Login retry after a 2FA timeout waits for the closed Chrome, then opens a new one

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Context.md

@@ -97,7 +97,8 @@ of 23–26/10 changed no other result.
 **2026-10-27 terminal runs:** 16:40 — 2FA mail never came in 10 min (eCW
 timed the approval out), then attempt 2 crashed: the run closed the Chrome
 it opened and re-attached to 9222 while that Chrome was still closing
-(`connect_over_cdp: Target page ... has been closed`; §10). 16:53 (after a
+(`connect_over_cdp: Target page ... has been closed`; fixed in `a89f689`,
+§10). 16:53 (after a
 120 s guard cooldown) — login 220 s, 2FA mail 42 s, **50/50, 38 Pass / 12
 Fail / 0 Review** (the 50-claim sheet; same as 26/10 17:40). 17:30 with
 `Mecor Ready 2 - 10262026.xlsx` (all claims, QUIET_BOXES, ECW_KEEP_OPEN=0)
@@ -213,7 +214,7 @@ Get-ChildItem tests\test_*.py | ForEach-Object { "$($_.Name): $((python $_.FullN
 - Tests run offline against fake HTML pages in `tests/fixtures/` and fake DB
   connections; they never touch eCW or SQL Server (`test_cdp_browser.py`
   starts a local Chromium on a debug port). As of 2026-10-27: 71 test
-  files, **1239 tests** (full run 2026-10-27: 1239/1239) (earlier flaky, each
+  files, **1239 tests** (full run 2026-10-27 after `a89f689`: 1239/1239) (earlier flaky, each
   failed once and passed on rerun: test_verify_no_live_writes,
   depression_dx_pointer, progress_note_wait — watch them) (about 5 minutes for all of them).
 
@@ -649,7 +650,9 @@ the Portal (restarted 2026-10-26, [portal] terminal tab).
 
 History: `git log --oneline`.
 
-Last code commits: `1da9e7c` (2026-10-27, Portal sign-in with the Ecare
+Last code commits: `a89f689` (2026-10-27, login retry after a 2FA timeout
+waits for the closed Chrome and opens a new one — §10), `1da9e7c`
+(2026-10-27, Portal sign-in with the Ecare
 office login — §2), `cc3a01a` (2026-10-27, Portal container on India time —
 §10), `69c51d8` (2026-10-27, Dockerfile: libgssapi-krb5-2 for ODBC Driver
 18 — §10), `70ba709` (2026-10-27, Portal server mode for Dokploy — §10), `f99dc17` (2026-10-27, Portal roles Super Admin /
@@ -787,11 +790,18 @@ test claim ids — new posts must not (§7).
   watchdog came after them and has not fired live yet. Suggested, not
   ticketed: a landing-page timeout with one retry, a 2FA retry, a week of
   daily runs to measure login times on the Runs page.
-- **Login retry bug (seen 2026-10-27 16:50, not fixed).** After a 2FA
+- **Login retry bug (seen 2026-10-27 16:50, fixed in `a89f689`, not yet
-  timeout the run closes the Chrome it opened and attempt 2 attaches to
+  seen live).** After a 2FA timeout the run closed the Chrome it opened and
-  9222 at once, while that Chrome is still closing → `connect_over_cdp:
+  attempt 2 attached to 9222 at once, while that Chrome was still closing →
-  Target page, context or browser has been closed`, run over. Fix: wait
+  `connect_over_cdp: Target page, context or browser has been closed`, run
-  until 9222 stops answering, then open a fresh Chrome (test-first).
+  over. Two causes: `Browser.close` only starts Chrome's shutdown, and the
+  "no Chrome → open it" step ran once, before the login loop. Now the
+  attached browser's `close()` waits (up to 30 s) until the port stops
+  answering, and `launch()` opens a new Chrome when this run opened one and
+  closed it (same extra args; the eCW proxy comes from `chrome_command`).
+  A Chrome the run did not open is still never opened or closed. The
+  "Chrome open" box is not shown again on the retry. Test:
+  `test_cdp_browser.py` (close → port quiet → launch again → close).
 - **Gitea:** one overall snapshot update at the end of a day, when the
   operator asks (from the clone `D:\ecw-charge-validation`; operator
   pushes; 2026-10-26: `ba15a57` → `120fd68`).

