Preset Skill Codex

codex preset revision

Use this child for the named built-in codex preset — the first-run default template. codexPreset in tui/internal/preset/preset.go ships provider codex, gpt-5.6-sol, https://chatgpt.com/backend-api/codex, ChatGPT OAuth with an empty api_key_env, thinking xhigh, web_search on DuckDuckGo, and vision inheriting the agent's own (Codex) provider. It is not the OpenAI API preset: an OpenAI API key or an OpenAI-compatible endpoint uses the openai family (reference/openai/SKILL.md), and several ChatGPT accounts behind one endpoint use an external pool such as sub2api / subs-pool through that same openai family.

Template-specific settings

Since 2026-09-30 the picker's model lineup is fetched, not hand-curated: see "Public model-directory picker" below before assuming the old docs-page-comparison workflow still applies to the live list. Never inspect or print OAuth token contents.

TUI surfaces to revise

Start at codexPreset in tui/internal/preset/preset.go, which pins the constructor/default to config.DefaultCodexModelOptions()[0] (tui/internal/config/codex_public_models.go) — currently gpt-5.6-sol: public Astra documentation is positive, but exact authenticated OAuth-route availability is not, so the offline default does not silently promote Astra. Revise codexThinkingOptions in tui/internal/tui/preset_editor.go when Codex's reasoning choices change; the service tier row offers normal/fast (fast is sent as priority). Preserve the /codex base_url suffix, empty api_key_env, OAuth identity (codex_auth_path account binding), and explicit xhigh default. Follow the Codex-specific checklist in tui/internal/tui/SKILL.md. The latest-two-generation rule in tui/CONTRACT.md no longer governs Codex — see that contract's "Codex exception" and the next section — but still governs Claude Code's CLI aliases.

Public model-directory picker

Codex's model row is per-editor state (PresetEditorModel.codexModels), not a providerModels map entry — there is no static list to edit when a new model ships. On preset-editor entry, config.RefreshCodexPublicModels (tui/internal/config/codex_public_models.go) fetches the fixed public directory https://raw.githubusercontent.com/openai/codex/main/codex-rs/models-manager/models.json (bounded 5s/2MiB, last-good cache under <globalDir>/cache/, offline static fallback on any failure) and retains entries whose display_name starts with the exact case-sensitive prefix GPT — no numeric-generation, visibility, or account-entitlement filter. The retained display_name is the displayed label; the entry's slug is what gets persisted into manifest.llm.model. This is public metadata, not proof the current account can use a listed model — treat every entry as a suggestion and let the actual run be the entitlement check. A saved/custom id absent from the live/cached list keeps rendering (hollow) and keeps working; pressing c on the model row opens a free-text custom entry for any account-specific or brand-new id. A completed refresh never changes a saved preset's model, an agent's active preset, or its thinking level, and never gates Save. To inspect live OAuth quota instead, complete the app-server initialize handshake, send account/rateLimits/read with structurally null params, and optionally observe account/rateLimits/updated; read usedPercent, windowDurationMins, and resetsAt without exposing secrets.

Reviewed deterministic revision

Prepare an evidence-bound manifest and explicit input, then run lingtai-tui presets revise --manifest PATH --input PATH --mode dry-run|check|apply [--output-dir PATH]. Review the JSON plan, use dry-run/check before apply, and apply only to a new explicit output directory. revision.go validates hashes, route bindings, and evidence and preserves unowned bytes. The constructor/default stays at gpt-5.6-sol behind the exact-route availability gate.

Maintenance: If the relevant TUI preset/page is revised, check whether this sub-skill also needs revision and, if so, include it in the same PR.

Operations

For save, endpoint/capability, availability, activation/refresh, or troubleshooting, use the five shared operation children under reference/operations/; for live OAuth quota/rate limits, use reference/operations/endpoint-capabilities/SKILL.md and never expose auth paths or tokens. This child owns Codex preset facts.

Purchased Codex credits

In either the terminal or desktop preset editor, select codex and turn Paid Codex credits On to allow requests after included ChatGPT usage runs out. The shared field is manifest.llm.codex_allow_credits: true (JSON boolean). It defaults Off, including for existing presets; Off omits the field. Saved presets retain the choice. Switching away from Codex clears it. This is for purchased Codex credits on the signed-in account, not an OpenAI API balance, and requires a kernel that implements the option.

The kernel checks included usage before requests with the option Off. If usage is exhausted or unverifiable, that turn stops; retry after reset or service recovery, or explicitly opt in. On uses the same Codex connection and lets OpenAI decide allowance, credits, debit, and spending limits. The local check cannot guarantee zero charges when usage changes during a request, and does not prioritize credits ahead of included allowance. External pools through openai manage their own billing policy.