01 — Quiz: runtime lifecycle

Pick an answer. The quiz checks it and shows the explanation.

Part 1

#

1. Put these steps in order:

1. result added to context
2. PreToolUse hook runs
3. model requests a tool
4. host loads CLAUDE.md
5. model writes the final answer
6. MCP client sends tools/call

a) 3, 4, 2, 6, 1, 5

b) 4, 3, 6, 2, 1, 5

c) 4, 3, 2, 6, 1, 5

d) 4, 2, 3, 6, 5, 1

c — init first (4), then the model requests (3), the host checks (hook, 2), the host routes to the server (6), the result enters context (1), the model answers (5).

#

2. Model, host, or MCP server? Who does each of these?

1. decides to run npm test
2. runs it
3. calls the GitHub REST API
4. enforces a deny rule
5. turns the tool request into JSON-RPC

a) model, host, MCP server, host, host

b) model, host, host, model, MCP server

c) host, host, MCP server, model, host

d) model, MCP server, MCP server, host, model

a

1. model — it decides and requests.
2. host — it executes built-in tools like Bash.
3. MCP server — the GitHub MCP server calls GitHub's API.
4. host — enforcement lives outside the model.
5. host — its MCP client builds the tools/call message.

#

3. True or false: "Every time Claude uses a GitHub MCP tool, the server first sends the full tool list."

True

False

False. tools/list runs at connect (or from cache, on list_changed, or on reconnect). Each use is only tools/call → result. name and arguments are separate fields in that call.

#

4. You connected a GitHub MCP server with 80 tools, but they take up almost no context. Why? What happens when one tool is needed?

a) the host never discovered them: it fetches each schema from GitHub on first use

b) tool search: the model sees only the names, then calls ToolSearch to load the one schema it needs

c) the server compresses its schemas, and the model calls the tool directly

d) the host sends tools/list again each time a tool is needed

b — tool search. The host discovered all 80 tools, but the model sees only their names. When it needs one, it calls ToolSearch to load that one schema, then calls the tool.

#

5. How do Read config.yaml and a GitHub MCP call differ?

a) they don't: the host executes both

b) hooks and permission rules apply only to the built-in

c) only the MCP result goes into context

d) for MCP, the host sends JSON-RPC tools/call and the server executes. Everything else is the same

d

• Same: the model requests, hooks and rules apply, the result goes into context.
• Different: a built-in runs inside the host. For MCP, the host sends JSON-RPC tools/call over a transport and the server executes.

#

6. You have a deny rule Bash(git push --force:*). Claude reasons: "the user wants it fixed, force-push the branch." What happens?

a) it runs: the reasoning shows the user wants it

b) you get a permission prompt

c) the host blocks it, and the denial goes back to the model

d) the session ends

c — blocked by the host. Reasoning can't change it: enforcement is outside the model. The denial goes back to the model, so it can try another approach. Side note: an ask rule triggers a permission prompt, not a hook.

#

7. Your MCP server moves from stdio to HTTP. What changes? What stays?

a) how messages travel and the config change. The JSON-RPC messages, tool names, results and what the model sees stay

b) the JSON-RPC messages and tool names change. The config stays

c) everything changes, including what the model sees

d) nothing changes, not even the config

a

• Changes: how messages travel (local process + pipes vs URL, network, TLS, auth) and the config (command vs type: http + url).
• Stays: the JSON-RPC messages, tool names, results, and what the model sees.

#

8. npm start fails with "PAYMENTS_DB_URL is not set". Then Claude calls a GitHub tool instead of answering. Why?

a) the host ran the GitHub tool on its own after the error

b) the agent loop: the result entered context, the model reasoned again, and it still needed the root cause

c) the error ended the turn and a new one started

d) the model hallucinated a tool call

b — the agent loop. The result entered context, the model reasoned again, and it still needed the root cause. The final answer goes to you, not to the host.

#

9. You said "never touch prod" in chat. After auto-compaction, Claude seems to forget it. Why, and what fixes it?

a) the model hallucinated. Repeat the instruction more often

b) the session expired. Restart Claude Code

c) the instruction never reached the model. Send it again in chat

d) compaction summarized the chat-only instruction. Put it in the root CLAUDE.md, or enforce it with a deny rule or a PreToolUse hook

d — compaction summarized the chat-only instruction. It is not a hallucination.

• Fix 1: put it in the root CLAUDE.md. It is reloaded after compaction (advisory).
• Fix 2: a deny rule or a PreToolUse hook (enforced).
• Strongest: credentials with no prod access.

#

10. The nightly backup job fails. Claude reads backup.tf, uses an AWS MCP tool, runs aws s3 ls, then answers. Which description of who runs each step, which controls apply, and where the results go is right?

a) the model runs aws s3 ls itself, and only the MCP result enters context

b) the host sends tools/list before the AWS MCP call, then executes the AWS tool itself

c) the host runs Read and aws s3 ls after its checks, the AWS MCP server executes its tool after the host sends tools/call, and every result enters context

d) hooks and rules apply only to the Bash step

c

Step	Who / controls	Result
Read backup.tf	Host runs it. Hook + rules apply.	Content enters context (uses tokens).
AWS MCP tool	Model loads the schema via ToolSearch, then requests. Host checks (likely a prompt), sends tools/call. Server executes.	Result → context. No tools/list here: discovery already happened.
Bash aws s3 ls	Hook → rules → mode → prompt. Sandbox if enabled.	Output → context.
Answer	Model writes it, host displays it.	Context keeps growing; compaction near the limit.

Part 2 — Retest

#

1. The GitHub server connected at start. Claude calls its tools 4 times. The tools didn't change. How many tools/list requests happen during those 4 calls?

a) 4

b) 1

c) 0

d) 8

c — discovery happened at connect. Each use is only tools/call.

#

2. Tool search is on. Only list_buckets' name is in context. What happens first?

a) the host sends tools/list

b) the model calls ToolSearch; the host loads that schema from what it already discovered

c) the model calls the tool without a schema

d) the server pushes the schema

b — the host already has every schema from discovery. ToolSearch only loads one into context.

#

3. You ran claude mcp add --transport stdio db -- npx -y @bytebase/dbhub --dsn .... Who connects to PostgreSQL?

a) the Claude Code process

b) the model

c) a Bytebase cloud service

d) the DBHub process on your laptop, started by Claude Code

d — a stdio server is a separate local process. It holds the DB connection and executes the queries.

#

4. A built-in Read on a 400-line file. Does the result use context?

a) yes, every tool result does

b) no, built-ins are free

c) only MCP results do

d) only above 10,000 tokens

a — every tool result enters the context.

#

5. You read a file 20 minutes ago. No compaction since. Is it still in context?

a) no

b) only a summary

c) yes, until compaction or session end

d) only if re-read

c — results stay in context until compaction or the end of the session.

Quiz Progress

0 / 0 questions answered (0%)

0 correct

Quiz Complete!

0%

Reset quiz

Quiz Progress

• Answered: 0 / 0 (0%)
• Correct: 0 / 0 (0%) Reset