feat(browser): agent-native payload — network bodies, html tree budgets, extract command

[jackwener]

Why

Three independent agent-usage gaps in the browser surface, addressed as one complete change per direction from #opencli-browser thread:

1. network --detail silently cut response bodies at 4000 chars, breaking JSON.parse on any non-trivial API response. Purely invisible to the agent.
2. get html --as json had no budget knobs, so the agent either got a full-page tree dump or nothing — no middle ground when scoping down.
3. No content-reading channel existed. Agents were misusing get html to try to read articles, which returns DOM structure, not prose.

All three are about the same thing from first principles: agent-readable payload shape, with visible truncation so the agent can continue rather than fail silently.

What changed

browser network — body truncation becomes visible

• src/browser/cdp.ts + extension/src/cdp.ts: raise the silent 4KB slice to an 8MB memory-guard cap and record responseBodyFullSize / responseBodyTruncated (extension path also gets request-body mirrors).
• src/browser/network-cache.ts: CachedNetworkEntry.bodyTruncated propagates through.
• src/cli.ts browser network:
  • List view now emits body_truncated_count + per-entry body_truncated flag.
  • --detail envelope emits body_truncated, body_full_size, body_truncation_reason: 'capture-limit' | 'max-body'.
  • New --max-body <chars> flag for explicit caller-side cap (default 0 = unlimited).

browser get html --as json — budget knobs

• src/browser/html-tree.ts: add depth / childrenMax / textMax budgets to the in-page serializer. The result envelope includes truncated: {depth, children_dropped, text_truncated} only when a budget was actually hit.
• src/cli.ts: expose as --depth <n> / --children-max <n> / --text-max <n>.

browser extract — new agent-native reading channel

• src/browser/extract.ts (new):
  • buildExtractHtmlJs(selector) — in-page clone + denoise (strip script/style/nav/header/footer/aside/iframe/form/button/... + role=navigation/banner/contentinfo/complementary + aria-hidden=true; scrub on*/style/data-* attrs).
  • htmlToMarkdown (existing helper) → structure-aware chunkMarkdown that snaps chunk end to the nearest \n\n (or \n) within a 15% boundary window, falling back to a hard cut.
  • Stateless resume via next_start_char cursor.
• src/cli.ts: new browser extract [url] --selector <sel> --start <n> --chunk-size <n> command.

Truncation signal contract (all three channels)

Every channel makes cuts visible to the agent in the envelope:

Channel	Signals
network list	body_truncated_count, per-entry body_truncated
network detail	body_truncated, body_full_size, body_truncation_reason
get html json	truncated: { depth?, children_dropped?, text_truncated? } (only when hit)
extract	total_chars, chunk_size, start, end, next_start_char

Test plan

• pnpm test src/browser/extract.test.ts — 13 tests pass (chunking boundaries, resume chain, min-clamp, envelope wrapping)
• pnpm test src/browser/html-tree.test.ts — all green incl. 5 new budget tests
• pnpm test src/cli.test.ts — all green incl. 5 new body-truncation-signal tests
• pnpm typecheck — clean
• Manual smoke: browser network --detail ... --max-body 500 returns body_truncated: true + body_full_size
• Manual smoke: browser extract <url> on a real article page + resume via --start <next>

Scope discipline

• No new dependencies (turndown was already in package.json; htmlToMarkdown already exists in src/utils.ts).
• No changes to list-view default behavior except the new body_truncated_count field (additive).
• No backwards-incompatible changes to existing flags.

cc @codex-mini1 @First-principles-1 — requesting review per thread assignment.