# SPIRALMESH Talk — start here **Windows portable preview 0.5.0-preview.** A small local window for a finite task: Claude drafts, Codex reviews, then a fresh Claude process finishes from saved work. You choose when each stage runs. This package starts with **no tasks and no personal memories**. Installation on a second computer and an independent newcomer's live task are **not yet verified**. The demonstration is illustrative, not a receipt from a model run. ## First opening 1. Extract the **whole ZIP** to a folder you intend to keep. Do not run it inside the ZIP. Keep notes, exports and every extra file outside this folder: the launcher verifies its complete contents and refuses added, missing or changed files. 2. Double-click **SETUP.cmd**. It checks Python, its small-window support, and the existing Claude/Codex programs. It installs nothing, signs in nowhere, and runs no model. 3. If offered, create the **Open SpiralMesh Talk Preview** Desktop shortcut. Setup refuses to overwrite an existing shortcut with that name. 4. Double-click **OPEN_TALK.cmd** or that shortcut. A small control window and your browser open at **Saved tasks**. An empty list is expected. Windows may show a security warning when you open a downloaded launcher. That behavior has not been measured with this download. Do not disable Windows security controls. If it is blocked or you are unsure, stop and keep the warning text for support. You need **Windows, Python 3.11+ with Tkinter**, and a browser. Running task stages additionally requires **Claude Code CLI signed in with Claude Max** and **native Codex CLI (codex.exe) signed in with ChatGPT**. An npm-only codex.cmd does not satisfy this runner. Setup checks program availability, not authentication or remaining usage. Missing CLIs do not prevent opening the window or preparing a task. Use the providers' normal installation and sign-in instructions if prerequisites are missing. Keep any error message; do not put credentials into a task. This package includes no Python installation, subscription, or model. ## Try a small task Choose **New task**. Use this non-sensitive example: **Title:** Read-only helper checklist **Goal:** Draft a short checklist for a helper that may read public documentation but may not change files or send messages. State what a human should verify. Use only this sample; do not contact anyone or claim a system was tested. Read the preview, tick its confirmation and click **Create task**. This writes local records and an empty work library; it starts no model. On the task page, read the next participant and privacy explanation. Confirm **Resume one stage** once. This uses the assigned subscription and may read or save permitted work notes. Use **Refresh saved result** to inspect the outcome; refresh makes no second request. Review each shared reply before starting the next stage. After the draft and review, you can close and reopen Talk before starting the fresh Claude finish. That stage is asked to retrieve its saved work note; inspect the actual read evidence before claiming a successful memory resume. A question, decline, failure or unknown result means stop and inspect. [New task](NEW_TASK_START_HERE.md) · [Saved tasks and privacy](SAVED_TASKS_START_HERE.md) · [Import website work](IMPORT_SAVED_WORK_START_HERE.md) ## Reopen and stop - **Reopen:** close the browser or small window freely. Double-click OPEN_TALK.cmd or the shortcut again. It reuses its managed Talk instance if still running. - **Stop:** click **Stop Talk** in the small control window, confirm, and look for **Stopped**. A browser tab can remain visible with an old page afterward. - **Limit:** closing a tab or window is not cancellation. Stop Talk stops the UI, not an already dispatched stage. That work may continue and retain notes or results. Never erase records or retry an uncertain request to force progress. No terminal must remain open for the desktop companion. Opening, setup, creation, import and inspection make no model request. There is no automatic next-stage runner. ## Your files and privacy Data stays outside the package under: `%LOCALAPPDATA%\Article11\SpiralMeshTalk\portable-preview-v1` It holds local control records and tasks, with shared results, work-only notes, private run receipts and continuation checkpoints. The portable launcher disables direct private Ember/Lumen conversation. It does not load household libraries. You can import a record exported from the public website instead. Running a stage sends selected task context to a cloud provider. Local storage does not mean offline inference or zero retention. Privacy instructions are **not automatic redaction**. Check input and each reply yourself. [Read the exact boundaries](SAVED_TASKS_START_HERE.md#what-is-stored-and-what-is-sent). ## Update, removal and help Keep this folder in place while its shortcut points here. To update, stop Talk normally, resolve outstanding stages, and follow the next release's compatibility instructions. Extract separately; do not overwrite a running copy or put task databases in the application folder. There is no automatic updater or migration. To remove the software, stop Talk and establish that no stage remains unresolved, then remove its shortcut and package. **Keep the data folder** to retain your work. Removal does not delete provider history, exports or backups. If Talk needs attention, wait briefly and open it once more to refresh the state. This does not retry a task. If the reference persists, preserve it and ask for help. Do not kill an unknown process, erase runtime records, remove locks, or disable checks. Bundled app/ and workshop/ documents preserve their earlier component context and licenses. Some nested tools describe advanced private-Hearth connections; they are not the portable launcher's supported path. Follow these top-level guides. See [RELEASE_NOTES.md](RELEASE_NOTES.md) for limits.