PocketMind 1.18.1

PocketMind 1.18.1 turns the phone into a model backend for every other app on it, and restores the on-device NPU backend.

New since v1.18.0:

• OpenAI-compatible local server: any app or script on the device can use PocketMind as its model endpoint (/v1/chat/completions and friends, served from the active model). Loopback by default; a LAN mode can be enabled with an API key. The server runs as a foreground service, so generation keeps running with the app in the background.
• Share sheet and text selection intake: select text in any app or use the system share sheet to send it to PocketMind. Shared text lands in the composer as a draft you can edit before sending, never auto-sent.
• NPU acceleration restored: the 1.18.0 APK unintentionally lacked the QDSP6/Hexagon backend libraries (its build machine had no Hexagon SDK, so llama.rn fell back to CPU-only). This release ships the full variant set again: on Snapdragon devices with an HTP DSP, supported models run on the NPU; everyone else keeps the usual CPU/GPU paths, selected automatically.
• Build gate for the NPU backend: scripts/verify-android-payload.js now passes against the shipped APK (arm64-v8a + x86_64 rnllama libraries, HTP DSP assets verified as QDSP6 objects, backend symbols exported), so a release can no longer silently ship CPU-only.
• Internal rename completed: Java/Kotlin tree, database file names, keychain service names, and deep-link handling now use the PocketMind identifiers end-to-end. Existing data migrates in place on first launch; the legacy pocketpal:// deep-link scheme keeps working alongside pocketmind://.

No signing changes: updating from any 1.17.x/1.18.x build keeps chats, models, and settings. Obtainium updates as usual.

APK signer certificate SHA-256: b79b7e5967ed09a53a6a2912b91d9fdf358c936a7e6ccd888846d706cd963e3c

PocketMind 1.18.0

PocketMind 1.18.0 lets your Pal work with files outside its sandbox, and removes the last third-party network SDK.

New since v1.17.2:

• Device folder access (Android): grant a directory through the system picker and the file tools gain a second root. A Pal can list, read, and grep files in the granted folder (root: "device"), while the sandboxed workspace works as before. Write access is a separate opt-in and every fresh grant starts read-only. Paths resolve one segment at a time inside the grant, so nothing outside the chosen folder is reachable. Settings, Agent file access.
• No Firebase: the Firebase SDK and app check wiring are gone. The app now ships no third-party analytics or crash SDKs at all; the benchmark and feedback cloud features that pulled them in are not deployed for this distribution and fail closed.
• "Context is busy" fix: rapid consecutive sends (stop then send again, fast tool-call turns) no longer surface a native context error; the request is retried briefly and completes.
• Zip guard: archive attachments whose declared uncompressed size exceeds the extraction budget are rejected before extraction instead of being trusted.

No package id or signing changes: updating from 1.17.2 keeps your chats, models, and knowledge base. Obtainium updates this entry as usual.

APK signer certificate SHA-256: b79b7e5967ed09a53a6a2912b91d9fdf358c936a7e6ccd888846d706cd963e3c

PocketMind 1.17.2

PocketMind 1.17.2 ships the workspace UI and the non-blocking knowledge base, and completes the rebrand under the fork's own Android identity.

New since v1.17.1-fgs.1:

• Workspace screen: browse everything a Pal wrote in its sandboxed file workspace. Preview files, share them, or delete them from the drawer. The sandboxed talents (list, read, write, grep) that landed last release now have a UI.
• Non-blocking attachments: oversized documents no longer stall the first answer. Your question goes out immediately with a budget-capped head slice while the full document indexes in the background; a progress strip above the input shows extraction and indexing ("Reading report.pdf", "chunk 12/85"). The next question gets full retrieval.
• Faster indexing: the embedding context runs on the CPU's full recommended thread count instead of a fixed 4, and is pre-warmed during file extraction, hiding model-load latency.

Housekeeping: the app is now io.github.tokenbleed.pocketmind and signed with the project's own release key.

Install note for existing users: the new package id means Android treats this as a new app. It installs alongside the previous one; your chats, models, and knowledge base do not carry over automatically. Move what you need (export chats from the old app, point the models directory at the same folder or re-download), then uninstall the old one. Obtainium users: update the app entry if it pins the old package id.

APK signer certificate SHA-256: b79b7e5967ed09a53a6a2912b91d9fdf358c936a7e6ccd888846d706cd963e3c

v1.17.1-fgs.1

Background runs: every generation now runs under an Android foreground service with a progress notification, so long agent runs survive backgrounding and screen-off. The notification follows agent steps and tool calls; notification permission is asked once and denial only hides the notification.