You Should Know

You Should Know

The plugin dcc-plugin-you-should-know shows one thing that you should know before Claude changes something. It shows the thing as a toast, and the change runs as usual. It replaces the built-in plugin cc-plugin-you-should-know, which shows its notice after Claude acts. It works without the core plugin, and it imports nothing from another plugin.

The plugin is opt-in. The setup of dotclaude does not turn it on, because each check costs a model call.

The source is in src/plugins/dcc-plugin-you-should-know/. The one slice is notice, and its hooks are in slices/notice/adapters/mod.ts.

Install

1. Install the plugin with /plugin install dcc-plugin-you-should-know@dotclaude.
2. Turn off the built-in plugin cc-plugin-you-should-know@builtin in /plugin.

When It Checks

The plugin checks the first call of each turn that changes something: an Edit, a Write, a NotebookEdit, or a Bash command that is not a read.

These calls get no check:

• a read, such as Read or cat a.txt
• the second and each later changing call of the same turn
• a call of a subagent
• a call that the permission rules deny

The model gives one topic of three kinds, or none: a trade-off that you possibly do not see, a design decision that has effects later, or an edge case that is not clear. Most checks give none, and then no toast shows. A topic that showed in the session does not show again.

In a TUI session, the toast showed 23 ms before Claude Code wrote the file. So the notice shows before the action that it is about.

What a Check Costs

The plugin makes at most one model call per turn. The tool waits for the call, so a check adds its wait before the tool runs.

Each check is a separate call to Haiku, with your prompt and the pending action, and no transcript. The output is one sentence or NONE, about 4 to 50 tokens. The sandbox sessions measured these checks:

Account	Wait of One Check	Input Tokens of One Check
Subscription	2,216 ms and 2,430 ms	412 of Haiku, against your usage limits
API key	3,017 ms	411 of Haiku

The tool waits at most 5 seconds for a separate call and at most 6 seconds for a fork. Each check sends the newest 10 topics that showed, so the input stays bounded. A failed or slow call shows no toast, and the change runs.

When CLAUDE_CODE_ATTRIBUTION_HEADER Is 0

On a subscription, a separate call from a hook gets HTTP 429 when CLAUDE_CODE_ATTRIBUTION_HEADER is 0. The request then carries no billing block and no identity block of Claude Code, so the API refuses it.

After a 429, the plugin uses a fork of the main conversation for the rest of the session. A fork sends the main request again, so it works. The fork has no output bound and no cancel. $.model.fork takes only { prompt }, so the plugin cannot set a token limit or a timeout on it. The plugin stops waiting after 6 seconds, but it cannot stop the request. The sandbox session measured this path:

• The first check waited 2,850 ms, because Claude Code tried the separate call 3 times before the 429 stood.
• The fork read 16,731 input tokens on the main model, nearly all of them a cache read of the conversation.

So a fork costs much more than a separate call. Keep the attribution header on to get the cheaper check. The core plugin does not set CLAUDE_CODE_ATTRIBUTION_HEADER=0.

Limits

• The toast is for you only, and Claude does not see it.
• The cost of a session does not show the cost of its checks, because the cache state of the main requests differs between runs.
• Each number above comes from one or two sandbox sessions with claude-sonnet-5-5 and one short edit.

The evidence is in .dotclaude/spec/changes/1.0.0/evidence/you-should-know.md.

Related Pages

• AGENTS.md: the other plugin that replaces a built-in plugin.
• Add-On Plugins: the list of the add-on plugins.