frontend-patterns/research/reviews/NavigationFocusExperiment.md

Keyboard continuation and navigation feedback — 2026-09-27

LEARN-0012; Codex author-model review. Preregistered protocol committed as 2b1c9a4 before implementation/execution. New stream-F comparison group F-navigation-focus-v1, not comparable to the preceding HTTP/cache question's group. Runner, authored policy, receipt.

Evidence and result

Pinned htmx 2.0.11, Playwright 1.63.0; same binary versions/hashes as the previous navigation cycle: Chromium 151.0.7922.34 and Firefox 155.0. One run, no harness corrections or discarded failures. 22/22 registered rows completed: two native cases, ten raw observation cases, ten managed-policy cases. A passing raw observation row means it was recorded successfully, not that it satisfies the managed obligations. All browser actions initiating navigation use Tab/Enter; background update is explicitly script-triggered and browser history uses Playwright's history command.

Claim	Cases	Observations in both browsers
NF01	K01	With scripts disabled, Tab reaches the native link and Enter loads beta with its full-document title
NF02	K02	Raw htmx retains focus on the new same-ID next link, but this fragment response leaves alpha's title over beta content and produces no completion status. Managed policy focuses beta's heading, updates title and writes Loaded beta
NF03	K03	Enhanced 503 preserves alpha URL/content and the invoking link's focus in both variants. Only managed policy writes the failure/recovery message; native recovery works by keyboard in both
NF04	K04	While a response is held at the server, the user tabs to persistent search. Both variants preserve search focus on release; managed policy reports completion without moving focus
NF05	K05	History restores alpha. Raw focus is the same-ID link; managed focus is the heading with matching title/status
NF06	K06	An unrelated background region update preserves search focus and does not emit navigation feedback in either variant

The raw result narrows our anticipation of focus loss: stable IDs let htmx restore focus to a replacement element. Here that link now leads back to the other article. DOM identity retention alone is not proof of an appropriate next task, but neither is this observation proof that retaining the link is always wrong. The server deliberately returns a fragment without title metadata; this is an application contract gap, not a claim that htmx cannot update document titles.

Mechanism and interpretation

The authored policy remembers the request source, checks focus ownership immediately before a zero-delay swap, and applies the continuation synchronously on afterSwap, after the library's own focus restoration. It updates heading/title/status only for registered content navigation. It leaves a surviving search control alone if focus has moved there. Errors keep focus on the invoker and provide a visible polite status message with a native recovery action already present. History uses a separate explicit heading policy. It is a small experiment, not a production navigation library.

W3C focus-order guidance allows more than one meaningful focus order and does not prescribe one universal target. Status-message guidance informs making feedback programmatically available without requiring focus movement. htmx events supplies the lifecycle hooks. Living primary pages accessed 2026-09-27; exact documentation revisions unknown. Package implementation is pinned and hashed. DOM status/ARIA observations are not evidence that a screen reader spoke the message; no assistive technology or human evaluation was run.

Disposition and limits

Corroboration, one bounded outcome. Existing Progressive Enhancement already calls for focus, feedback and recovery semantics; Focus Return explicitly excludes automatic generalization from modal closure to route transitions. The experiment makes that boundary concrete and adds an ownership-sensitive candidate mechanism. No newly invented pattern, canonical promotion, independent recurrence or usefulness gain.

Add scoped implementation guidance: distinguish successful route continuation, failed navigation, history and background updates. Test whether a late response still owns focus before moving it. Returning to a modal invoker and continuing into a new page are different decisions. Keep both candidates as hypotheses.

Unverified: screen-reader speech (including possible heading/status duplication), focus visibility, complete names/order audit, actual usability, pointer modality, nonzero swap delays, overlapping/superseded route requests, delayed history responses while the user moves focus, script-late activation, interrupted transport and offline availability. The explicit history policy always focuses a heading in the tested immediate restore; it is not proven safe for an asynchronous history miss. Public single-response success does not settle these ownership races. No writes or production application were exercised.

Effort receipt separates monotonic active agent time from experiment/check subprocess time; final administrative refresh/commit/snapshot/sync overhead excluded. No token/cost estimate. One cycle in this new comparison group cannot establish a trend, even though the ledger now has two prospective cycles in different groups.

WP-0003 T01 retains delayed history restoration/overlapping navigation ownership and interrupted enhancement. WP-0003 T02 retains independent human/AT validation and the offline branch. WP-0006 T03 retains comparable-window and remaining review-date work.