web-fetch

# Web Fetch With a URL in hand: - Static page text: `./ax <url>`; run `./ax agent-context` for the full agent-facing guide - X/Twitter, Bluesky, YouTube posts/transcripts: `./tgrab <url>`; `./tgrab --help` carries the agent contract Searching, in this order: 1. `exa` — semantic search with filters, plus extraction in the same call. Default choice: one HTTP round trip, and it returns sources to judge rather than a conclusion to trust. [references/exa.md](references/exa.md) 2. `grok -p "<prompt>"` — tweet search; it has the native X index. Feed the URLs it returns to `./tgrab` when full content is needed. 3. `codex` — escalation, not a first stop: use the environment-appropriate Codex path when the question needs multi-step digging or synthesis across sources, when exa's results are ambiguous and an independent reading helps, or when the 1Password key exa needs is unavailable. [references/codex.md](references/codex.md) Inside T3 Code, run grok and codex as [references/inside-t3.md](references/inside-t3.md) describes instead. If `ax` or `tgrab` fails, fall back to the host's web fetch/search tool or `curl`. A browser is the last resort, only for a page the tools above cannot read — JavaScript rendering, a login, or interaction. Which browser to reach for, in order: [references/browser.md](references/browser.md). Delegate the reading, not only the fetching: hand the whole task to the current agent's own subagent (the Agent tool in Claude Code, a native subagent in Codex) and ask for a digest, so whole pages and JSON payloads never enter this context. Two things stay here: `exa`, whose 1Password key no Codex sandbox mode reaches, and anything wanted verbatim, such as the exa docs pages above, where a digest rewrites the exact names you came for.