Greptile Summary

This PR fixes a forward-compatibility bug where openai-codex/gpt-5.4 was incorrectly inheriting the legacy 272 K context window from the gpt-5.3-codex / gpt-5.2-codex templates instead of using GPT-5.4's own 1 M context window. The fix is clean and well-targeted.

Changes:

• model-forward-compat.ts: Adds OPENAI_CODEX_GPT_54_CONTEXT_TOKENS = 1_050_000 and OPENAI_CODEX_GPT_54_MAX_TOKENS = 128_000 constants, and introduces a patch object that is spread over the cloned template (and also drives the no-template fallback path) only for the gpt-5.4 case. Spreading undefined in the non-5.4 branches is a valid JS/TS no-op, so gpt-5.3-codex behavior is unchanged.
• model-compat.test.ts, model.test-harness.ts, list.list-command.forward-compat.test.ts: Test expectations updated from 272_000 to 1_050_000 to match the corrected metadata.

No logic errors or regressions were found. The change is minimal and surgical—two new constants plus a three-property patch object—with all affected code paths covered by updated unit tests.

Confidence Score: 5/5

• This PR is safe to merge — it is a narrow, well-tested metadata correction with no functional side-effects on other model resolvers.
• The change is minimal and surgical: two new constants + a three-property patch object. All affected code paths are covered by updated unit tests. No existing behavior for other model IDs (gpt-5.3-codex, gpt-5.2-codex) is altered. The patch approach elegantly uses falsy checks (patch?.contextWindow ?? DEFAULT_CONTEXT_TOKENS) to ensure both template and fallback code paths work correctly.
• No files require special attention.

_{Last reviewed commit: 0d21607}

Merged via squash.

• Prepared head SHA: c410207
• Merge commit: f624391

Thanks @yuweuii!