Doing much blog work previously does not mean a post requested from a cafe should go to a blog today. I separated the request-time writing destination from accumulated workflow experience. A familiar procedure must not override the user's current intention.

More experience can strengthen the wrong default

A successful blog skill or program can accelerate the next run. Applied to a different destination, however, that useful experience becomes a misleading shortcut. A generic writing request should not automatically send work to the most familiar editor.

I treated this as more than a wording repair. The design needed to connect destination selection, program recommendation and the check immediately before input. The same question has to reach both descriptive context and execution boundaries, preventing old conversation material from silently determining the final click.

Do not merge cafe and blog into one service bucket

Site keys distinguish cafe and blog destinations. Desktop and mobile cafe surfaces can belong to the same workflow while remaining separate from a blog. With a writing request and bound target, target checking compares the candidate with the initial or explicitly requested destination.

This is not a universal policy that blocks every possible string. Missing targets and unresolved site keys have separate assumptions, while account, permission and editor checks live elsewhere. The function narrows a particular leakage boundary when execution has a request target; it is not the complete authorization system.

Apply the same boundary to saved-program suggestions

Program filtering uses declared sites or a start destination to exclude unrelated definitions. A generic cafe-writing request should not automatically recommend an old blog procedure. A plausible program name does not make its different starting location appropriate for this work.

The filter cannot infer every internal step of a definition without site metadata. Inspection and current-input review remain necessary. Reducing an unsuitable suggestion differs from guaranteeing rejection under every possible definition. Missing metadata calls for more inspection, not unsupported confidence.

Reading elsewhere differs from writing elsewhere

Research on another site need not be blocked. Material for a cafe post may be read from a blog. Conversely, an explicit blog or multi-destination writing request should be honored. Defaulting to the current tab does not justify ignoring a declared destination.

Tests distinguish references, prohibitions, explicit destinations and multiple destinations. A site word inside requested content cannot automatically become a target. Readable material and the writing location have different roles even when both appear in one tab inventory.

Retain the exact scope of the historical live test

The version-0.1.77 cafe test read the current tab and returned a draft in conversation; it did not input, publish or schedule. The original problem run could not be identified in the same database evidence. That record is neither a complete reproduction of the initial failure nor a successful cafe publication.

Current external MCP has no embedded CLI. Historical natural-language routing tests cannot establish an internal model interpreting requests today. The caller must read current context and explicitly bind the actual destination. This account compares the earlier repair, retained contracts and tests rather than reporting new live publication.

Keep company defaults subordinate to the current request

Customers have familiar forms and frequently used accounts. Remembering them is useful, but a changed request must override the default. A system giving old experience more authority than a new instruction can misroute work in the name of convenience.

I connected current intent, selected destination, recommended procedure and actual input checks. The linked context module offers a narrower synthetic separation of destination and references without accounts. It is not the product's complete language-routing behavior, but exposes trusted inputs and missing choices for inspection.