feat: add metadata-only app/read

[stevenlee-oai]

Codex Thread 019f39de-e02c-79d0-9263-2447105cae69

Why

M3 needs a fast, fresh, consistent way for app-server clients to read metadata for selected connector IDs without rebuilding connector runtime state or reloading MCP tools. This adds that metadata-only seam while preserving existing app/list and app/list/updated behavior.

What changed

• Adds ConnectorMetadataStore, partitioned by ChatGPT backend base URL, account, ChatGPT user, and workspace classification, using the existing connector-directory TTL.
• Adds v2 app/read { appIds } -> { apps, missingAppIds }.
• Enforces a 100-ID input cap, de-duplicates repeated IDs in first-request order, and returns both apps and misses in that order.
• Reads fresh cache entries first, then makes at most one batch request for missing or expired IDs with include_actions=false and include_model_descriptions=false.
• Treats unknown or unauthorized IDs as partial misses. Backend or transport failures return an RPC error without replacing cached records.
• Keeps the metadata shape limited to id, name, description, distributionChannel, branding, appMetadata, labels, and nullable installUrl; it excludes runtime state, MCP tools, actions, model descriptions, and icons.
• Parses optional backend install_url so mixed-version responses that omit it remain safe.

Public backend dependency

Depends on openai/openai#1097857, commit 9593a5ca75201. That change leaves the existing internal /v2/connectors/batch source, tests, and OpenAPI unchanged and adds only the authenticated public projection at /public/connectors/batch.

Codex calls POST https://chatgpt.com/backend-api/ps/connectors/batch. The ChatGPT gateway validates Identity Edge app_v2, injects the actor biscuit, and rewrites /ps/* to Plugin Service /public/*. Plugin Service requires an authenticated user, forces the trusted ChatGPTAuthorizationSchema, and delegates to the existing resolver visibility logic. Codex continues sending the captured auth headers: Authorization, ChatGPT-Account-ID when applicable, optional X-OpenAI-Fedramp, OAI-Product-Sku: codex, and JSON content type; no extra auth-schema field is added.

The auth snapshot also defines the cache scope before the request is awaited, so late responses can commit only to the backend/account/workspace key under which they were requested.

Scope and non-goals

• No behavior changes to app/list or app/list/updated.
• No connector-tool reloads or runtime-state reporting from app/read.
• No M4/M5 work or unrelated refactors.
• No new codex.apps.read.duration_ms metric; the current app-server metric seam appears to depend on M1 and is intentionally out of scope.

Validation

• just fmt
• just test -p codex-connectors
• just test -p codex-app-server-protocol
• just test -p codex-chatgpt
• just test -p codex-app-server app_read
• just fix -p codex-connectors -p codex-chatgpt -p codex-app-server-protocol -p codex-app-server
• git diff --check origin/main...HEAD

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 03433b693e

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]

Gate experimental app/read behind capability

Because the new AppsReadParams/response types are documented as EXPERIMENTAL, adding the variant here without #[experimental("app/read")] makes ClientRequest::experimental_reason() return None for app/read. The app-server will therefore accept this experimental surface from clients that did not initialize with experimentalApi, and it will be omitted from the experimental client-method metadata; add the method annotation here (or make the API explicitly stable).

AGENTS.md reference: AGENTS.md:L286-L287

Useful? React with 👍 / 👎.