← Plugins

Next message suggestion

Predict one editable next message on demand with Claude or Codex, following the session by default

$ paseo plugin add 52216108/next-input
Report a bug
52216108Utils39installs880e2b1fe91eSource
Next message suggestion screenshot 1

Shows one context-aware next message below a completed reply when a client opens that conversation and mounts its suggestion row. Unopened conversations do not call the prediction model. The full sentence wraps when needed. Click the paper plane to send directly, or click the text to edit in place before sending or copying it into the main input. Suggestions use small, muted text without a card border or background. When there is no useful prediction, nothing is shown.

Suggestions containing commands, paths, links, credential-related terms, or deletion, push and deployment actions must be edited before sending; their one-click send icon is hidden. This conservative rule also applies to ordinary questions containing those terms and is not a complete safety assessment.

Requires Paseo 0.13 or later and a signed-in Claude CLI or Codex CLI on the daemon host. The "Prediction tool" setting defaults to "Follow session": Codex sessions are predicted with Codex and everything else with Claude, falling back to whichever is installed. You can also pin Claude or Codex; a pinned tool that is missing means no prediction. For Claude choose Haiku (the default), Sonnet, Opus or a custom model ID; leave the Codex model empty for Codex's default; it runs with low reasoning effort. Each Codex prediction runs in a temporary CODEX_HOME that holds only a copy of your login (auth.json), the settings that decide where requests go (model_provider, [model_providers.*], chatgpt_base_url and similar), and a config turning off shell and code execution, image viewing, browser and computer use, apps and plugins, skills, hooks, memories and web search; your MCP servers, plugins and AGENTS.md are not loaded, it runs in an empty directory in a read-only sandbox, and the directory is deleted afterwards. If Codex refreshes its login tokens during the run, they are written back when your own credentials have not changed meanwhile, so you stay signed in. Only one Codex prediction runs at a time. Known limitations: the current Codex version cannot turn off spawning subagents, which share the same isolated config and have no tools either; a login kept in the system keychain (cli_auth_credentials_store = "keyring") cannot be used for predictions, so use Claude instead. Both use the existing account, endpoint and proxy and consume that account's usage. Claude is searched on PATH, then ~/.local/bin/claude, ~/.claude/local/claude, /opt/homebrew/bin/claude and /usr/local/bin/claude; Codex on PATH, then codex in ~/.local/bin, ~/.npm-global/bin, ~/.bun/bin, /opt/homebrew/bin and /usr/local/bin. If no usable executable is found, no suggestion appears in the conversation. Settings shows the status of both and the searched locations, with an option to check again.

The placeholder is appended about 3.5 seconds after a reply ends, to follow the generation-speed row when available. Each eligible turn shares one prediction request and cached result across clients, including simultaneous opens. Claude requests time out after 30 seconds and Codex requests after 45 seconds; neither retries automatically. A new turn cancels the previous prediction and invalidates its suggestion. Subagent turns and failed or cancelled turns are excluded. The client does not poll in the background. A mounted row can trigger prediction even if the app is behind another window; this is not an operating-system foreground check.

Each request reads up to eight recent user messages and assistant replies from the current conversation, capped at 12,000 characters. Tool results, reasoning and attachment files are excluded. Private text contained in visible messages is still sent to the Claude or Codex service in use. Claude predictions disable tools, MCP, skills, custom hooks and thinking and keep no session; Codex predictions turn off shell and code execution, load none of your MCP servers, plugins or hooks, and use low reasoning, but may still discover global skills and keep some utility tools (see the known limitations above). Neither performs tasks or creates Paseo subagents. The prediction instructions are in English and ask for a sentence in the user's language and voice. Suggestions are displayed as timeline components, whose storage and restoration are managed by Paseo. They are not sent to the main model as user messages until you press Send. Conversation text and raw model responses are not logged by this plugin.

Automatic prediction can be disabled in settings. Changing settings cancels pending predictions; the new settings apply after the next completed reply. A failed prediction shows an unavailable message instead of a fixed suggestion. The interface defaults to English and follows the app's Simplified Chinese locale. Suggestions scroll with the conversation. Paseo does not expose a public composer draft API, so this plugin does not fill the original input or provide Tab-to-accept completion.