How Meta's Muse Solves CAPTCHAs on Google, Amazon and Cloudflare

A topical view of this archive. The extractions sort by kind — embedded files by what they are, curated strings by the curation that produced them — so one subject scatters across several trees. This directory regroups everything about one subject, the browser agent's bot-challenge machinery, as symlinks to the canonical files, so the topic reads in one place. Nothing here is a copy: follow a link to the real file, and edit it there. The string material below was produced by the bot_challenge topical curation (extract_binary_strings.py); in binary-map.json its entries carry "topic": "bot_challenge", which is what distinguishes a topical curation from the generic file extraction's entries.

When the browser agent meets a CAPTCHA or a bot wall, it does not improvise. A classifier — the Browser Task Tagger, whose prompt is in prompts/classifier.txt — first decides what the page is: a solvable challenge, a hard bot block, an ordinary page, and so on. If it is a challenge, the work hands off to a dedicated specialist role, browser_bot_challenge, whose registry entry is in prompts/roles.txt. The specialist takes over the tab with a small toolset of its own (tools/): bot_challenge.assess reads the challenge's state and classifies the block, bot_challenge.guide fetches the handling guide for that challenge type, browser.bot_challenge performs the actions inside the challenge, and bot_challenge.recover reloads the page or replaces the environment when an attempt is spent.

The guides the specialist fetches are the per-challenge-type playbooks in skill/ — one for each challenge family, from image grids and sliders to press-and-hold and waiting rooms. The browser agent's own standing instructions for recognising checks and challenges, and for conducting the handoff, are in browser-instructions/. What varies by site — how a given site's bot detection behaves and what has worked there — is recorded in sites/, one note per site, split out of the full site guides. reference/ holds the surrounding vocabulary: the Browser* enum and telemetry names, the dispatch-string fragments that name the tool outside the main string pool, and the browser error lexicon the regions yielded.

skill (20)

The per-challenge-type handling guides — the playbooks bot_challenge.guide serves to the specialist.

The browser agent's own embedded instructions about bot checks and challenges: when a page is a challenge, and how the handoff to the specialist works.

The tool definitions: the specialist's three tools, and the browser tool through which it acts inside a challenge.

The prose that steers the machinery: the specialist's role-registry entry, and the Task Tagger classifier prompt that routes pages to it.

Per-site bot-detection notes, split out of the site guides — what each site's defences do and what the guide says about them.

Supporting vocabulary: enum/telemetry names, dispatch fragments, and the browser error lexicon.