My/AI/Agent/Fleet/Browser

• Fleet — Browser

  • Browser-shaped work has two routes, and they fail for different reasons. Claude in Chrome is the first route. Playwright is what is left when it is genuinely unavailable.
  • Stated by Person/codekiln twice on 2026-08-24 Mon: Claude in Chrome by default, and Playwright only for a harness that is not Claude — CursorAI or Codex. That is narrower than the line above it, and the difference is the part to keep. It makes the reason to reach for Playwright the harness doing the work, rather than a seat's judgement that Chrome was unavailable — and that judgement is exactly what went wrong twice the same day, once as four classifier refusals recorded as a missing capability and once as a route's constraint reported as a blocked source. A seat running under Claude has no Playwright case to argue.
  • A Playwright browser left running from 09:16 was closed at the noon wrap by the seat that found it. A browser nobody is driving is a stale worker like any other.

  • Claude in Chrome, first

    • The worked shape:
      • sh

        claude --chrome -n '<name>' --permission-mode auto '<brief-pointer>'

    • It reaches an SSO-gated source with no login handshake, no debugging port, and no disturbance to the running browser. Demonstrated 2026/08/24 by reading a protected Splunk index directly.
    • A report that fell back to Playwright says so, and says which route it tested. "The browser was unavailable" describes two different situations with different fixes.

  • Which profile it attaches to is not settled

    • A session that answers proves that a session works. It does not establish which profile that session is using, and the question has not been settled either way. Record it as open rather than assuming it is the profile below.

  • The Playwright route, and what actually blocks it

    • A Chrome profile directory opens in one process at a time. Person/codekiln's own browser holds Default, and no --remote-debugging-port is open on it, so this route does need him to quit or relaunch.
    • That is a fact about the route, not about the source. On 2026/08/24 two seats measured this constraint and reported the source as blocked on him. It was the route. The correction cost a request for his attention that had to be withdrawn.
    • Profile directories and profile labels are different things, and the labels do not identify these profiles. Read from ~/Library/Application Support/Google/Chrome/Local State, checked 2026/08/24:

      • Directory	Label	Use
        Default	Person 1	the live sessions — use this
        Profile 9	names another account	leave alone
        Profile 3	names another account	leave alone

    • Default is the last_used profile and the one holding the live sessions. Its label is Person 1 — Chrome's placeholder for a profile that was never renamed. The profile that matters is therefore the one whose label says least about it; the other two carry labels naming separate accounts, and those are the two to leave alone.

  • An unauthenticated profile renders a blocked page as empty

    • A fresh or isolated browser profile carries none of the signed-in sessions. An SSO-gated page opened in one does not report a refusal — it renders as though the content is not there.
    • The failure therefore arrives as emptiness rather than as an error: no exception, no status code worth noticing, no prompt. An agent reading that result concludes the search found nothing, and reports silence.
    • This is the shape My/AI/Agent/Chief of Staff records under a check that reports a break where none exists. Well-formed output carries no signal that it is a fragment, and a confident empty result earns no second look.
    • So before reporting that a page held nothing, establish that the client was signed in. An empty result from an unauthenticated client is not a measurement.

  • A capability recorded as absent decays like any other dated claim

    • On 2026/08/24 a seat met four classifier denials on claude --chrome and recorded the capability as unavailable. After bedding down, the identical command worked on the first try.
    • So bedding down is itself a retry — My/AI/Agent/Fleet/Bed Down. A handoff doc carrying "X is unavailable" hands the next window a conclusion it cannot distinguish from a permanent fact, and the next window has no reason to test it.
    • An absence is a measurement with a timestamp, the same as a count. Write it as one: what was tried, when, and what the failure looked like, so a successor can tell a standing limitation from a transient refusal.

  • Working in someone else's live window

    • One browser-driven task at a time across the whole fleet, and the seat taking it says so.
    • The window is in use. Tabs that an agent did not open stay open, accounts stay signed in, and a tab he is working in is not navigated.