Codex Cloud and ChatGPT Work

HypAware can discover and import selected categories of hosted chats from a local collector, using the existing local ChatGPT login. It does not need a browser extension or an agent to export history. This integration is in beta: it uses an observed hosted interface that may change. Hosted collection is a separate choice from local Codex recording. The Codex desktop app is not required; the collector uses the Codex CLI and runs while the collector computer and HypAware daemon are running.

By default only codex_cloud coding chats are collected. Three origin categories can be selected independently, alone or together:

The origin identifies the originating client, not a restriction to where you later view the chat. These labels are observed backend metadata, not a documented third-party contract. Web/mobile collection can include everyday personal assistant conversations. Delegated codex_work_cca, hidden/unknown origins and all unselected categories are excluded, including when you supply explicit IDs.

Requirements: Node 22.12 or newer, the @hypaware/codex-cloud plugin enabled in HypAware, an installed codex executable on the collector's PATH, and a file-backed ChatGPT login in $CODEX_HOME/auth.json (default ~/.codex/auth.json). The collector adds the running Node executable's directory to its helper PATH, so a Codex installed beside Node remains discoverable under launchd or systemd. The collector checks the effective Codex credential-store setting and rejects keyring, auto, ephemeral and API-key authentication. It does not change that setting.

Choose during setup

hyp setup offers two hosted choices:

• Codex Cloud (beta) is a separate row from local Codex. On first setup it is checked when Codex is detected and included in Record everything (or Record and sync everything).
• ChatGPT Work (beta) bundles Work web and Work mobile. It starts unchecked and is only added through Customize.

Choosing either hosted row enables its collector without enabling local Codex recording. Both use the existing file-backed Codex ChatGPT login. If that login is unavailable or unsupported, setup reports the incomplete step and continues; sign in through Codex and rerun setup to select the desired hosted rows.

Reconfigure shows the choices saved by the hosted plugin. Clearing a hosted row stops future collection for that category. Keeping a row preserves existing thread IDs, history bounds and a narrower web-only or mobile-only CLI grant. Cancel and Back do not save changes. Unattended setup does not authenticate or grant hosted access.

Upgrading an earlier install

Earlier builds ran hosted collection inside the @hypaware/codex plugin and kept its grant at ~/.hyp/hypaware/plugins/@hypaware/codex/cloud-connection.json (under $HYP_HOME when set). The collector now lives in @hypaware/codex-cloud, which reads only its own grant, so after upgrading hosted collection stops until you choose it again. Rerun setup and select Codex Cloud (beta), ChatGPT Work (beta) or both:

hyp setup

The old grant is left in place. It is not read, migrated or deleted, and you can remove that file once the new selection is saved. A narrow scope saved at the old location (specific chats, only Work web or only Work mobile, or a --from-now bound) is not migrated either: rerunning setup re-selects hosted collection with a new account-wide grant. To narrow it again, run hyp codex-cloud connect with the options you want after setup.

Connect and import

For automatic account-wide discovery, using your existing history window:

hyp codex-cloud connect --all

Select a different category or a comma-separated combination explicitly:

hyp codex-cloud connect --all --origins codex_work_web
hyp codex-cloud connect --all --origins codex_work_mobile
hyp codex-cloud connect --all --origins codex_cloud,codex_work_web,codex_work_mobile

Add --from-now only to exclude older activity. To collect specific chats instead, use UUIDs from their hosted URLs:

hyp codex-cloud connect <thread-id> [thread-id...] [--origins <origin,...>]
hyp codex-cloud status
hyp backfill codex-cloud --dry-run --json
hyp backfill codex-cloud --json

connect replaces the previous grant and binds it to the current ChatGPT user and account. Explicit selection supports up to 20 chats; --all discovers active and archived chats within the selected origins automatically. It does not import immediately. The daemon checks connected chats every five minutes. Restart an older daemon after upgrading. Imported records follow the existing HypAware export settings, including prompts and tool outputs from any selected Work/dot chats. Status shows the active origin selection. No credentials are sent to HypAware servers.

