2.7.1

Cursor integration: real-catalog model picker

The Cursor model picker and per-family mode dropdown now match the live cursor-agent --list-models output. Previously the single-step family resolver collapsed only one trailing token, so multi-suffix ids (e.g. claude-opus-4-7-thinking-low-fast, gpt-5.5-extra-high-fast, claude-4.6-opus-max-thinking) ended up as their own pseudo-families and the mode dropdown was missing most variants. After this release the chat sidebar shows 25 clean families with rich, runnable mode lists.

Fixes

• Iterative family resolver with strong tokens (none/low/medium/high/xhigh/thinking/fast) and the multi-token extra-high. max is treated as a weak token that only strips when its candidate base has other strong-mode siblings, so claude-opus-4-7-max collapses into the Claude Opus family while gpt-5.1-codex-max keeps its own family.
• Standard-mode safety: a family's mode dropdown only includes Standard when the bare family id is itself a real model. Picking a family without a bare id now defaults to the first runnable variant, so the runtime never sends a --model value the CLI rejects (the previous default would have spawned --model claude-opus-4-7 which Cursor refuses).
• Compound mode UX: labels render as space-separated title case (Thinking Low Fast, Extra High Fast); the dropdown sorts non-thinking variants first, then by effort, with fast last.
• Parser hardened: drops the trailing Tip: footer line that previously leaked into the cached catalog as a phantom model id, prefers the explicit - separator from the native <id> - <label> format, and tightens the id pattern so trailing punctuation cannot survive.
• Stale effortLevel repaired: when the persisted mode is no longer valid for the selected family, resolveCursorModelSelectionForCli falls back to the first runnable variant instead of producing an invalid CLI invocation.
• Label polish: dotted version numbers (Composer 2.5, Kimi K2.5, Grok Build 0.1) are preserved, and the GPT/Gemini/Grok formatters keep size and qualifier tails so the picker no longer collides on lookalike names (GPT-5.4 Mini vs GPT-5.4 Nano vs GPT-5.4, GPT-5.1 Codex Max vs GPT-5.1 Codex Mini). Kimi K2.5 moves out of the OpenAI vendor bucket.

Tests

A real-catalog fixture (tests/fixtures/providers/cursor/realCatalog.ts) and 14 new tests pin the 25 families and key compound-mode shapes, so any future change in Cursor's id taxonomy fails this suite loudly instead of silently degrading the picker.

Also in this release

• [codex] Install Superpowers Codex skills (#6, @Luis85)

Full Changelog: 2.7.0...2.7.1