Pimp My IDE / garage dispatch
Back to garage
October 2, 2026 | Agent extensions / authority

A mod is inside the steering column.

Claude Code mods can watch, alter, or take over events inside the agent interface. Review them as behavior-changing code, not as decorative add-ons.

A plugin can package instructions and tools. A mod can also intercept the moment when the agent draws, asks, or acts. That changes the installation question.

The extension boundary moved inward.

Claude Code 2.1.287 added Mods. Its release note says plugins may now modify deeper behavior. The same release added a built-in mod that runs a side agent and flags things it thinks the user or main agent may have missed.[1]

The mod documentation supplies the sharper definition. A mod is a plugin whose JavaScript or TypeScript handlers run inside Claude Code. A handler can watch an event, change it, or take it over. Named events include a tool call, a submitted prompt, and interface drawing.[2]

The right review unit is not the menu item. It is every event the mod can intercept.

List behavior before benefits.

The documentation says a mod can draw panes, replace parts of the existing interface, hold or answer a tool call, route a request to another model, and run a command without a model turn. Its handlers can also share variables inside the file.[2]

Those powers may produce useful controls. They can also change what the operator sees, what the agent sends, and whether a tool runs. A useful install review names each handler, its event, the data it reads, the result it can replace, and the code path that disables it.

Package scope is runtime scope.

The plugin overview says an enabled plugin becomes part of every session, not only the sessions where a person invokes it. Plugin MCP servers run beside enabled sessions, hooks fire on their events, and plugin code runs with the user's permissions.[3]

That makes provenance and update policy part of behavior review. Record the source, exact version, installed scope, update route, loaded surfaces, and ejection command. Repeat the review when the package changes.

Test the disagreement path.

A mod that advises, blocks, reroutes, or redraws needs an adversarial fixture. Feed it a tool call that should pass and one that should stop. Check the original request, the changed request, the visible notice, the final result, and the behavior after the mod is disabled.

  1. Inventory every event handler and direct command.
  2. Run allow, block, error, and disable fixtures.
  3. Compare the original event with the event that continued.
  4. Restart a clean session and prove the mod is absent.

The fuseboard below drafts that review. It does not inspect, install, enable, test, or approve a mod.

Interactive makeover / mod authority fuseboard

Wire the review boundary.

This replaces a vague plugin checklist. Choose the authority you intend to permit, select the circuits the review must cover, and copy a test sheet with the real evidence fields still open.

Review controls

These controls write a test plan. They do not configure Claude Code or grant a mod permission.

Maximum intended authority
Circuits to include
Behavior backplane

No circuits selected

0 of 4
Observe ceiling

The review asks for read-only observation. The fixture must still prove that no event or result changed.

Test sheet incomplete.

No behavior circuit is selected.

A lit circuit means its section appears in the test sheet. It does not mean a handler was found, the code stayed within that boundary, a test passed, or the mod was disabled.

Mod authority test sheet

The completed control state means four sections are selected. Every bracketed source, handler, event, output, and disable field still needs real evidence.

Sources read

Source log and evidence boundary
  1. Anthropic, Claude Code v2.1.287 release notes, published October 1 and read October 2, 2026. This supplies the release claim that Mods let plugins modify deeper behavior and the description of the built-in side-agent mod.
  2. Claude Code documentation, "Mods overview", read October 2, 2026. This supplies the in-process handler model, named event and interface powers, direct command behavior, shared state, trust guidance, and enable or disable concepts.
  3. Claude Code documentation, "Plugins overview", read October 2, 2026. This supplies plugin composition, all-session scope, process behavior, permissions, marketplace delivery, and footprint guidance.
  4. Hacker News top stories, read October 2, 2026. We reviewed the current front page for the wider tool and agent signal. It did not supply claims about Mods.

Evidence boundary. We read the release notes and documentation. We did not enable a mod, inspect the source of the built-in mod, capture handler events, compare requests, or run a disable drill. The fuseboard produces a review template only.