From a recording to a reusable flow

This is how log learns to do something on a website once and then repeat it on any account. You do the task step by step with log inspect, which records everything; log turns that recording into a flow, a state machine, a command and an MCP tool. Read this if you want to automate a new task, or understand where the saved flows come from.

It works on every platform with a browser path (Facebook, Reddit). Every run passes the same gates as any other: proxy, device, locale, self-check, active hours, launch budget and a raw recording of the network (HAR).

The pipeline

Inspect actions

Details:

• A step settles when the network has been quiet for 1.5 s (up to 15 s, longer while an expect is pending).
• Typed values are hidden in snapshots.
• A HAR over rea's 32 MB limit gets a copy without static file bodies for rea; the raw HAR is kept.
• --now skips only the 10-minute gap between launches.

The mouse

Every session moves the cursor with python/human/humanMouse.py, which is fitted to a recording of a real person's mouse made with python/human/recordMouse.py. What it reproduces:

• one smooth speed curve per move, peaking early and settling slowly;
• overshooting and coming back about half the time;
• loose loops around whatever is being looked at;
• short pauses, and staying still about 25% of the time.

Every step writes mouse.json (its moves) and mouseSamples.json (every point, in the recorder's format), so the same analyser can compare the two:

python python/human/recordMouse.py 60                      # record your own mouse for 60 s, then analyse it
python python/human/recordMouse.py --analyze <run>/steps/02_wait60s/mouseSamples.json

Camoufox's own humanize setting stays on, but is capped at 20 ms per hop in these sessions (launch_kwargs(…, own_trajectory=True)). Its default cost about 100 ms per 5 px hop, which made earlier movement crawl.

How a flow step runs

A flow never depends on a page's exact wording or on a person stepping in (rule 79). Each control is one scored match by role and name inside the dialog the step names, never a guess (controlAmbiguous, controlMissing, controlDisabled); every step waits for the call it must cause and checks its answer before the next one fires (rule 73); and before the first click, fill, upload or location the known overlays are pressed away with Escape, each checked, or the flow stops with overlayNotCleared. Each of these is a requirement in the flows spec.

Saved flows

Every flow below is in src/browser/flow/definitions/, has its doc in docs/platforms/facebook/flows/, and is an MCP tool (flow<Name>, a dry run) with its live twin (flow<Name>Live, needs confirmLive). log flow list prints them.

profileSetup was recorded on one account and then ran 13 out of 13 steps on another.

What early recordings showed

• A first sign-in on a new account can land on "We suspect automated behaviour" (/checkpoint/601051028565049/). Pressing Dismiss sends FBScrapingWarningMutation and lands signed in.
• A locked account went through these screens in order (recorded before a checkpointed account was deactivated instead of pressed through, rule 82; see the page states spec): "Confirm that you're human to use your account" [Continue] → useUFACSubmitActionMutation (5 s) → a text captcha → solver (1.5 s) → [Continue] → useUFACSubmitActionMutation (3 s) → "Confirm your identity with a video selfie". Only a person can clear the last screen. A second visit picks up at the step it was left on.
• Changing exit IPs stood out. One account signed in from three exit IPs in 90 minutes over 15 launches; accounts whose exit IP stayed fixed fared differently.
• Connect read the page too soon. It read 1 s after pressing Enter, saw the login form still there, and tried again, although the first Enter had worked. It must wait for the page to change.
• The listing's location isn't set by any request in a run. It arrives in the create page itself (sell_location), so a wrong Marketplace location is an account setting.

Where things are

The rules behind this are in AGENTS.md.