a CLAUDE.md per package, so each package's conventions stay out of the window until Claude reads a file there. Which of these are true of @ references in that repository?
A file reference adds the CLAUDE.md from that file's directory and its parents, so one @ into packages/web/src/ brings that package's conventions in even though the task never touches itA directory reference can return a listing large enough to be its own context cost while still containing no codeAn @ into another package loads that package's skills and rules along with its instructions, the way --add-dir doesclaudeMdExcludes guarantees no package instructions can arrive through an @, since exclusions are matched before anything loads
6 / 12 · Choose one
You resume a five-week-old session, pick a checkpoint from the list /rewind offers, choose **Restore code**, and get No files were restored: 1 file failed (backup missing, or the file could not be updated). The file is writable and nothing else has it open. What happened?
You resume a five-week-old session, pick a checkpoint from the list /rewind offers, choose **Restore code**, and get No files were restored: 1 file failed (backup missing, or the file could not be updated). The file is writable and nothing else has it open. What happened?
Something held the path open while the restore ran, so the fix is the ordinary one — close whatever has it and run /rewind again, which succeeds once the write can go throughThe checkpoint fell off the 100-most-recent cap and its snapshot files went with it, which is why a session with a few hundred prompts in it cannot be rewound to its first afternoon however recently you resumed itNothing that cannot be undone: raise cleanupPeriodDays and run /rewind again, since that setting is what governs how long backups are keptThe retention sweep deleted this session's backups about 30 days after it last saved one, so a second /rewind fails the same way and the undo has to come from somewhere else
7 / 12 · Choose one
One shared type plus its call sites in six packages. A colleague proposes one session per package, arguing that each stays short enough never to compact. What is wrong with the argument?
One shared type plus its call sites in six packages. A colleague proposes one session per package, arguing that each stays short enough never to compact. What is wrong with the argument?
It buys a smaller problem with a bigger one: each session re-derives the reasoning from a different starting context, so naming, error handling and the migration shim come out subtly different — while a long single session keeps its plan across compactions anywayThe six sessions would collide on the shared type unless each ran in its own worktreeNothing is wrong — compaction is lossless for anything Claude read from disk, so a long session loses nothing either wayOnly the token cost, which is real but is not the objection that matters: six sessions each reload the repository's instructions and each re-read the shared type from disk, so you pay for the same context six times over — and every one of those sessions really does stay short enough never to compact
8 / 12 · Choose one
A code intelligence plugin is installed and its language server binary is on $PATH. What does Claude gain that it did not have before?
A code intelligence plugin is installed and its language server binary is on $PATH. What does Claude gain that it did not have before?
An index of the repository that makes searches skip generated and vendored code automaticallyA faster, smarter search that returns fewer irrelevant grep hitsA compile-and-test step Claude can run to confirm a change behaves correctlyClaude Code's built-in LSP tool: jump to definition, find references, and type errors reported back automatically after each edit
9 / 12 · Choose one
The first adapter is converted, its recorded-fixture diff is green, and eleven remain. What is the cheapest way to get the rest converted the same way?
The first adapter is converted, its recorded-fixture diff is green, and eleven remain. What is the cheapest way to get the rest converted the same way?
Ask for a plan describing the intended shape of all twelve files before any more are convertedPoint at the file: "convert the remaining files in src/adapters/ the same way you converted src/adapters/stripe.ts"Write the conversion rules into the package's CLAUDE.md so every future session follows themDescribe the target style more carefully than you managed to the first time
10 / 12 · Choose all that apply
The repository root denies Read(./**/*.generated.*). A teammate reports that Claude can no longer open generated files, and then you watch it print the contents of one. Which routes get past the rule?
The repository root denies Read(./**/*.generated.*). A teammate reports that Claude can no longer open generated files, and then you watch it print the contents of one. Which routes get past the rule?
The Read tool, with the generated file's pathA script or codegen tool Claude runs that opens the files itselfAn LSP find references call that lands inside the generated fileA Bash grep -r or find over a directory that contains denied filescat packages/api/src/types.generated.ts in Bash
11 / 12 · Choose one
A session edits six files with the Edit tool, runs a formatter over the whole package in one Bash call, then hands the imports to a subagent. You rewind to before all three. What single action taken beforehand would have covered all of them?
A session edits six files with the Edit tool, runs a formatter over the whole package in one Bash call, then hands the imports to a subagent. You rewind to before all three. What single action taken beforehand would have covered all of them?
Raising cleanupPeriodDays, so the file snapshots survive longerNaming the session, so its checkpoints are kept rather than sweptA commit — git checkout then covers every channel, including the two a checkpoint never heldRunning the subagent in the foreground, which puts every change back inside the checkpoint
12 / 12 · Choose one
From packages/api/ you need to read and edit packages/shared/, but you do not want the shared package's CLAUDE.md, rules or skills entering your window. Which grant does exactly that and nothing more?
From packages/api/ you need to read and edit packages/shared/, but you do not want the shared package's CLAUDE.md, rules or skills entering your window. Which grant does exactly that and nothing more?
Start at the repository root with { "claudeMdExcludes": ["**/packages/shared/**"] }{ "permissions": { "additionalDirectories": ["../shared"] } } in packages/api/.claude/settings.jsonclaude --add-dir ../sharedCLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=0 claude --add-dir ../shared

12 questions still to answer