turno-browserActions
Flows, recordings and traces
Save what an agent did on a site and run it again, and record sessions as flows, traces and screenshots.
Flows: record and replay
An agent can save what it did on a site as a flow and run it again later without working the page out:
browser_record with action: "start" and a name, the actions, then action: "stop". browser_flows
lists the saved flows and browser_replay runs one. Flows are files per identity under
~/.local/share/turno/browser/flows/<id>/.
Each step keeps the element it acted on four ways (Playwright's own locator, role and accessible
name, visible text, a CSS path), tried in that order on replay; the snapshot ref is kept for
reading only, since refs change with every snapshot. When no locator finds exactly one visible
element, the replay stops and reports the step, each locator and why it did not fit, the steps
that did not run and a fresh snapshot: the agent does that step itself and calls browser_replay
again with from set to the next step.
Text typed into a field becomes a parameter, with what was typed as its default. A sensitive one
keeps no value: a password field, a one-time code or card field, a field named like password,
otp, pin or token, a vault value typed by hand, or any parameter the agent lists in
sensitive when it stops. A vault sign-in is a browser_login step with the site name only.
Replay needs the sensitive values again; a value written keychain:<item> is read from the secret
store (turno-browser secret set <item>), so the agent never sees it.
turno-browser flows school # saved flows and their parameters
turno-browser flows school sign-in # its steps
turno-browser replay school sign-in --param password=keychain:school-passwordturno-browser replay runs the flow on the identity's Chrome through turno-browser-mcp, the
same way an agent's browser_replay does; a sensitive parameter is a keychain: reference or
is typed at a hidden prompt, never a value on the command line.
Recordings and traces
Off by default. Per identity:
turno-browser record school --flows --traces --screenshots # what to record, from the next Chrome start
turno-browser record school --off
turno-browser set school --record-traces false --record-max-size 500MBflowssaves every session's actions as a flow namedauto-<time>; the newest 20 are kept.tracesrecords a Playwright trace (DOM snapshots, network, a screencast) of each browser session.screenshotstakes a JPEG after each action.
Traces and screenshots go to ~/.local/state/turno/browser/identities/<id>/recordings/, one
folder per browser session, together at most record.maxSize (200MB by default; the oldest go
first). turno-browser traces school lists them and turno-browser traces school --open shows the
last trace in the trace viewer that patchright-core ships, served on 127.0.0.1 in your default
browser. A trace is saved when turno closes the browser (idle limit, browser_close, end of the
session) and before a sign-in; if the server is killed, the trace in progress is lost and the
screenshots stay.
A recording never holds a vault sign-in or a password. Before browser_login fills a form, or a
value goes into a password field, or a replay types a sensitive parameter, the trace is stopped
and saved; it starts again only after every page that holds the value has navigated away, so
neither the filled form nor the request that sends it is traced, and no screenshot is taken of
that page. Each saved trace is then searched for those values and deleted if it holds one. There
is no video recording, because a video cannot leave a sign-in out. A trace still holds what the
pages showed and their cookies: treat the folder like the profile itself.
With the other settings:
"sharing": "tabs": a flow recording belongs to the agent session that started it, and only that session's actions go into it;browser_replayruns in that session's own tabs. A trace covers the whole browser, so it holds every session's tabs.- Attach mode: flows work; traces and screenshots are refused, because they could show your own
tabs. turno says so before it connects, and
turno-browser setrefuses the combination. - Read mode (
"engine": "lightpanda"): flows work, but a replayed step that needs Chrome fails like any such tool call; traces and screenshots are refused, since Lightpanda renders no page. humanize: a replay runs each step as the stock tool call, so it moves and types like the agent's own calls on that identity.