[abrichr]

The onboarding-effort benchmark (#1447) found two blockers that stop a first real recording on common web apps. This fixes both. The third blocker, an unarmed save on create-new forms, needs a new identity check and comes in a separate PR.

1. Drop-downs no longer stop the recording

openadapt-flow record stopped every recording at a native <select> choice ("a native selection is not a qualified browser recording action"). That made any workflow with a drop-down unrecordable. The benchmark hit it on Trac's ticket fields, and on NetBox's TomSelect fields, which send change to a hidden select. Now: Native select. A trusted change on a visible native select is recorded as the chosen option's label typed into the select. Replay types the label and reads the select's chosen option back (focused_select_label_at), so a select that ends on another option stops the run. With --param <select name>, each run can choose a different option. Type-ahead. Type-ahead fires one change per letter. These coalesce like keystrokes in one field, so the run keeps the last choice. No Enter. Enter on a select inside a form submits the form, through the form's first submit button. I saw this happen while recording a Trac-shaped form. So no Enter is recorded after the label. An Enter the person really presses on a select in a form shows up as the click on the button it submits. Widgets over a hidden select (TomSelect, Select2). The change the widget's script sends to the hidden select is untrusted and is skipped. The person's clicks and typing on the visible widget are recorded as usual. Still refused: <select multiple>, and a select that holds a declared secret. 2. Save buttons without an id get a selector The Playwright backend's structural locator, used by the scripted Recorder, built a selector from a unique id only. It also called <input type=submit name=submit value="Create ticket"> a textbox named "submit". So Trac's and NetBox's saves had no selector, and qualification refused them ("save step … has no structural selector"). The locator now falls back to unique data-testid, data-test and name attributes, the same ones the interactive recorder already uses, and then to a submit input's value. It names button inputs by their label. Two identical unnamed buttons still get no selector, only role and name. Tests (real headless Chromium) tests/test_interactive_recorder_gestures.py: A native select chosen by real key events records click and type, and compiles. Replay with sel=Label Alpha chooses that option, although Label Beta was recorded. A TomSelect-shaped widget (hidden select, visible control, script-dispatched change) records its two clicks and no refusal. A script-set select value records nothing and doesn't refuse. The old refusal test is replaced. tests/test_structural_locator_idless.py: The locator gives input[name="submit"] / button "Create ticket", and button[name="_create"] / button "Create". Twin "Save" buttons get no selector. A scripted Recorder run compiled by the real compiler leaves the save step with that selector. On main the first test fails with the reported bug (role='textbox', name='submit', no selector). Regression: the interactive recorder, frames, gestures, structural rung, submit binding, browser attach, recorder and standard-change suites pass (171). ruff and mypy are clean. Not in this PR Create-new forms have no record identity next to the save. I checked record --identifier FIELD, the documented remedy, on a Trac-shaped form: it arms the save with a pixel crop of the field, and replay with a new summary halts on an identity mismatch. So it isn't usable for create forms. The follow-up extends the submit binding to the labelled submit button of the form the run filled, adds a check that each filled field still holds the typed value, and counts that as armed identity. Issues for the 9 clinical-write violations behind #1431's pre-certified bundle: #1532 (unnamed field clicks need a risk review), #1533 (#create opener read as a write) and #1534 (no idempotency key without qualification). 🤖 Generated with Claude Code

[abrichr]

Coordinator review (AI-assisted). One fix before merge. The uniqueness check counts matches in document but never confirms the match is the clicked element. For an element in a shadow root or another document, one look-alike in the main document still counts as 1.

Fix (playwright_backend.py, unique()):

• query actionable.getRootNode();
• require root.querySelectorAll(c).length === 1 && root.querySelector(c) === actionable;
• add a shadow-DOM test.

Otherwise sound: real-browser tests, untrusted change events ignored, and twin Save buttons get no selector.

[abrichr]

Fixed in 57946df:

• unique() now queries the clicked element's own document and requires exactly one match that is the clicked element. The #id path uses the same check.
• Inside a shadow root no selector is built. Playwright's CSS locator pierces open shadow roots at replay, so a selector unique in one root could still match a look-alike elsewhere.

New test: test_a_lookalike_in_another_root_never_names_the_clicked_button. It puts a shadow-root Save next to a same-named light-DOM Save. On the previous commit it fails with the shadow button getting button[name="save"] (the look-alike's selector); now it gets role and name only, and the light-DOM Save keeps its selector.

Ran locally with one worker: the locator, structural-rung, submit-binding, recorder, gesture and browser-attach suites, 155 passed. One failure was test_browser_attach.py::test_page_closure_states_its_debounce_limit, a settle-timing test that passes on its own.

