yoi/.yoi/tickets/closed/20260531-003743-codex-gpt55-effective-context-window/item.md

id	slug	title	status	kind	priority	labels	created_at	updated_at	assignee
20260531-003743-codex-gpt55-effective-context-window	codex-gpt55-effective-context-window	Provider: make codex gpt-5.5 context window effective	closed	task	P2	provider models catalog	2026-05-31T00:37:43Z	2026-05-31T01:58:24Z	null

Background

refresh-builtin-model-catalog updated the codex-oauth/gpt-5.5 builtin entry to record OpenAI's advertised 1.05M context window while retaining max_context_window = 272000 as an effective clamp. That made the catalog entry technically accurate to both OpenAI docs and observed Codex OAuth behavior, but it also made the model metadata harder to understand.

For a provider-specific builtin entry, context_window should represent the effective window for that provider route. codex-oauth/gpt-5.5 is not the generic OpenAI API route; it is the Codex OAuth / ChatGPT backend route. If that route effectively fails above ~272k input tokens, then the builtin catalog should say context_window = 272000 directly and explain why in a comment.

Requirements

• Change the builtin codex-oauth/gpt-5.5 model entry so context_window is the effective Codex OAuth route limit (272000).
• Remove max_context_window = 272000 from that builtin entry.
• Add a nearby TOML comment explaining that OpenAI documents GPT-5.5 with a 1.05M context window, but Codex OAuth / ChatGPT backend access has been observed/known to be effectively limited around 272k, so the provider-specific builtin entry records the effective limit.
• Keep the default profile on codex-oauth/gpt-5.5.
• Preserve the max_context_window manifest/config field and its code/tests if still useful for explicit inline overrides or future providers; do not remove the field globally in this ticket unless doing so is clearly tiny and mechanically safe.
• Update provider/manifest tests that currently expect catalog-level max_context_window clamp for builtin gpt-5.5.
• Do not change OpenRouter openai/gpt-5.5 unless there is separate evidence for that route's effective limit; this ticket is about codex-oauth.

Non-goals

• Re-refreshing the full model catalog.
• Changing provider definitions or auth behavior.
• Changing compaction/pruning algorithms beyond tests that derive known model context.
• Removing max_context_window from public manifest/config types wholesale.

Acceptance criteria

• resources/models/builtin.toml has codex-oauth/gpt-5.5 with context_window = 272000 and no max_context_window field.
• The builtin catalog comment makes the 1.05M-vs-272k route distinction explicit.
• Tests no longer rely on catalog-level max_context_window for codex-oauth/gpt-5.5.
• Inline max_context_window clamp behavior remains tested if the field remains supported.
• cargo fmt --check, focused provider/manifest tests, ./tickets.sh doctor, and git diff --check pass.