feat(app): add GPT-6 Sol and GPT-6 Luna to the Codex model picker

[chphch]

OpenAI released GPT-6 Sol and GPT-6 Luna to Codex on 22 September, and the Codex picker still offers only GPT-6 Astra from that generation. This adds two rows, gpt-6-sol and gpt-6-luna, directly under Astra, plus their entries in CODEX_EFFORTS_BY_MODEL. Existing rows and the Codex default (gpt-5.6-sol) are unchanged.

Where the values come from

Codex's own registry, codex-rs/models-manager/models.json on openai/codex main:

field	gpt-6-sol	gpt-6-luna
visibility	list	list
minimal_client_version	0.155.0	0.155.0
supported_reasoning_levels	low … ultra (6)	low … max (5, no ultra)

Sol publishes ultra and Luna does not — the same split GPT-5.6 Sol and GPT-5.6 Luna already have — so each row gets exactly the levels its registry entry lists, not a shared list.

Both are in one PR because the rows sit on adjacent lines: two separate PRs would conflict with each other the moment either merged.

Caveat: access is still rolling out

On codex 0.155.1 with a ChatGPT account, codex exec -m gpt-6-sol and -m gpt-6-luna both currently get The '<model>' model is not supported when using Codex with a ChatGPT account from the server — the same gate Astra hit at launch (see #1748). The rows are listed anyway for the reason Astra was: both models are public in Codex's registry, and the rows start working the moment an account gets access, with no further app release.

Proof

Screenshot in the comments below: the /new Codex model picker with both new rows under Astra, and the effort picker for each, read back from the live DOM as well as screenshotted. vitest run sources/components/modelModeOptions.test.ts passes with the assertions updated alongside, and tsc --noEmit is clean.

🤖 Generated with Claude Code

[chphch]

Proof — the Codex picker offers GPT-6 Sol

Captured locally from this PR's commit: a throwaway standalone happy-server (PGlite) + Expo web, authenticated through the dev_token URL bypass and driven headlessly with Playwright.

Read back from the live DOM:

model:  ["GPT-6 Astra","GPT-6 Sol","GPT-5.6 Sol","GPT-5.6 Terra","GPT-5.6 Luna"]
effort: ["Low","Medium","High","xHigh","Max","Ultra"]   (with GPT-6 Sol selected)

How Codex was reachable in a daemon-less harness, since it affects what this proves: the agent picker offers nothing until a machine is selected and online, so the capture registered a machine via POST /v1/machines and held a presence-only socket that sends the daemon's machine-alive heartbeat. No session was created. This proves the picker's contents and the per-model effort levels — not that a session can start on gpt-6-sol, which the rollout caveat in the description covers.

[chphch]

Proof — updated for GPT-6 Luna (supersedes the Sol-only screenshot above)

Re-captured from the current head, same harness as before (standalone happy-server + Expo web, dev-token bypass, headless Playwright, a presence-only machine so Codex is selectable; no session created).

Read back from the live DOM:

model:              ["GPT-6 Astra","GPT-6 Sol","GPT-6 Luna","GPT-5.6 Sol","GPT-5.6 Terra","GPT-5.6 Luna"]
effort, GPT-6 Sol:  ["Low","Medium","High","xHigh","Max","Ultra"]
effort, GPT-6 Luna: ["Low","Medium","High","xHigh","Max"]      (no "Ultra" node anywhere in the DOM)

The Luna frame is the one that matters most: it shows each row takes its own registry entry's levels rather than a shared list.