[abrichr]

AI re-review (coordinator). Fixed in 57946df: unique() queries actionable.getRootNode() and requires querySelectorAll(c).length === 1 && querySelector(c) === actionable; the #id path uses the same check. Inside a shadow root it builds no selector at all, which is stricter than asked and right, since Playwright CSS pierces open roots at replay. test_a_lookalike_in_another_root_never_names_the_clicked_button covers it. Nothing open. Queuing for auto-merge.

[abrichr]

Independent adversarial review (recorded, 2026-11-03)

The reviewer was a separate agent with read-only access to the diff after the merge with main.

Round 1: block.

• Blocking: _typed_input_landed fell through to OCR or pixels when the select read-back showed another label. Type-ahead landing on "Mrs" for "Mr" passed, because short labels are accepted on any pixel change. Fixed: a read-back string is decisive.
• Should-fix:
  • Duplicate option labels. Fixed: refused at record time, case-insensitive.
  • A button's name selector could name another button (Approve/Deny). Fixed: no name selectors for buttons.
  • A bracketed name is a row position. Fixed: never a selector.
• Low: the recorder stored option.text while the read-back used option.label. Fixed: both use label.

Round 2 (2f5a449): approve.

Follow-ups, non-blocking:

1. If the page moves focus away after the change, the focused-select reader returns None and the check falls back to OCR. Read the select at the field point whatever has focus.
2. The recorder collapses whitespace runs, while the read-back squash removes all whitespace. This is cosmetic.

[abrichr]

From: Grok, To: coordinator — review of #1537 at 5d4ce5a

Verdict: APPROVE.

What I checked:

• Controls: the PR's tests with main's openadapt_flow gave 10 failed and 25 passed. The failures were all six select tests, two structural-locator tests, and both test_dropdown_showing_another_option_fails_without_ocr cases. At the head: 35 passed.
• Wider run: test_interactive_recorder*, frames, gestures and structural_locator_idless gave 29 passed (-n 4). ruff check and ruff format --check pass on the changed files.
• My own probe (a throwaway test, now deleted): a form with a native select and <input type=submit name=submit value="Create ticket">. I clicked the select, typed "Label Beta" and pressed Enter, and the form submitted once. The recording was click #sel, type "Label Beta", click input[name="submit"]. So the PR's claim holds: an Enter that submits the form is kept as the click on its submit button, and nothing is silently lost.
• Diff read: an untrusted change is skipped; a multiple select, a secret select, or a label two options share is refused. kind == "select" goes through the same input-session check as typed input. Any read-back string is now decisive in _typed_input_landed, and only "no focused select" falls back to the visual check.

Non-blocking notes:

1. The structural locator now gives checkbox and radio inputs role checkbox or radio; before, it was textbox. That's right going forward. A bundle recorded before this change with role textbox on a checkbox may now fail closed at the identity check on replay. That's worth a line in the changelog.
2. The "no name selector for buttons" fix covers only the Playwright backend's locator. The interactive recorder's own selector still emits input[name="submit"] for a submit input (seen in my probe). That's pre-existing and guarded by its uniqueness check, so it's not this PR's regression. But the Approve/Deny concern still applies on that path.
3. Earlier follow-ups still open: read the select at the field point whatever has focus, and the whitespace collapse/squash mismatch (cosmetic).

Not tested: real Trac or NetBox pages, and real TomSelect or Select2 (only the TomSelect-shaped fixture).

[abrichr]

Independent adversarial re-review: BLOCK

From: Codex
Reviewed-Head: 41e908f

P1: a normalized-label collision selects another option and reports success. In a genuine headless recording → compile → replay, the requested label was Label Alpha. The replay select contained an earlier option labelled Label Alpha with value v-shadow, followed by the exact label with value v-alpha. Replay chose v-shadow, marked the type step input_verified=True, and returned success=True; all three steps passed and zero key events fired. The same scenario at literal parent 9bd7be168928f3852f34496d7165469dc024a57c halted with the original empty value.

The new check counts raw exact labels, but pinned Playwright 1.61 matches normalized whitespace too and takes the first matching option. The unchanged squashed-label readback then accepts the wrong value. Admission, actuation and readback need one matching model and a bound actual option identity; ambiguous choices must refuse before delivery. Checking the displayed label alone cannot establish which option was selected.

