v3.50.1

Changed

• A replay whose same-name population merely changed SIZE no longer stops when
  the recorded element can be positively identified by its container, which
  turns some of the previous refusals back into successful runs. (#1185)

• browser_repl snapshots return a diff again, with the refs still
  complete. A browser.snapshot() or browser.smart_snapshot() inside a
  script used to be forced to render the whole tree, because the script's
  refs[] was parsed out of the returned text and a diff names only what
  moved — so an act-then-verify loop paid for a full listing on every
  iteration (a Node.js API page cost about 78 KB and 7–8 seconds each time,
  where the direct tool answered with a few lines). The handler now hands the
  bridge the complete listing alongside whatever it returned, so the script's
  text is the ordinary diff-or-full the direct tools give — its first line
  still says [snapshot: full] or [snapshot: diff …] — while refs[] keeps
  every element on the page. full: true still forces the whole tree, and the
  entries are unchanged: ref is a string for snapshot, a number for
  smart_snapshot. (#1188)

• A touch preset now dispatches touch. Under browser_emulate {device: "Pixel 7"} the page reported a touchscreen — maxTouchPoints: 5,
  (pointer: coarse) — and then received mouse input, so anything gated on
  touchstart, or reading pointerdown.pointerType, saw a desktop pressing
  the button. browser_click now sends a real touchStart/touchEnd pair on
  the session the preset is already held open on, and browser_drag sends
  touchStart → moves → touchEnd. Measured on the same fixture and click:
  before, pointerdown: mouse and no touch events at all; after,
  pointerdown: touch, touchstart, touchend, then the click Chrome
  derives from them, all isTrusted. The element is still checked for
  actionability exactly as a mouse click checks it, and the approach the mouse
  walks is skipped — a touchscreen has no pointer resting on the page between
  gestures. Two gestures a touchscreen does not have stay on the mouse and say
  so in the result: a double click, and browser_hover, which would otherwise
  turn every hover-revealed menu into a silent failure. Nothing changes without
  a device preset. (#1190)

Fixed

• A replay stops when the number of same-name elements changed, instead of
  clicking by position anyway. A step recorded against the first
  button "Submit order" on a page used to replay against whatever sat at that
  position later, on the reasoning that the first element cannot have been
  displaced. It can: a look-alike with the same accessible name inserted above
  the real one takes over position one, and the run clicked the look-alike,
  never fired the real submit, and still reported the step as ok. A live
  population whose size differs from the recorded one now stops the run at that
  step — at every position, the first included — saying which way the count
  moved. The count is also re-checked against the live page at the moment of
  acting, so an element that appears between the internal snapshot and the click
  is caught too. A stop of this kind no longer counts toward the failure streak
  that retires a flow, since a page growing a banner says nothing about whether
  the recording is still good. The page is left where it stopped, so the
  recovery is unchanged: take a snapshot, finish the flow live, and save under
  the same name to re-record the healed path. A change that leaves the count
  unchanged — one look-alike added above, one element removed below — is still
  not detectable from what a recording stores. (#1179)

• browser_repl snapshot refs are the type their tool wants. A snapshot
  inside a script handed back refs[i].ref as a number, but every browser tool
  types ref as a string — so the tool's own documented usage, passing
  refs[i].ref as the arg named refs[i].param, was rejected as "expected
  string, received number" before the click ever happened. Refs now come back
  as strings where the argument is ref, and stay numbers for smartRef. Note
  for existing snippets: a snapshot-lane refs[i].ref is now a string, so a
  comparison written as r.ref === 3 has to compare '3'. (#1180)

• browser_repl shows the hints its calls carried. A page with a promoted
  flow advertises it to a direct browser_navigate call — [skill] fixture-submit — proven 4-step flow — but the same call inside a script said
  nothing, exactly where a proven flow saves the most steps. A run now collects
  those [replay]/[skill] hints across its calls, dedupes them, and prints a
  --- hints --- block beside --- calls --- and --- console ---. (#1180)

• A navigation that carries a Referer no longer contradicts it. When the
  agent left one page for another, the Referer header said a link had been
  followed while Chromium's own Sec-Fetch-Site: none said the URL was typed
  into the address bar — a pair no real browser sends. With a referer to send,
  the navigation now happens from inside the page, so the whole header set
  agrees; without one, page.goto is used as before. (#1181)

• browser_emulate device presets apply the whole device. Choosing
  iPhone 13 set the User-Agent and viewport and stopped there, leaving no
  touch points, a desktop pixel ratio and desktop viewport semantics behind the
  mobile UA. The preset now carries device metrics, screen size, orientation
  and touch on both backends, holds them across navigations, and device: null
  clears all of it. One limit the result now states outright: a non-Chromium
  preset such as an iPhone cannot fill navigator.userAgentData or the
  Sec-CH-UA* headers, which keep answering as Chromium — a Chrome-based
  preset gives a consistent mobile identity. (#1181)

• Humanlike typing holds each key for a human interval. Keys were pressed
  and released within a millisecond; the dwell is now drawn from the same
  log-normal rhythm as the inter-key delay (median 70 ms). (#1181)

