Next Question

Next Question suggests one follow-up question below a completed assistant reply. On desktop and web, click the question or Use in composer to fill the original conversation input. An existing draft is read on the current device and preserved, with the suggestion appended after a blank line. Repeating the click does not insert the same suggestion twice. Review or edit the draft and send it normally. Edit opens an inline editor with an explicit Send question button, including on mobile; Copy uses your device’s clipboard and Dismiss hides the suggestion on that client. Inserting text never sends a message.

The plugin works with the standard Paseo client. It adds a timeline row after a successful top-level agent turn; generation starts when a connected client mounts that row. A mounted row can generate while the app is behind another window. For an older conversation or after a daemon restart, the /next-question command adds a row for the latest completed reply. Suggestions are cached per reply and shared across clients. A new turn cancels obsolete generation; stale replies cannot be sent through the plugin. Filling rechecks that the suggestion belongs to the latest completed reply. The input adapter targets a single visible, editable composer for the same host and agent. It relies on private Paseo 0.13 UI details and may need updating after app changes. If unsupported or ambiguous, it leaves the input unchanged and offers Copy/Edit. Composer text is used only on the current device and is not sent to the prediction model. Mobile uses Copy/Edit. There is no Tab-to-accept or composer placeholder integration.

Requires Paseo 0.13.x and a current Codex CLI on the daemon's PATH, already signed in or configured with a working provider. The CLI must support exec, --ephemeral, --ignore-rules and structured output. Generation uses the model, account and endpoint from that host's Codex configuration, with low reasoning effort. It works for conversations from any Paseo provider; prediction itself always uses Codex. No additional application or model account is created. macOS desktop is tested; the UI uses React Native components for compact and native clients, but physical iOS and Android devices have not been tested.

Each prediction sends up to eight recent user/assistant messages, at most 6,000 characters, to the configured Codex service and consumes that account's model allowance. Tool output, reasoning and attachment file contents are excluded; private information present in message text is still included. Prompts treat the transcript as untrusted data and request no tools. Codex runs in a temporary empty directory with a read-only sandbox, without project rules, in an ephemeral session. This is not complete isolation from the user's Codex configuration or every read-only tool. Your configured provider determines data retention. The plugin does not log conversation text or model responses and never reads credentials itself.

The host creates a temporary output-schema file and removes its own temporary directory afterward. Requests time out after 30 seconds. Failures show a Retry button and do not post a fallback message. No software is downloaded or installed by the running plugin. Only an explicit send from the original composer or Send question posts to the conversation. Duplicate sends from multiple clients are rejected for the same source reply; after an uncertain send error, inspect the conversation before retrying.