src/ecw_bot/core/browser.py

@@ -30,7 +30,8 @@ only), so the run can be watched with the Claude extension in that Chrome.
 No Chrome there when the run starts → the bot opens it (`start_chrome`,
 operator 2026-10-22: Start on the Portal is all it should take), and closes
 it again at the end (`CDP_CLOSE_AFTER_RUN`) — never a Chrome it did not
-open. The bot opens its own tab, clears eCW's cookies and storage when it
+open. A login retry after a 2FA timeout waits for that Chrome to be gone and
+opens a new one. The bot opens its own tab, clears eCW's cookies and storage when it
 attaches and again when it is done, and closes only its own tabs.
 """
 
@@ -72,7 +73,12 @@ def launch(pw, incognito: bool = False):
     a clean, cookie-less browser is the point of that retry.
     """
     if cfg.BROWSER_CDP_URL and not incognito:
-        return _attach(pw, cfg.BROWSER_CDP_URL)
+        url = cfg.BROWSER_CDP_URL
+        # This run opened Chrome and closed it (a 2FA timeout): open a new
+        # one. Never one it did not open — that is the operator's to start.
+        if url in _CLOSED and not start_chrome(url, _CLOSED.pop(url)):
+            raise RuntimeError(f"Could not open Chrome again at {url}")
+        return _attach(pw, url)
 
     args = list(CHROME_ARGS)
     if incognito:
@@ -111,8 +117,12 @@ def launch(pw, incognito: bool = False):
 PROXY_BYPASS = "claude.ai;*.claude.ai;*.anthropic.com"
 
 # Debug-port URLs of the Chromes this process opened — the only ones it
-# closes when it is done.
+# closes when it is done — with the extra arguments they were opened with.
-_STARTED = set()
+_STARTED = {}
+
+# The ones it opened and has closed again: a login retry (2FA timed out)
+# opens them afresh instead of attaching to nothing (2026-10-27 16:50).
+_CLOSED = {}
 
 # Started through a Python that exits at once: Chrome's parent is then
 # gone, so a Stop on the Portal (taskkill /T on the bot) never takes Chrome
@@ -204,7 +214,7 @@ def start_chrome(url: str, extra_args=()) -> bool:
                      stderr=subprocess.DEVNULL)
     for _ in range(60):
         if cdp_ready(url):
-            _STARTED.add(url)
+            _STARTED[url] = list(extra_args)
             print(f"[BROWSER] Opened Chrome at {url} "
                   f"(profile {cfg.CDP_PROFILE_DIR})")
             return True
@@ -225,6 +235,16 @@ def cdp_ready(url: str) -> bool:
         return False
 
 
+def _gone(url: str, seconds: float = 30) -> bool:
+    """Wait until nothing answers at `url`. False if something still does."""
+    end = time.monotonic() + seconds
+    while cdp_ready(url):
+        if time.monotonic() > end:
+            return False
+        time.sleep(0.25)
+    return True
+
+
 def ecw_sites(urls) -> set:
     """eCW's registrable domains ("ecwcloud.com"), from its configured URLs.
 
@@ -297,11 +317,16 @@ class _AttachedBrowser:
                 print("[BROWSER] Closed the Chrome this run opened")
             except Exception:
                 pass                    # the connection goes with it
-            _STARTED.discard(self._url)
+            _CLOSED[self._url] = _STARTED.pop(self._url)
             try:
                 self._browser.close()
             except Exception:
                 pass
+            # Browser.close only starts the shutdown; the port answers for
+            # a while yet. Attaching then gets a Chrome that dies under it.
+            if not _gone(self._url):
+                print(f"[BROWSER] The Chrome this run opened still answers "
+                      f"at {self._url}")
             return
         try:
             mine = [p for p in self._context.pages if p not in self._theirs]

tests/test_cdp_browser.py

@@ -354,7 +354,17 @@ def _run_bot_starts_chrome() -> int:
                     page.set_content("<p>x</p>")
                     b.close()
                     if close_after:
-                        assert _gone(url), "Chrome still answers"
+                        # close() returns once Chrome is gone, not while it
+                        # is still shutting down.
+                        assert not browser.cdp_ready(url), \
+                            "Chrome still answers after close()"
+                        # A login retry (2FA timed out) launches again: a
+                        # fresh Chrome, not the dying one (2026-10-27 16:50).
+                        b, ctx, page = browser.launch(pw)
+                        page.set_content("<p>again</p>")
+                        b.close()
+                        assert not browser.cdp_ready(url), \
+                            "the reopened Chrome still answers"
                     else:
                         time.sleep(1)
                         assert browser.cdp_ready(url), "Chrome was closed"