• wmux --help now shows mcp register --profile. The core profile
  shipped in 3.50.0 but the help text still listed mcp register bare, so the
  flag could only be found in the changelog. The row names it and says what
  full and core mean. (#1184)

• browser_replay now records the accessible name of the nearest named
  container each element sat in, and checks it before acting. A flow whose
  page was restructured underneath it — a look-alike inserted above the
  recorded element and another removed below, which leaves the same-name count
  unchanged and used to resolve silently by position — now stops at that step
  and hands the page back live instead of clicking a stranger. Recordings made
  before this release keep replaying under the previous rules. (#1185)

• A device preset now holds navigator.platform. Under browser_emulate {device: "Pixel 7"} the page answered MacIntel from the next tool call
  onward, while its User-Agent, Client Hints, pixel ratio and touch points all
  said Android. Of everything a preset installs, the platform is the one value
  Chromium keeps as a page-wide setting rather than as state owned by the
  session that wrote it — and every debugger session that detaches from the
  page clears that setting, including a session that never set a platform of
  its own. wmux opens and closes short-lived sessions on the page while
  resolving which tab a tool is talking to, so the preset was being undone by
  wmux's own bookkeeping. The override is now re-sent once the page has been
  resolved, so navigator.platform agrees with the rest of the identity across
  navigations and tool calls, and device: null still returns the host's.
  (#1186)

• wmux mcp check now says when it is speaking for an isolated instance.
  An instance launched with WMUX_DATA_SUFFIX never registers itself in the
  agent configs, but check still listed ~/.claude.json, ~/.codex/config.toml
  and ~/.gemini/settings.json with no hint of that — so the production entry
  read like this instance's own, which is exactly the confusion that made the
  original hijack hard to spot. check now prints a heading naming the suffix
  and saying the rows below belong to your production wmux (--json gains a
  matching isolated field), and the register / unregister warnings read
  the shared dataSuffix() helper instead of the env var directly, so every
  subcommand agrees with the rest of the suffix-aware paths. (#1187)

• A replay can now tell two identical siblings apart by the attribute the
  page gives them. The neighbourhood check added in the previous release
  abstains on the one shape it was built around: two elements with the same
  role, the same accessible name, and the same container are under the same
  neighbourhood, so it has nothing to say about them and the count rules decide
  by position. Real pages almost always distinguish those siblings on the
  element itself — a data-testid, an id, a name, or an aria-label — and
  a recorded step now carries whichever of those the element had, as a second
  verify-only signal. Where the neighbourhood gives no verdict, the sibling
  carrying the recorded attribute wins if it is unique and still at the
  recorded position; if it moved, or if the population carries attributes but
  not that one, the run stops and hands the page back live instead of clicking
  the twin. Nothing is ever located by the attribute, and a population that
  carries none at all abstains exactly as before, so no existing flow changes
  behaviour. Both recording lanes — browser_snapshot and
  browser_smart_snapshot — mint the value from the same pass, so a flow
  recorded on one replays against the other. Traces recorded before this
  release replay unchanged. The irreducible case is now genuinely irreducible:
  identical siblings that carry no identifying attribute at all.

• A flow recorded from browser_smart_snapshot no longer stops on a page
  that has not changed. A replay refuses to act when the number of
  same-named elements differs from the recording, which is right — but the two
  snapshot tools do not walk a page the same way: browser_snapshot's listing
  is depth-capped and trimmed when the output runs long, while
  browser_smart_snapshot's is neither. A step recorded from the smart
  listing therefore had its count compared against a different measurement
  at replay, and could stop on an untouched page, reporting a population
  change that had never happened. Each step now records which listing its
  numbers came from, and a disagreement is re-read from that same listing
  before it is treated as a change — so a real change still stops the run, and
  a difference that is only between the two ways of counting no longer does.
  The extra reading happens only on a disagreement, so ordinary replays are
  unaffected, and traces recorded before this release keep the behaviour they
  had. (#1191)

What's Changed

• docs(cli): list --profile on mcp register in the help text by @openwong2kim in #1184
• fix(browser-replay): stop a replay when the same-name element population changed by @openwong2kim in #1179
• fix(browser_repl): ref type contract, and the hints a run was swallowing by @openwong2kim in #1180
• fix(browser): make the emulated identity agree with itself by @openwong2kim in #1181
• fix(browser): let a device preset keep navigator.platform by @openwong2kim in #1186
• fix(browser-replay): verify a replayed element by its neighbourhood, not only its position by @openwong2kim in #1185
• fix(cli): wmux mcp check names the isolated instance it is reading for by @openwong2kim in #1187
• feat(browser): dispatch real touch under a device preset by @openwong2kim in #1190
• perf(browser): keep browser_repl snapshots diffed, with a complete refs list by @openwong2kim in #1188
• fix(browser-replay): verify by the element's own attribute, and count with the recorder's enumeration by @openwong2kim in #1191
• chore(release): 3.50.1 by @openwong2kim in #1192

Full Changelog: v3.50.0...v3.50.1