Rewrite Cursor article to use Claude Sonnet 5 and Haiku

[ethanwestfall]

Summary

• Replaces the GPT-5.6 Sol/Terra/Luna framing throughout with anthropic/claude-sonnet-5 (main coding) and anthropic/claude-haiku-4-5, across all four languages
• Testing found Token Station's GPT-5.6 routes can read/discuss code in Cursor's Agent mode but consistently fail to apply file edits, a tool-call response format gap on Token Station's side that's being addressed; Claude Sonnet 5 and Haiku work end to end through the identical BYOK setup, including real Agent-mode file edits
• Adds a new Step 4 showing Sonnet 5 implementing a real feature in httpie (adding the effective URL to --meta output), delegating research and verification to two custom subagents
• Subagents are named bill-the-explorer and jill-the-test-runner rather than explore/test-runner, after discovering explore collides with one of Cursor's own built-in agents and gets silently misrouted there instead of ours
• Documents a second, more significant finding from running the actual demo: Cursor's subagent system does not currently honor a custom model: field for BYOK/custom models at all (confirmed directly by the agent's own tool-spec introspection, and matching multiple independent, still-open reports on Cursor's community forum). Every subagent actually runs on the parent conversation's model, not a cheaper one, regardless of what model: specifies. Step 3, Step 4, and the closing section describe this accurately: subagents are real and useful for role/permission scoping (name, description, readonly are all honored) and for delegation via automatic or explicit (/name) invocation, but not currently a cost tier
• Updates the closing section to reflect current status: Claude confirmed working end to end (including Agent-mode edits), the subagent model-routing limitation, GPT-5.6 support in progress, xai/grok-4.6 mentioned as another supported option
• Removes the Tab-focus workaround note from Step 1 since a recent Cursor update fixed that input bug
• Cover art regenerated to match (2 model chips instead of 3, updated headline/subhead)
• Sitemap lastmod bumped for all four language URLs
• All four videos are in place, recorded and edited against the final Claude-based flow: registering the provider, adding the two custom models, creating the two subagents, and the full httpie demo

Verification

An independent adversarial review checked all 8 factual claims in the article (Cursor's settings UI path, the subagent model-routing bug against Cursor's own forum, the explore built-in-agent collision, the Cursor Pro gating, etc.), confirmed all four videos exist at the correct paths with distinct, non-trivial file sizes and no accidental duplication, verified byte-for-byte consistency of model IDs, subagent names, video paths, and cover path across all four languages, confirmed structural parity (headings/code fences/figures/tables) across all four, and checked house style (no em-dashes, correct punctuation conventions, no AI-voice phrases). No blocking issues found.

Test plan

• npm run build succeeds
• Verify at ~375px width: no horizontal scroll, videos and code blocks stay within their container
• Confirm language switcher navigates correctly between all four versions
• Confirm this article leads the /blog/ listing as the newest post (date 2026-08-21)