core/app-server: preflight image modality mismatch with typed error

[ccy-oai]

Summary

• Add an authoritative core preflight that blocks turns when the selected model is text-only and the final outgoing payload contains image inputs (including history/tool outputs).
• Introduce a typed non-retryable modality error surface (unsupportedInputModality) through protocol and app-server v2.
• Regenerate app-server protocol schema/TS fixtures for downstream clients.
• Update docs/model-update.md with the implemented contract and client handoff details.

Rationale (from spec findings)

• UI-only gating cannot reliably predict final payload after history normalization/compaction/tool outputs.
• Core has authoritative visibility into what is actually sent, so correctness should be enforced there.
• Typed errors let clients handle this path explicitly instead of generic retries/fallbacks.

Scope

• codex-rs/core: preflight check + error mapping.
• codex-rs/protocol: image detection helpers + new error variant.
• codex-rs/app-server-protocol: v2 error variant + generated schema/types.
• docs/model-update.md: implementation-focused spec notes for consumer clients.

Trade-offs

• Added schema fixture churn from the new error variant in protocol exports.
• Preflight currently checks the assembled outgoing payload; future optimization can cache modality summary state, but this keeps behavior correct today.

Client follow-up

• Consumers should treat codexErrorInfo.unsupportedInputModality as non-retryable.
• Show remediation UX (switch to image-capable model or fork/reset context).
• Keep existing composer gating as best-effort UX; rely on server preflight for source-of-truth enforcement.

Testing

• cargo fmt --all
• cargo check -p codex-protocol -p codex-core -p codex-app-server-protocol -p codex-app-server
• cargo test -p codex-app-server-protocol unsupported_input_modality -- --nocapture

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f9af4d9576

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Apply modality preflight beyond sampling requests

This new preflight only runs in run_sampling_request, so any other model calls that build a Prompt directly (e.g., compaction via run_compact_task_inner/drain_to_completed in core/src/compact.rs) still send image-bearing history to text-only models. In practice, a thread with images that triggers auto‑compaction (or a manual compact) will bypass the new typed error and still hit provider 400s. If the goal is an authoritative guard for all outgoing payloads, the same check needs to be enforced in the compaction path (or centralized before any prompt is streamed).

Useful? React with 👍 / 👎.