Audit Codex model availability and selector defaults for 5.2, 5.3, Spark, and 5.5

[cbusillo]

Problem

code-gpt-5.3-codex-spark / gpt-5.3-codex-spark appears to be drifting from an always-available built-in selector to a limited or unavailable model path.

Current repo behavior still advertises Spark in several user-facing and runtime defaults:

• code-rs/core/src/agent_defaults.rs includes code-gpt-5.3-codex-spark as a built-in agent selector and maps it to --model gpt-5.3-codex-spark.
• code-rs/common/src/model_presets.rs, code-rs/tui/src/bottom_pane/model_selection_view.rs, and code-rs/tui/src/remote_model_presets.rs expose gpt-5.3-codex-spark in model selection paths.
• README.md, docs/agents.md, docs/config.md, and docs/slash-commands.md still document Spark as a default or supported selector.
• code-rs/core/src/codex/streaming.rs uses Spark as the auto-context judge primary model.

Evidence suggesting this may no longer be safe:

• OpenAI's public model index currently lists GPT-5.2-Codex, GPT-5.1-Codex, GPT-5.1-Codex-Max, GPT-5-Codex, and GPT-5.1 Codex mini, but not gpt-5.3-codex-spark: https://platform.openai.com/docs/models
• The Spark launch post describes GPT-5.3-Codex-Spark as a ChatGPT Pro research preview on Cerebras, not necessarily a generally available API model: https://openai.com/index/introducing-gpt-5-3-codex-spark/
• The code already has tests/handling for The requested model 'gpt-5.3-codex-spark' does not exist. in code-rs/core/src/client.rs, which is a strong hint that this failure mode is real.
• Local API probe could not be run from this session because OPENAI_API_KEY is unset.

Suggested Scope

1. Verify current availability for ChatGPT-auth Codex, API-key Codex, and Pro-only capability paths.
2. If Spark is unavailable or account-gated, stop advertising code-gpt-5.3-codex-spark as a default built-in selector for unsupported auth contexts.
3. Replace Spark-dependent internal defaults, especially auto-context judge primary model, with a stable available model or a capability-aware fallback.
4. Update docs and prompt examples so users are not steered toward a selector that fails immediately.
5. Add/adjust tests around selector filtering and model-not-found fallback behavior.

Acceptance Criteria

• Unsupported users do not see or receive code-gpt-5.3-codex-spark as a default agent/model selector.
• Spark remains available only when capability detection confirms the account can use it, or it is removed entirely if no supported path exists.
• Auto-context/model fallback behavior avoids hard failures when Spark is missing.
• Docs match the new selector behavior.

[cbusillo]

Additional model-lifecycle signal to include in this audit:

A user-reported announcement says GPT-5.2 and GPT-5.3-Codex are being sunset from Codex on June 2 when logged in with a ChatGPT account, while GPT-5.5 becomes the default frontier model for free plans. I could not find an official OpenAI page with that exact June 2 Codex wording yet; the strongest public hit is a Reddit/social repost. Treat the date as unverified until we find the source banner/email/release note.

Official adjacent signals found today:

• OpenAI Help Center says GPT-5.5 Instant is the default for all logged-in ChatGPT users, and GPT-5.5 is available to all ChatGPT tiers: https://help.openai.com/en/articles/11909943-gpt-52-in-chatgpt
• The same page says GPT-5.2 Thinking remains available under Legacy Models for 90 days after GPT-5.5 Thinking launch for Plus and Pro users.
• OpenAI API docs still list gpt-5.3-codex as an API model with Responses/Chat Completions support, which matches the claim that API availability may continue even if ChatGPT-account Codex access changes: https://developers.openai.com/api/docs/models/gpt-5.3-codex
• Codex pricing still lists GPT-5.3-Codex-Spark as a Pro research preview, reinforcing that Spark should not be broadly advertised as a default selector: https://chatgpt.com/codex/pricing/

We do use GPT-5.2 in current repo paths. Relevant examples:

• code-rs/core/src/agent_defaults.rs keeps code-gpt-5.2, code-gpt-5.2-codex, gpt-5.2, and gpt-5.2-codex as upgrade/alias paths.
• code-rs/common/src/model_presets.rs exposes gpt-5.2 and gpt-5.2-codex presets.
• code-rs/core/src/model_family.rs includes gpt-5.2-codex_prompt.md and maps gpt-5.2* families.
• code-rs/core/src/config/validation.rs upgrades older 5.2 aliases toward 5.4 in some contexts.
• code-rs/core/src/reasoning.rs clamps/handles gpt-5.2* reasoning settings.
• code-rs/tui/src/model_migration.rs still has GPT-5.2/GPT-5.2 Codex upgrade copy.
• code-rs/tui/src/lib.rs has auth-mode filtering for gpt-5.2-codex.
• code-rs/protocol/src/config_types.rs test/default fixtures include gpt-5.2-codex.

Implication for scope: this should be broader than Spark. We need an availability matrix by auth mode:

• ChatGPT account / free
• ChatGPT account / paid tiers
• ChatGPT Pro with Spark capability
• API key
• cloud agents, if they have separate model availability

Likely product behavior after audit:

• Prefer GPT-5.5 as the default frontier ChatGPT-account path if Codex has moved there.
• Keep API-only 5.2/5.3 selectors only where auth mode is API key or explicit override.
• Keep legacy aliases for config migration, but do not surface retired ChatGPT-account models as normal choices.
• Gate Spark behind explicit capability detection or remove it from defaults.