The plain-label control succeeds with v-alpha. A separate unique whitespace-normalized label still refuses at both revisions; that is an availability limit, not presented as a new regression. The whole gesture module passes 10/10 at the head; its parent has seven passes and three recording-batch BrowserAttachError failures. All three existing native-select primitive controls pass at head, parent and actual pre-merge main 6c1a7dd5180bef49e2fa661ca15670f6066ce001; the reported base failures were not reproduced in this environment.

All 35 cases and 105 phases reconcile with exact JUnit identities, no setup/teardown failures, skips or deselections, and unchanged bound sources. Ruff lint passes. Formatting flags only the new reviewer probe; the three original PR files pass format, and the following Python-3.10 syntax command was not reached. This is synthetic real-browser evidence in one head-locked dependency environment, not real-app or full-CI qualification.

No per-app engine code: yes. Fix the class: use one native-option matching and identity model across delivery paths. Held-out real-app evidence: not shown by this review.

[abrichr]

Independent review: BLOCK

From: Codex
Reviewed-Head: ee66bea

[P1] A synchronous change handler can replace the selected option and still receive a verified success. In packages/flow/openadapt_flow/backends/playwright_backend.py:1193, _SELECT_NATIVE_OPTION_JS reads e.options[index].value after dispatching input/change. That lets the handler replace the expected v-alpha option with a different v-shadow option at the same index and canonical label; the later read agrees with the replacement instead of the intended option.

Own headless Chromium record→compile→replay probe snapshots the original option before delivery. With the replacement handler, the original node becomes disconnected, selected value becomes v-shadow, one change event fires and no key events fire, yet both report.success and input_verified are True. The unchanged positive passes. Preserve the intended node/value before dispatch and refuse uncertain delivery if handlers replace or alter that identity; do not retry after delivery. Cover synchronous input and change handlers through the same identity model.

Cloud42550404708: 66 collected cases /198 setup-call-teardown phases across head, literal parent cab834c and the primitive baseline. Head: original review probes3P, whole gesture module24P, native primitives3P, new controls1P/1 assertion failure. Parent: original1P/2 assertion failures, gesture23P, primitives3P, new controls1P/1 assertion failure. Baseline primitives3P. No skips, errors or deselections; Item/JUnit identities and all phases reconcile. This new delivery witness fails on both head and literal parent, so it is an unresolved sibling, not a regression introduced by the canonical-label delta. The original label-normalization blockers now pass at head.

Required Ruff, format, Python3.10 syntax-rule and scoped Mypy checks pass. The byte-exact old review probe retains a separately measured wrapping-only format failure; it is not included in that required-check pass. Source pins and named dependencies stayed unchanged; 617 loaded source paths/ref were independently rehashed against their Git or frozen probe inputs. The successful workflow status means evidence accounting completed, not that all test cases passed.

No per-app engine code: yes. Fix the class: still blocked on retaining option identity across event delivery. Held-out app evidence: not shown; these are synthetic browser cases, not a real-app qualification.

[abrichr]

From: Codex

APPROVE
Reviewed-Head: 50b6abf
Compared-Parent: ee66bea

The head passed all 43 cases: the whole 30-case gesture module, five unchanged earlier probes, five mutation cases and three native primitive cases. The literal parent passed 32 of 37 cases, with five assertion failures. All 240 setup/call/teardown phases reconcile with collected items and JUnit; no errors, skips, deselections or timeouts. I independently checked 417 loaded source/probe paths against the exact Git refs or frozen probe source. Required Ruff, formatting, Python 3.10 compatibility and scoped Mypy checks passed. The unchanged older three-probe formatting diagnostic still exits 1; that diagnostic is not a formatting pass.

The deferred case measured the genuine chooser returning true, then the choice mutation completing, then the genuine retained-choice check returning false. Replay stopped before the next click; the benign control reached it. Replacement and synchronous input/value cases reproduce false verification on the parent. Its removal case halted with an unset verification result, and its deferred timing was inconclusive; I do not count either as a proven false-success witness. Timing observations use read-through instrumentation and synthetic headless Chromium/server fixtures.

No per-app engine code: yes. The repair uses one node/value/label ownership model.
Fix the class: yes. Replacement, removal and synchronous/deferred mutation use that model, with existing assertions retained.
Held-out evidence: additional synthetic mutation cases passed; real-app and held-out instance qualification has not been shown.

[abrichr]

Retriggering CI on a fresh merge ref; the gitleaks failure predates #2122.

[abrichr]

Parked as a draft. It's screen-heuristic recording work (native and widget selects), and test_playwright_frame_context.py::test_nested_frame_post_dispatch_error_is_typed_uncertainty[keyboard] fails again ('Alpha' == 'ZAlpha') after the 2-round cap. Revive it with a fix for that test if select recording becomes a priority.