Configure Codex

Account and model source

Choose Official, Model provider, or Manual configuration as the model source, and select ChatGPT or No account independently. Sign in through Codex; choosing ChatGPT in OcHub does not sign in or verify a subscription.

1. Add the third-party URL, API key, protocol, and models under Model providers.
2. Add a Codex connection, choose that provider, and select ChatGPT to retain an existing real login.
3. Choose a model and leave context management on Auto.
4. Review the configuration, save and switch the connection, then start a new Codex session.

A provider connection keeps upstream credentials in the supplier. Manual connections use the address and API key entered in the connection. Keeping ChatGPT login alongside an independent API key does not qualify for experimental context management.

Save stores a Codex draft without rewriting the live connection files. Save and activate saves the draft and switches to it. If the currently active connection changed on disk, OcHub first shows the same version diff and asks whether to save or revert those live edits. Check configuration checks local configuration requirements only; it sends no inference request and does not verify subscriptions or backend capabilities. Remote workspaces require an updated node supporting draft operations.

Virtual login is under Advanced for compatibility with clients without a real account. It grants no subscription or official quota and cannot replace real credentials.

Context management

Pending Codex verification is a configuration state, not proof of backend availability. Independent API-key connections do not enable this account extension.

For a distribution server implementing History / Notes, select ChatGPT, bind a supplier with exactly one Responses channel, and choose On. OcHub preserves real login credentials and explicitly writes [features.token_budget] enabled = true and use_history_notes_extension = true. This uses Codex’s explicit configuration path without editing ID-token plan claims. Codex must include the extension and a known model context window; managed feature restrictions still apply. Remote control continues using the official chatgpt_base_url, which OcHub does not redirect. Provider capabilities and official pairing still require live verification.

Keep the same distribution API key for a history session; start a new session after changing keys.

Automatic activation requirements

The corresponding Codex source checks all of these in apply_experimental_context:

• [features.context_management] experimental_mode = true.
• Exact provider name OpenAI.
• No base URL, or a URL ending in /backend-api/codex after removing trailing slashes.
• requires_openai_auth = true.
• No env_key, experimental_bearer_token, auth, or aws provider fields; empty strings still count as present.
• Chatgpt authentication with a Plus, Pro, or ProLite plan.
• Token Budget can be enabled and is not pinned off by managed requirements.

Passing this check enables Token Budget and use_history_notes_extension. Actual use also requires a known model context window, an installed History / Notes extension, and working history ingestion, retrieval, and notes services. Explicit Token Budget configuration and model defaults are separate activation paths.

History and compaction are separate

OcHub forwards the ten History / Notes operations, including thread_hint, under its scoped /backend-api/codex/alpha/ route to the supplier’s Responses endpoint sibling alpha/ path. Requests need no model field. Opaque argument bytes, encrypted results, and upstream statuses are preserved; real OAuth and account headers are replaced with supplier authentication. Mixed or multiple enabled channels are rejected for these operations, and writes are never retried on another channel.

The distribution server must implement history ingestion, storage, retrieval, notes, user ownership checks, and persistent session-to-account affinity. OcHub does not supply that backend or verify it when saving. Keep the supplier unchanged for existing sessions. Auto remains conservative for gateways. The older 0.5.16 claim of complete history support from compaction alone was incorrect.

/responses/compact only provides remote compaction. OcHub forwards /v1/responses/compact and the Codex backend equivalent to Responses upstreams that support it. Chat or Messages conversion cannot substitute for this capability.

Model capability declarations

Under Model providers → Edit → Models, open Model capabilities for a selected model.

These settings are marked User declared · Unverified. Fetching a model list does not verify Lite, compaction, or history services. Declarations are scoped to the supplier and upstream model; aliases use the mapped upstream model’s declarations.

Lite changes Responses request behavior. It is neither an authentication mode nor a context-management prerequisite. Lite requests require native Responses channels and cannot fall back to Chat / Messages conversion.

Route isolation

New real-login connections use an address such as:

http://127.0.0.1:4180/connection/rd-…/backend-api/codex

The URL’s local gateway credential identifies the supplier independently of rotating account tokens. Account credentials remain in auth.json; OcHub uses the supplier API key upstream and removes the local routing credential. Different suppliers no longer share a single OAuth route binding.

The URL contains a local credential and should not be shared publicly. The listener remains loopback-only. Route authentication does not verify a ChatGPT subscription. Disabling its key rejects the connection.

Legacy URLs without /connection/… still support the manually configured global OAuth binding. Save and activate the connection again to migrate.

Local history, MCP, and skills

Local history unification and migration manage session files, not the experimental History / Notes backend. Exit Codex and retain backups before changing history locations. Enable Codex in the MCP or Skills page, sync, and restart sessions after configuration-directory changes.

See Codex advanced configuration and Model providers.