[earakely-scale]

What

A website browser loads its start tabs again after a load of cookies or local storage without tabs, so their pages
read what was just set. A start tab already on its start URL's document, as every hash-routed page's is
(https://<site>.test/#/<page>), wasn't loaded again.

• Why: Page.navigate to a URL that differs from where the tab is only in its fragment, or not at all, moves
  within the document. It answers with no loaderId and loads nothing.
• What followed: _StartLoad never settled, so the tab was timed_out after its wait, and the page never read
  the cookies.
• Where it shows:
  • On a browser with its own sites, _refuse_failed then refuses the load.
  • On a website app, the start tab records timed_out, so its readiness goes false.

How

• BrowserControl._start_navigation: a start tab navigates as before. When the browser's answer names no loader,
  the browser has stayed in the document, so the tab then sends Page.reload and the page loads again and reads what
  the load set. The browser decides what counts as the same document, so URLs it spells otherwise are covered too, for
  example https://shop.test#/orders for a tab at https://shop.test/#/orders.
• Waiting for the move: Chrome answers a navigation that stays in the document before the move lands. A reload
  sent at once reloads the route the tab was on (seen on real Chromium). So the tab waits for its main frame's
  Page.navigatedWithinDocument, up to MOVE_WITHIN_S (5 s), before it reloads.
• _StartLoad: a reload's answer names no loader either, so a reloading load takes its main frame's first
  document request or commit after the reload as its document. A reload that fails before it commits (for example
  net::ERR_ABORTED) therefore fails the tab at once, instead of waiting out its time.
• The log: the within-document move is taken as agentenv's, as before. The reload's navigation is also taken as
  agentenv's (Causes.navigates_next), and forgotten if the reload fails or is cut short.
• Causes.navigated_within: it now uses up a pending reload or history move only when it returns that one. A
  reload asked for once the move's answer arrived was otherwise used up by the move's own log entry, so the reload was
  logged as the page's own.

Proof

• Unit tests (packages/agentenv-protocol/tests/test_website_browser.py). The Navigating stub now answers a
  same-document navigation as Chrome does, with no loader and Page.navigatedWithinDocument, and loads on
  Page.reload.
  • test_a_start_tab_already_on_its_pages_document_is_reloaded_there_and_loads_again covers a tab at its URL, at
    another route of its page, and at its URL spelled otherwise. It fails on main, timed_out, and passes
    here.
  • test_a_reload_asked_for_once_a_move_within_a_document_answered_is_left_for_the_navigation_it_makes (in
    test_browser_log.py) fails on main and passes here.
  • test_a_start_tab_whose_reload_fails_before_it_commits_fails_at_once_naming_why checks the failed reload.
  • test_a_start_tab_whose_reload_never_answers_times_out_and_leaves_no_cause_for_a_later_navigation checks that a
    reload cut short by the tab's wait leaves no cause in the log.
  • The stub behaves like Chrome: it answers a move within the document, and a reload, before they land, so the
    tests fail if the reload doesn't wait for the move.
  • test_a_start_tab_on_another_page_of_its_site_navigates_and_isnt_reloaded checks the usual path is unchanged.
• Real Chromium (tst/integration/computer/chromium_app_test.py, on an amd64 host):
  • New: test_a_load_of_cookies_alone_reloads_a_start_tab_on_a_hash_route_so_its_page_reads_them. Chromium sits on
    http://127.0.0.1:8783/app#/trips, a loopback page whose title shows its sid cookie. A load_artifact of
    cookies alone must bring the page back signed in, without waiting out the live timeout. A second load, with the
    tab moved to app#/home, must bring the page back at app#/trips with the new cookie.
  • The data-plane test (check_browser_data) covers the same load on a start URL without a fragment.
  • On this branch, with a venv installed from it, so the computer's image has the fix: both pass. Before the
    wait for the move, the second load reloaded app#/home; that's how the race was found.
  • First run, with the image holding an older protocol from the host's venv (the composer takes the protocol from
    the installed distribution): the new test failed. The page stayed app signed out, which is this bug on real
    Chromium.
• Locally: the protocol package's tests all pass. Of tst/unit, one test fails, a Windows disk recipe test that
  fails the same way on main on macOS. CI runs the full unit suite on Linux.

Notes

• sp#1974: the website apps in synthetic-pipelines had worked around this at launch, by opening their window at
  the start URL plus ?launching. That workaround can stay; this fixes the path it didn't cover.

🤖 Generated with Claude Code

Confidence Score: 5/5

The PR appears safe to merge.

Summary

Reloads start tabs after same-document navigation so pages read newly loaded cookies and local storage.

• Start tabs reload when their URL stays in the same page.
• Browser logs keep the reload cause for the page it loads.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Load start URL] --> B{Navigation names a loader?}
  B -->|Yes| C[Track document load]
  B -->|No| D[Wait for move within document]
  D --> E[Reload page]
  E -->|Request succeeds| C
  E -->|Request fails or is cancelled| F[Clear pending log cause]

Loading

_{Reviews (5) · Last reviewed commit: "fix(browser): a start tab's reload cut s..." · Reviewed by Greptile}