Day 87 of 92 · Tuesday · Learning

Mods - a plugin can now watch your session (Claude Code 2.1.287)Capability

Claude Code 2.1.287 added mods: a plugin can now change what the terminal shows and does, and one built-in mod puts a side agent on your session to flag what you and Claude miss. Ten minutes to switch it on and watch one note arrive.

Catch-up progress
87/92
Why this matters to you

On 10-1 a security pass on the Scott Memo found three preview URLs and 86 old deployments serving meeting content without a login, plus a pay-detail leak in the snapshot. Every one of those was caught by a review you asked for. A watcher that reads along and says 'you should know' is that review, unasked. D2-B on the 10-4 Weekly Review asks whether to try it for a week.

Until now a plugin could add three kinds of thing: skills (instructions Claude reads), hooks (scripts that run at events such as a tool call), and MCP servers (tools). A mod is a fourth kind: TypeScript that runs inside Claude Code itself. It can draw a pane, a band above the prompt, a status line or a toast; it can intercept what a tool is about to do; it can replace a feature. Mods ship inside plugins, hot-reload while you work, and your skills list already carries the authoring skill for them (plugin-authoring: 'Make a mod: a live pane, band, status line, toast or hook').

'You should know' is the first built-in mod. A second, smaller agent reads your session alongside the main one and posts short notes when something looks off: a file about to be overwritten, a claim that was not verified, a risk nobody mentioned. It flags; it never blocks, edits or approves. Two limits matter for you: it runs only in first-party interactive sessions with telemetry on (so your headless plan-day, nightly and weekly runs never see it), and its tokens land on the same usage meter those headless jobs share — the meter that capped out on 9-11.

The bloat axis the weekly review weighs: a skill costs context every turn, a hook costs a script run every event, and a mod costs runtime for as long as it is enabled. That is why the 10-4 doc frames it as a one-week trial with a named off switch, not an install.

Worked example

Three commands: confirm the version, see what is installed, and turn the watcher on inside a session.

claude --version                 # 2.1.289 today; mods arrived in 2.1.287
claude plugin list | head -25    # what is installed and enabled right now

# inside an interactive session (not claude -p):
/plugin enable cc-plugin-you-should-know@builtin
# ... do one small real task, read any 'You should know' note ...
/plugin disable cc-plugin-you-should-know@builtin   # the off switch
▶ Do it now
  1. Run `claude --version` and `claude plugin list`; confirm 2.1.289 and note that no 'you-should-know' row exists yet.
  2. Open a session in /Users/tom/Claude, run `/plugin enable cc-plugin-you-should-know@builtin`, then give Claude one real, small task: 'list the FOR REVIEW documents on the device older than seven days'.
  3. Watch for a 'You should know' note during or after the task. Write on D2-B of the 10-4 Weekly Review whether that one note earned its tokens; disable it with the same command if you want it off before deciding.

Gotchas

Go deeper: Claude Code CHANGELOG — 2.1.287 (mods, You should know) · Claude product announcements
One-card takeaway

A mod changes the harness, not the prompt — treat enabling one like installing software, not like adding a line to CLAUDE.md.