feat: GPT-6 Sol and GPT-6 Luna - catalog rows, family prompt preset, recommended default

[code-yeongyu]

Fixes #2023

Summary

OpenAI's GPT-6 family is Astra, Sol and Luna; 2026.9.22-4 supported Astra only. This PR brings Sol and Luna to the same level on four surfaces, each with a failing-first test:

• Catalog (packages/ai): gpt-6-sol and gpt-6-luna rows on openai and chatgpt-subscription (hand-added; models.dev never carries the Codex backend), plus regenerated passthrough rows on Azure, OpenCode, OpenRouter, Venice and Vercel, and -fast Priority variants on the two first-party providers. Published prices with the >272k tiers, 128k output, text + image, tool search / additional tools, and the documented ladder none/low/medium/high/xhigh/max (minimal: null; unlike Astra, off stays selectable). Prompt budgets stamped family-wide: Luna 922k (the documented input cap), Sol 400k, Astra 600k unchanged.
• Effort inference (packages/ai/src/models.ts): map-less custom rows for both ids surface xhigh and max.
• Prompt preset (packages/coding-agent): every Sol/Luna id shape resolves to the GPT-6 preset (OpenAI ships one prompting guide for the family; the rendered core names no model). The preset keeps its gpt-6-astra name so existing settings.json pins keep working.
• Defaults: gpt-6-sol:medium joins the recommendation ladder directly under Astra and becomes the implicit default for openai / chatgpt-subscription; the high-reasoning warning covers Sol at xhigh/max (Luna excluded).

Which rows came from regeneration vs. by hand

Regeneration (generate-models --strict against live models.dev): azure-openai-responses, opencode, openrouter (the eight rows 22-4 already shipped, now with the ladder + Sol budget), venice, vercel-ai-gateway. Hand-added in generate-models.ts: the openai fallback rows and the chatgpt-subscription rows. Incidental upstream drift carried by the regeneration: opencode gains claude-opus-5-5 and grok-4.7, OpenRouter price/context refreshes on unrelated ids, Vercel gemini metadata. No model id was removed (per-file id delta audited).

Changelog

• packages/ai/CHANGELOG.md: Added (first-party rows) + Fixed (OpenRouter rows shipped in 22-4 without any effort ladder).
• packages/coding-agent/CHANGELOG.md: Breaking (recommended/default OpenAI model moves to GPT-6 Sol) + Added (family preset, warning coverage).
• changes.md trackers: packages/ai/changes.md, packages/ai/src/changes.md, packages/coding-agent/src/core/changes.md, prompt-preset/changes.md. node scripts/check-pr-changelog.mjs --base origin/main → PASS.

RED → GREEN evidence

Criterion	Test	RED	GREEN
Catalog rows, pricing tiers, ladder, -fast, family budget	packages/ai/test/gpt-6-family-catalog.test.ts	27 failed / 1 passed (ids absent)	105 passed with the neighbouring Astra/fast/input-cap suites
Map-less xhigh/max inference	same file, map-less block	part of the 27	green
Preset resolution	test/suite/prompt-presets-gpt-6-family.test.ts	19 failed (undefined preset)	156 passed with Astra/#1891/#1968/eval-routing suites
Recommended ladder, provider default, warning	recommended-models-extension, model-resolver, high-reasoning-warning tests	8 failed	189 passed

openai-input-cap-catalog.test.ts exempts exactly the gpt-6-sol @ 400,000 pairing (a deliberate budget equal to the 400k-tier total); a 1,050,000 Sol row still fails it.

Real-CLI QA (senpi-qa channel 3)

bun .agents/skills/senpi-qa/scripts/gpt-6-family-preset-mock-loop.mjs → 24/24. The real CLI in --print mode against a fake Responses server sent model: gpt-6-sol with reasoning.effort: medium and model: gpt-6-luna with reasoning.effort: none (the off level Astra cannot take), and both developer messages carry the GPT-6 sections (## Asynchronous Work, ## Instructions From Files), none of the GPT-5.6 ones, and no "Astra" token. Evidence under local-ignore/qa-evidence/20260923-gpt-6-family-preset-mock-loop/ (gitignored). Sandboxes and the fake server are torn down in-script; zero leftover dirs/processes after the run.

Gates

• bunx biome check --error-on-warnings . → 0 errors
• root bunx tsgo --noEmit → exit 0 (run from the canonical checkout path)
• bun run --cwd packages/ai check:model-data → valid
• test/suite/anthropic-subscription-naming.test.ts → 3 passed (no legacy provider literals introduced)

Not in this PR

• No live inference against gpt-6-sol / gpt-6-luna (the mock loop proves what the harness sends, not that a provider accepts it).
• configuration_update (mid-conversation effort change) stays Astra-only, matching the Codex backend's wire constraint.
• omo-side routing (deep-low, explore/librarian luna rungs) follows in the omo repository after this ships.

---

Summary by cubic

Adds GPT-6 Sol and GPT-6 Luna across the model catalog, effort inference, prompt preset, and defaults, so the whole GPT-6 family is supported rather than Astra alone. gpt-6-sol:medium becomes the recommended and implicit default for the openai and chatgpt-subscription providers; explicitly configured settings are untouched.

Catalog and effort inference

• New gpt-6-sol and gpt-6-luna rows on OpenAI, ChatGPT Subscription, Azure, OpenCode, OpenRouter, Venice, and Vercel, with -fast Priority variants on the two first-party providers.
• Rows carry the published prices with >272k tiers, 128k output, the full none/low/medium/high/xhigh/max effort ladder, and family prompt budgets (Astra 600k unchanged, Sol 400k, Luna 922k).
• Fixes OpenRouter rows that shipped in 22-4 with no effort ladder at all.
• Map-less custom rows for both ids now surface xhigh and max.

Prompt preset and defaults

• All gpt-6-sol / gpt-6-luna ids resolve to the GPT-6 family prompt preset, which keeps its gpt-6-astra name so existing settings pins keep working.
• The high-reasoning warning now covers GPT-6 Sol at xhigh / max; Luna is not covered.

The regeneration also carries incidental upstream drift (opencode gains claude-opus-5-5 and grok-4.7, plus OpenRouter and Vercel refresh), with no model ids removed.

^{Written for commit c0e27cd. Summary will update on new commits.}