The import supports the normal backfill --since, --until and retention options. Scheduled runs also honor the hosted plugin's backfill.window_days when configured. Repeat imports deduplicate against recorded part IDs.

hyp query sql "select count(*) from ai_gateway_messages where entrypoint = 'codex-cloud'"
hyp codex-cloud disconnect

Disconnect prevents future collection; it does not delete recorded data. hyp session ignore <thread-id> persists a session exclusion that both manual imports and scheduled sweeps reload at the start of each run. An import already in flight uses its run-start snapshot. Local hyp detach codex is independent of hosted recording. Use setup to clear a hosted row, or hyp codex-cloud disconnect to stop both. None of these deletes prior rows.

Capture and limits

This first version imports completed user text, assistant text, command calls and their full outputs, and MCP calls with their results, including structured content. Tool calls and results occupy separate rows. Reasoning content is excluded. Unfinished items are deferred, and unsupported items or content parts are reported as skipped. Image inputs, file-change items and other native item types are not yet mapped.

Usage is unknown when the history response lacks it. This collector neither subscribes to live usage events nor estimates tokens. Thread-level model metadata is retained in attributes, not presented as a per-request model. The native originator is retained in attributes.codex.originator; all rows from this collector use entrypoint = codex-cloud. client_name is codex-cloud for coding chats and chatgpt-work for Work chats, so their source export policies are independent. Cloud filesystem paths are metadata: local .hypignore rules cannot classify the remote /workspace directory.

The collector opens one authenticated connection per pass, with serial requests, a two-minute pass deadline, a 64 MiB received-message budget and at most 20 pages of 100 items per chat. If a chat exceeds the limits or pagination fails, that chat is not partially imported. Earlier successfully imported chats remain. Limits are reported as errors rather than silently truncating history. Large chats beyond these bounds currently require a later implementation extension.

Automatic discovery permits at most 20 list pages per archive state (up to 4000 metadata entries across all origins), filters the selected origins locally, and reads up to 20 eligible chats per sweep. Unselected chat messages are not fetched. Batches rotate, with newly discovered chats prioritized. Existing chats are revisited; updates can wait a full rotation. Restarting the daemon resets rotation. Chat-specific read failures are reported and skipped for that pass; other chats continue. Discovery or connection failures stop the pass visibly.

Authentication and troubleshooting

HypAware keeps the access token in memory. Near expiry, it asks the installed Codex app-server to renew the existing login; Codex updates its own credential store. There is no separate HypAware refresh-token copy. A dry run refuses a needed renewal rather than explicitly rotating tokens. Generic network failures do not trigger renewal loops.

If the account changes, collection stops until you reconnect the desired chats. If credentials are revoked, sign in through Codex normally and reconnect. HypAware never automatically logs you out or deletes your credentials.

Failures emit codex.cloud.failed with a fixed error_kind and the failed step. Account-mode chat failures emit codex.cloud.chat_failed and the cloud_chat_failed progress event. Other chats continue, then the pass reports some_chats_failed with a failed-chat count and a nonzero exit status. Discovery logs discovered, eligible, selected and deferred counts. Network errors and upstream rejection bodies are not logged. Success emits codex.cloud.completed; hyp backfill ... --json reports actual rows written.

The hosted WebSocket interface is an implementation observed and tested with a personal account, not a documented third-party hosted API contract. Upstream support for commercial integrations remains unresolved. Simultaneous refreshes by independent Codex processes and natural-expiry recovery still need broader acceptance testing. This implementation does not claim cross-process locking.

Validation and remaining limits

Tested on macOS with Node 24.2.0, Codex CLI 0.159.0 and a personal ChatGPT login: complete long command output and MCP results, repeat-import deduplication, growing-chat capture through launchd, restart, pagination, recovery after a connection drop, Codex-managed refresh with a simulated near-expiry clock, strict originator discovery and central export. The filtered live inventory contained four Codex Cloud chats, and the remote query matched 494 local rows.

Linux service behavior, natural token expiry, simultaneous refresh by separate Codex processes and managed-workspace accounts remain unverified. Token counts are unavailable from the persisted history responses tested so far.

Design: LLP 0511.