Add GPT-5.4 and GPT-5.5 model support

[Mindrocket42]

Summary

• Adds GPT-5.5 and GPT-5.4 model support (1.05M context, 128K output), plus new first-class GPT-5.4-mini / GPT-5.4-nano presets (400K context, 128K output), to MODEL_MAP and both config/opencode-modern.json (variants) and config/opencode-legacy.json (flat presets).
• Important: confirmed via testing against the real ChatGPT backend that, unlike GPT-5.1/5.2, there is no separate gpt-5.5-codex/gpt-5.4-codex model from OpenAI's side — sending those literal strings gets rejected with "model is not supported when using Codex with a ChatGPT account". So the OA-codex presets (GPT 5.5 (OA-codex), GPT 5.4 (OA-codex)) are our own config-side choice — same underlying gpt-5.5/gpt-5.4 model as the OA-chat presets, just with different system instructions and reasoning-effort defaults applied. resolveWireModel() in model-map.ts sends the real gpt-5.4/gpt-5.5 string on the wire regardless of which preset is selected.
• Generalizes getReasoningConfig in lib/request/request-transformer.ts to derive xhigh/none reasoning support from the parsed GPT-5.x point-release number instead of a hardcoded per-version boolean flag, since the same schema (introduced with 5.2) carries forward to later point releases. This means a future 5.x point release following the same schema only needs new MODEL_MAP/config entries, not a change to the reasoning logic.
• Mini/nano default to low reasoning effort (rather than inheriting the full model's high default) since they're a new lightweight tier, but still allow explicit xhigh opt-in.
• Bumps to 4.5.0 (package.json, CHANGELOG, OAuth success page banner) and updates the README model list/preset count (22 → 50).

Note: getModelFamily in lib/prompts/codex.ts (which selects the upstream Codex system-prompt file to fetch) was intentionally left unchanged — GPT-5.4/5.5 general purpose falls back to the gpt-5.1 prompt family and codex variants fall back to the generic codex family, since I couldn't confirm whether dedicated prompt files exist yet in openai/codex for these releases. Worth revisiting if you know those files exist upstream.

Test plan

• npx tsc --noEmit passes
• npx vitest run — all pre-existing tests pass (5 pre-existing failures in install-script.test.ts are unrelated Windows HOME/homedir() environment issues, reproducible on main without this change)
• Added unit tests in test/request-transformer.test.ts covering GPT-5.4/5.5 normalization, reasoning-effort behavior, and the wire-model resolution for the OA-codex presets
• Validated both config JSON templates parse and model counts match expectations
• Verified against the real ChatGPT/Codex backend — gpt-5.4, gpt-5.4-mini, and gpt-5.5 (OA-chat) confirmed working; the -codex wire-model rejection was caught and fixed this way during that testing

[Mindrocket42]

Update: confirmed via real testing against the ChatGPT backend that `gpt-5.4-codex`/`gpt-5.5-codex` are not valid model names — the backend rejects them with:
```
Bad Request: {"detail":"The 'gpt-5.5-codex' model is not supported when using Codex with a ChatGPT account."}
```
Unlike GPT-5.1/5.2, GPT-5.4/5.5 apparently have no distinct `-codex` backend deployment — it's the same underlying model either way.

Pushed a follow-up commit that decouples our internal "family key" (used to pick codex-flavored system instructions + reasoning-effort defaults) from the literal wire model name via a new `resolveWireModel()` in `model-map.ts`. The `GPT 5.4 (OA-codex)`/`GPT 5.5 (OA-codex)` presets now correctly send `gpt-5.4`/`gpt-5.5` as the model field while keeping codex-oriented prompt/reasoning behavior. Also relabeled the config entries from `(OAuth)`/`... Codex (OAuth)` to `(OA-chat)`/`(OA-codex)` to make clear these are prompt presets of the same model, not distinct deployments.

[Mindrocket42]

Withdrawing this PR for now — testing surfaced enough inconsistencies in the GPT-5.4/5.5 model-naming assumptions that this needs more verification against the real ChatGPT/Codex backend before it's ready. Will reopen (or resubmit) once that's sorted.

[Mindrocket42]

Reopening — confirmed this branch's exact code and config are what's synced and running locally: clean working tree at 892515f, dist rebuilt from it, and the global config's model list verified byte-identical to config/opencode-modern.json in this commit. gpt-5.4, gpt-5.4-mini, and gpt-5.5 (OA-chat) confirmed working against the real ChatGPT backend; the gpt-5.4-codex/gpt-5.5-codex wire-model rejection was caught and fixed as part of this same testing.

[o7ux]

the goat

[o7ux]

codex oauth'd gpt5.4 and gpt5.5 are working as expected with all variants.
chatgpt oauth'd models returning "model is not supported when using Codex with a ChatGPT account". however, same rejection occurs verbatim in codex-cli when calling those models.