docs: rewrite What is BitRouter? around bitrouter/auto

[SPIKESPIGEL404]

Rebuilds /docs/overview/what-is-bitrouter so the positioning line — context-aware LLM router that continuously improves your agent workflows — is the page's spine, with each section defending one clause of it.

Why

The page opened with 190 words of origin story before defining the product, so the first quotable sentence was "We started BitRouter out of frustration." Its value props (cost, open source, not-just-models, Cloud) sat beside the tagline as co-equal claims rather than laddering up to it, and the mechanism it described — v1 opening / after_<tool> / midstream fingerprints — is superseded by the v2 projection.

Shape follows Diátaxis (explanation: no step-by-step, no close-up machinery) and answer-engine research (answer-first lede, assertion headings, short paragraphs), since "What is BitRouter?" is a query LLMs answer on our behalf.

What changed

• Lede is a 60-word answer-first definition leading with bitrouter/auto, framed as taking model selection off your plate without taking it out of your hands. Pain reframed onto the moving cost/latency/accuracy frontier rather than static routing alone.
• Headings are assertions, each carrying a contrast: One model id replaces every model decision (vs. gateways that start after the hard choice) · Your production traffic is the training signal (vs. observability tools that hand you a dashboard and leave the homework) · It optimizes the workflow, not the request (vs. every gateway's request-scoped unit).
• "Context-aware" now describes the v2 agent_trace projection — trajectory position × risk band — replacing the v1 fingerprint scheme.
• New third section covers the workflow-level optimize loop (your command, your success contract, one routing change per run, publish/rollback by digest). The old "not just models" section was a breadth claim that missed this entirely; the three routable primitives are now a supporting clause.
• Cost / OSS / Cloud demoted to single clauses. Self-host details, callouts, and the policy evolve commands dropped to keep the page hosted-first.

Net +19 lines; ~940 words vs. ~780.

Merge order

Both dependencies land first, per author:

• feat: publish bitrouter/auto as the router slug bitrouter#788 — publishes the bitrouter/auto slug this page is built on.
• docs: rewrite the quickstart around bitrouter/auto and the optimize loop #65 — creates the #start-optimizing anchor the Next-steps card targets, and documents the optimize surface this page introduces conceptually. It also touches this file (stripping #adaptive-routing anchors), so expect a small conflict here.

Known gaps

• The ~30% Terminal-Bench 2.1 figure has no public report to link. Stated without a source, and without a quality-held claim — workflow-optimization.md explicitly warns against promising fixed quality-loss or latency percentages.
• bitrouter optimize is undocumented on this site (absent from content/, cli-overlays/, and .cli-snapshot.json). docs: rewrite the quickstart around bitrouter/auto and the optimize loop #65 fixes this; until then a reader who hits the unbound-policy error sees a command with no reference page.
• quickstart.mdx and configuration.mdx still document v1 fingerprints, so they disagree with this page once #788 lands. Separate follow-up.

Verification

• pnpm lint:docs passes (49 docs, 6 sections)
• Rendered locally: three assertion headings in the TOC, no callouts, code block and cards render, no page-specific console errors

🤖 Generated with Claude Code