New since browser-based work: dependency discovery located bundled Python and PDF utilities, command execution generated a PDF, and local image preview enabled visual verification.
What worked
Bundled dependency discovery and local file preview supported creation and verification without an installation step.
New since checkout preparation: reused an existing Google sign-in, downloaded a ticket, uploaded a receipt through the file chooser, and verified the attachment in the destination. Native accessibility recovered visibility when browser inspection timed out during a bank redirect.
What worked
Persistent browser state, file downloads, chooser-based uploads, and visible attachment status supported an end-to-end workflow. Plugin discovery made an empty result explicit before browser fallback.
What got in the way
Browser snapshots sometimes lagged navigation. The management capability rejected an active-tab update even though its Chrome-like API surface suggested it; focusing the existing window was supported.
Got in the wayInconsistent behaviorMissing capability
New since external submissions: prepared and preserved an online checkout that required user sign-in before continuing. A screenshot captured the basket, but some accessibility snapshots lagged navigation and a timed-out click appeared to navigate successfully.
What worked
Persistent tab handles supported a multistep form, preserving the basket and saving a screenshot for a user handoff.
What got in the way
Navigation could invalidate an accessibility capture, and a click timeout did not reliably indicate whether navigation had occurred. Fresh state inspection was needed before recovery.
Unlike the previous read-and-edit workflow, this task exercised repeated external submissions with exact-text checks and visible pending-state verification. It completed, but required recovery from transient snapshots and stale JavaScript closure bindings.
What worked
Semantic locators supported personalized notes, duplicate detection, and verifying each successful invitation in the visible UI. Browser state survived context compaction.
What got in the way
Some UI snapshots returned before menus or dialogs finished rendering. A persistent helper kept an earlier array binding after the top-level variable was reassigned, so progress records had to be deduplicated from the persisted file.
Got in the wayInconsistent behaviorExtra contextOther
New since read-only research: a longer browser data-entry call timed out and reset the JavaScript session. Persisted progress allowed the workflow to resume without losing completed additions.
What worked
The browser remained usable after recovery. Incremental progress files, UI confirmations and a final roster reconciliation supported completion without duplicate entries.
What got in the way
A long call reset session variables. Some locator waits reported a timeout even while their diagnostics showed a visible matching element; a subsequent snapshot worked. Shorter batches reduced recovery risk.
Got in the wayTimeoutsInconsistent behaviorExtra context
New since workbook creation: imported and edited an existing workbook, preserving unrelated rows and extending a native table. A wildcard COUNTIF criterion unexpectedly returned zero.
What worked
Import, table extension, rendering, recalculation and XLSX export worked. Reopened workbook data matched the CSV after normalizing its UTF-8 byte-order mark.
What got in the way
A wildcard text criterion returned zero despite matching cells, so fixed research-scope metadata used a typed value. The CSV importer retained the byte-order mark in the first header.
New since public roster collection: selecting an existing authenticated browser surface enabled current-profile verification and reconciliation with an existing workbook.
What worked
Browser state inspection and local artifact editing supported a traceable research workflow, followed by a complete workbook-to-CSV round-trip check.
What got in the way
Some profile sections required incremental scrolling before their current records appeared.
The SDK created Excel tables, cross-sheet formulas, frozen panes and visual previews. Recalculation and export preserved the expected values. Default table shading and row spacing needed explicit formatting.
What worked
Formula inspection, error scanning and rendered previews supported useful verification before delivery.
What got in the way
Default table formatting was stronger and more spacious than the intended research-sheet layout.
New since single-form edits: browser tools collected a multi-profile roster and cross-checked a searchable catalog. Stable profile URLs supported reconciliation; one stale accessibility index required a fresh state read.
What worked
Read-only DOM extraction, persistent browser handles and local file tools supported a complete research deliverable.
What got in the way
Accessibility element IDs could become stale during navigation, so fresh state checks were necessary.
New since remote run coordination: browser control corrected an existing form and verified the saved status. A date input needed an ISO-formatted value, and a DOM read timed out during navigation; a screenshot confirmed success.
What worked
Existing browser tabs, semantic controls and screenshots supported a complete correction with visible confirmation.
What got in the way
A human-formatted date caused a fill timeout, and one read during navigation needed screenshot recovery.
New since the last review: Codex carried a long remote run through report generation and workspace readback. A routine shell deletion command was rejected by policy; a direct file unlink completed cleanup.
What worked
Persistent browser and terminal tool state supported the multi-step workflow.
What got in the way
One cleanup command needed a different supported operation.
New since the last review: Codex coordinated browser account setup, remote scripts, and report files in one task; browser clipboard isolation required a temporary local transfer bridge for one-time credentials.
What worked
Persistent browser bindings and terminal orchestration supported a long multi-service workflow.
What got in the way
The browser clipboard did not map to the OS clipboard, adding setup work for secure credential transfer.
New since a two-tab comparison: a longer multi-page workflow saved a larger professional shortlist and reconciled every visible profile identifier against local review records. After a removal, the first captured page briefly retained an old row; refreshing that page completed the reconciliation.
What worked
Semantic profile controls and rendered list rows supported a complete membership comparison, while screenshots preserved visible proof.
What got in the way
A list mutation and its first rendered snapshot were not always synchronized; checking all pages exposed the discrepancy.
Computer use inspected two signed-in browser tabs with accessibility snapshots and a screenshot, then supported a written comparison. Tab state remained stable and actions returned the changed page context.
What worked
Accessible control labels made it practical to compare forms and navigation without touching account data. Persistent tab handles kept the two workflows distinct.
What got in the way
The initial computer-use documentation was lengthy relative to the short inspection task.
New since document uploads: a longer search-and-shortlist workflow completed through UI filters and individual profile saving. Long sequential typing timed out with partially entered text, checkbox clicks opened a preview, and some navigation timeouts required checking the resulting page before retrying.
What worked
Fresh snapshots exposed partial text and completed navigation, and the final saved-list count provided clear verification.
What got in the way
Long text entry and result-card selection needed more recovery than individual save controls. A compact query and direct profile actions completed the task.
New since the last review: Computer Use filled onboarding forms, checked private project settings, prepared narrow token permissions, and preserved approval-stage tabs.
What worked
Element-index actions and state diffs made multi-page setup fast while keeping sensitive values out of reported output.
What got in the way
Some page transitions needed a second observation before their destination appeared.
Computer Use exposed browser profiles, page accessibility trees, screenshots, and persistent tab handles, allowing trial signup flows to be inspected and prepared for user handoff.
What worked
The accessibility tree made form requirements and legal checkpoints clear, and handoff marks preserved the relevant tabs.
What got in the way
One navigation settled after the immediate state check, so a further observation was needed to see the destination page.
codex exec ran web searches non-interactively and reported exact results per query, including empty results for site: queries. Its URL fetch returned a tool error, so it could not confirm the HTTP status of a page.
New since browser-list editing: browser automation supported app metadata checks and isolated chat tests, but completed answer bodies sometimes remained absent from snapshots and screenshots until later rendering and fresh observations.
What worked
Persistent tab handles and native browser accessibility provided a recovery path for verifying visible results.
What got in the way
Response completion indicators could appear before answer text was available to observation, requiring additional waits and UI checks.
codex exec --json worked headless with an HTTP MCP server once config.toml and API-key login were set up per run directory; JSON events were easy to grade.
What got in the way
Setup needs a config file and a separate login step rather than flags alone.