snapsurflisted

# snapsurf — browse by reading semantic diffs instead of screenshots > MACHINE-GLOBAL install (~/.snapsurf). Refresh after source changes: > `node tools/install-global.mjs` SnapSurf is a Playwright daemon with the SnapSurf SDK injected on every navigation, plus a CLI driven from Bash and an MCP server that exposes the same verbs as tools. Read the semantic diff after each action; ask for pixels when the report cannot answer the question. NOTE: if the SnapSurf MCP tools (browser_open/find/act/verify/…) are available in this session, prefer THEM over the CLI — same daemon, structured output. Read the fields, not the prose: every reply carries `structuredContent`, and each tool description says which field it fills. Parsing the fenced text is never necessary. Four fields decide whether an answer is usable, and all four are honest about what they do NOT know: `blocked`+`challenge` (content withheld, not a thin page), `failure` {layer,code} (dns/tls — never reached HTTP), `authState` (always `unknown` until the workflow proves identity; `cookiesForOrigin` is evidence, not authentication), and `truncated` (escalate with browser_text, never record a cut value). For long text, call `browser_text` with `maxChars: 3000` (maximum 12000), then pass its returned `continuation` object directly to the next call. Text is retained from the first read (`capturedAt`); continuations keep the same observation and session, and fail after expiry. CLI: `text <id> --max-chars 3000`, then the re