Repository navigation
click/type/fill/get-text fail with "Invalid parameters" on v0.37.0, even with a fresh valid ref (regression from #121 fix) #123
Description
Version: wmux.exe 0.37.0 (ProductVersion confirmed via file properties)
Context: #121 fixed the ref_not_found bug caused by the @ prefix mismatch between buildAccessibilityTree and resolveRef. That fix works � the old error is gone. But a new regression appeared in the same command group.
Bug: browser click, browser type, browser fill, and browser get-text all fail with a generic Invalid parameters error, even when called with a ref that was just returned by browser snapshot and definitely exists.
Reproduction:
wmux browser open https://example.com
wmux browser snapshot
# вâ� â�� returns tree with @e1..@e14, e.g. "@e2: heading \"Example Domain\""
wmux browser click e2
# вâ� â�� Error: Invalid parameters
Tested and ruled out as causes:
@prefix / no prefix � same error withe2and"@e2"- Timing / race condition � same error even with a 5s delay between snapshot and click; when the ref genuinely doesn't match the latest snapshot, wmux correctly returns a different, honest error (
ref_not_found: eN � the last snapshot has @e1..@eM), so this isn't a "clicked before the tree settled" issue - Missing
WMUX_SURFACE_ID� present and correct in the test environment, no effect
Command matrix (same session, same fresh snapshot):
| Command | Result |
|---|---|
browser open |
OK |
browser snapshot |
OK |
browser eval |
OK |
browser wait <ref> |
OK |
browser click <ref> |
Invalid parameters |
browser type <ref> <text> |
Invalid parameters |
browser fill <ref> <value> |
Invalid parameters |
browser get-text <ref> |
Invalid parameters |
Notably wait uses the same ref param as the failing commands and succeeds � so this doesn't look like a client-side param-building issue (the CLI builds {ref: args[2]} identically for all of them per dist/cli/wmux.js). Looks like a server-side validation regression scoped to the interaction/read command group, likely introduced alongside the ref-normalization fix in #121.
Workaround: browser eval with direct JS still works fine, e.g. browser eval "document.querySelector('a').click()".
Activity
fix: browser refs, drop marker, in-app update, dead popup links (#123 #…
amirlehmam commentedon Jul 28, 2026
Fixed in v0.38.0. Your triage was right on every point, including that
waitsucceeding ruled out the client side — thank you, that narrowed it a lot.Root cause.
Accessibility.getFullAXTreereturnsAXNodes whosenodeIdis an accessibility-tree id — a string — while the DOM handle lives in a separately named integer field,backendDOMNodeId.buildAccessibilityTreereadnode.backendNodeId(a fieldAXNodedoes not have) and fell back tonode.nodeId, so every ref carried an AX string intoDOM.getBoxModel/DOM.resolveNode, which require an integer.Confirmed against a live page rather than by reading code:
That last line is your error, verbatim.
Why it looked like a #121 regression. It isn't one — it predates #121 and was simply unreachable. Every ref lookup failed with
ref_not_foundbefore any DOM call was made, so the bad id never reached CDP. Fixing the lookup exposed the next bug in the same path, which is exactly why the failure moved but stayed in that command group.waitkept working because it only checks whether a ref resolves; it never touches the DOM.evalandsnapshotnever take a ref at all.What changed
backendDOMNodeIdas an integer. A hand-built node list usingbackendNodeIdstill works, and a plain numericnodeIdis still accepted, so nothing that worked before stops working.clickfalls back toscrollIntoView+getBoundingClientRect()whenDOM.getBoxModelthrows — elements with no layout box used to fail here too.fillalso dispatcheschange, not justinput; some frameworks ignored the value otherwise.no_dom_node: @e1 (RootWebArea) is an accessibility-only node with no DOM element — pick a ref that maps to oneinstead of being handed to CDP to fail asInvalid parameters. Generic protocol errors were the thing that cost you the most time here, and that class is now named at the source.Regression tests cover the wire format with string ids, assert an AX id can never leak out as a DOM handle, and drive a
clickend to end assertingDOM.getBoxModelreceives an integer. Verified live afterwards:e1/e2/e3from a real snapshot each resolve, box-model, and read back text.