SDK guides

Walkthrough

A Walkthrough is one narrated pass over your own site: you talk while you browse, pick or circle what you mean as you go, and press Stop. Fixback transcribes what you said, cuts the walkthrough into one item per thing you raised, and files each as its own Issue — with your words as its comment, your marks on its screenshot, and its own replay and Trace.

Who can start one

Turn on Walkthroughs in the Project's settings, under Capture & analysis. Then, on any page running the web SDK, a Member who is connected sees Talk & mark in the report panel's footer. The SDK learns this from its boot answer — the server decides, never an init option — and re-checks it on every request the walkthrough makes, so a Member removed from the workspace is refused on their next one.

Press Talk & mark (or ⌥ ⇧ W from anywhere on the page). Your browser's microphone prompt is the only confirmation. The panel folds into a small recorder pill in the launcher's corner, and you are recording.

What it records

Your narration — the microphone, at 32 kbps, uploaded in one-second pieces as you go, so a closed tab loses about a second.
Your marks — each element you pick and each area you circle, stamped the moment you make it with its page and viewport.
The page, as a report sees it — the session replay and the console and network Trace the SDK already records, masked and scrubbed exactly as a report's are.

A walkthrough follows you across full page loads in the same tab. It ends when you press Stop, after 30 minutes, or about 90 seconds after the tab goes quiet — closing the tab without Stop still files what arrived. After Stop a read-only drawer shows the analysis, then the Issues it filed, each with its key.

How it is cut

Once your narration is transcribed, your workspace's Analysis model reads it as numbered utterances beside your marks and the pages you visited, and cuts it into items — each a run of what you said plus the marks that belong to it. The cut names utterances, never rewrites them, so each Issue's comment is your own words. Anything you said that belongs to no item, like "okay, now the cart", is kept with the transcript but files nothing.

If the model can't produce a usable cut after its retries — or you said nothing at all — the walkthrough still files: each mark you kept becomes its own Issue.

Marking, and the keys

The pill's buttons do everything. With a keyboard, while a walkthrough records:

Key	What it does
`P`	Pick — click an element to mark it (⌖ on the pill). One pick, then you are back to browsing.
`C`	Circle — draw a loop around an area to mark it (◯ on the pill).
`Z`	Remove the last mark. The toast offers Undo for a few seconds.
`Esc`	Leave Pick or Circle; else collapse the marks tray. It never stops the walkthrough.
`⌥` `⇧` `W`	Start a walkthrough from anywhere on the page (Alt+Shift+W).

⌥ ⇧ E and ⌥ ⇧ S work too, as Pick and Circle. The single letters are claimed only while a walkthrough records, and never while you are typing in a field, so your page's own inputs keep every key. The Page shortcuts switch in the panel's settings turns them all off. On a phone or tablet there are no keys: tap the pill's buttons, and a finger draws the loop.

Browser support

Current Chrome, Edge, Firefox and Safari, on desktop and mobile. A walkthrough needs a microphone, MediaRecorder and a secure (https) page; where one is missing, Talk & mark says why and the panel stays as it was. After a page load the microphone resumes on its own where the browser remembers your permission; where it asks again (Safari after a typed URL or a reload, for one), the pill offers Tap to resume, and your marks keep counting meanwhile.

The recorder follows the SDK's standing rules: it honours reduceMotion and the OS reduced-motion setting with a short fade in place of every movement; it grows its buttons to touch size under a finger, and the results drawer becomes a bottom sheet on a phone; every control is named for a screen reader, which also hears marks added and removed; and it never takes focus from your page.

Data terms

Transcription. Narration is transcribed by AssemblyAI in the EU, on Fixback's own AssemblyAI account — there is no key of yours to bring. It is not used for training, and AssemblyAI deletes the audio within 48 hours. Once Fixback has stored the transcript, it deletes AssemblyAI's copy.
Scrubbed before it is stored. The transcript passes the same free-text and secret scrubbers as other untrusted text: emails, bearer tokens and long digit runs are redacted. Each Issue's comment is your own words, never a model's rewording.
Kept like replay. The narration audio and the walkthrough's streamed replay and Trace are Evidence: kept with no time-based expiry, counted in your workspace's Evidence storage, and deleted when their Project or workspace is purged.
Your hooks. beforeBreadcrumb still filters every crumb a walkthrough streams. beforeSend is not called for a walkthrough's uploads — it reshapes whole reports, and a walkthrough streams pieces, not reports. See Privacy & data.

Limits

One walkthrough runs for at most 30 minutes, holds at most 200 marks (removed ones included), passes through at most 100 pages, and is cut into at most 50 Issues. A Member records one walkthrough per Project at a time — a second tab is told so. The recorder stops short of each limit and says why; the rest is the server's to enforce.

When Fixback cannot be reached, the SDK keeps what it has not sent and sends it once the connection returns — but only about 5 minutes of it, so memory stays bounded on a long outage. Past that the narration pauses (the pill says so) and the replay and Trace stop growing; marks keep stamping. When a send gets through, recording resumes and the replay picks up again from a fresh full snapshot, so an Issue's replay may skip the stretch Fixback was out of reach.