Many open tabs do not require every body in the conversation. A title, destination, project and account clue can be enough to locate a writing surface. I preferred metadata for initial connection and actual content reading only where the task needed it.
Narrow the first question to locating the tab
Finding the blog editor differs from understanding its manuscript. Collecting every body first enlarges context and mixes unrelated documents into the task. Separating location metadata from decision material makes the reason for each subsequent read explicit.
Routing uses tab IDs, titles, profiles and reduced destinations. The MCP tab inventory is also a different entry from body observation. Reading that inventory cannot establish the actual account or editor state. Metadata locates a candidate; it does not replace understanding the page.
Navigation comparison and descriptive metadata differ
Route metadata uses full destinations locally while retaining origin and pathname for its descriptive role. It deliberately omits session queries and fragments and bounds title length. The purpose is separating useful identity from incidental values.
That function does not define every API's destination-output policy. Context inventory and other tools need their own contract review. Paths and titles can themselves contain sensitive clues, so public material still needs separate inspection. Removing a query is not complete anonymization.
Read the actual page when metadata is insufficient
Metadata may not distinguish several boards or documents with equal titles. The next read should inspect that candidate's current structure, account and relevant body. A compact initial reply cannot justify typing into unresolved state.
Search, extraction and bounded observations provide more detail on large pages. A returned fragment is still not the entire body. I care about explaining what remains missing and why another read is needed. Metadata-first organizes reading order rather than eliminating necessary reading.
Distinguish preparation from collecting every open body
Page preparation separates automatic active/reference-tab preparation from explicit reads and applies dormant, loading and window-visibility conditions. Entry count, bytes and age are bounded. It is not a continuous transfer of every open body's content into conversation.
Local draft recovery is another responsibility. Collecting drafts for shutdown differs from supplying research text to an AI. Their purposes and disclosure scopes differ even inside one app. The presence of both does not establish that every local stored value enters model context automatically.
Separate historical context assembly from external calls
Earlier embedded-CLI documentation describes supplying a default page and bounded references to a model. Current external MCP returns inventories and operation results without an internal model. Historical automatic-attachment wording cannot be applied unchanged to every external conversation.
I read role metadata, current context responses and preparation code without collecting every current body. This article also measures no new token-saving percentage. A structure reducing supplied bytes and actual model cost need distinct evidence rather than one assumed number.
Build a workflow that opens the needed material
Opening every company document at once is not always convenient. Locating the relevant account and document before inspecting it can be more intuitive. Agents also need an interface separating discovery from detailed reading.
I treat metadata as a thin guide connecting the current task to its material. Correct location and justified reading matter more than a large collection count. That distinction helps manage context growth and unrelated-data exposure as the number of tabs increases.