Arbor 0.22.4

Arbor 0.22.4

• Replace the raw Google Drive UNREGISTERED_ON_API_CONSOLE message with an actionable OAuth-registration diagnostic.
• Show and copy the installed package name, signing SHA-1, signing SHA-256, required drive.appdata scope, and setup guide.
• Document the exact public GitHub release package/signing identity while keeping client secrets out of the app and repository.
• Embed the GitHub source repository and source commit into builds.
• Check that embedded repository's latest GitHub Release automatically once per day and manually from About Arbor.
• Make fork releases follow their own fork while rehosted APKs continue trusting the repository they were built from.
• Publish a signed release manifest containing package, version code, APK checksum, signing-certificate digest, and source commit.
• Offer direct APK downloads only when package and signing certificate match the installed build; otherwise open the release page safely.

Arbor 0.22.3

Arbor 0.22.3

• Correct app-bar scroll physics so content tracks the finger one-to-one while the title collapses or expands.
• Restore Material nested-scroll consumption for chat and Settings.
• Remove continuous list-offset-to-app-bar synchronization that added app-bar inset movement on top of the list's own scroll.
• Restore saved title collapse once when reopening a chat or Settings page, then leave live movement to a single scroll owner.

Arbor 0.22.2

Arbor 0.22.2

• Give the Android IME first ownership of predictive Back while the keyboard is visible.
• Prevent drawer and page navigation from consuming the keyboard's Back gesture.
• Replace Search's oversized collapsing header with a compact, pinned, IME-aware search field and clear empty states.
• Remove competing nested-scroll ownership that made fully collapsed chat and Settings titles snap back to expanded.
• Keep a collapsed Settings title stable through transient pre-measure and IME-resize states.
• Parse Google authorization result data before classifying the account picker as canceled.

Arbor 0.22.1

Arbor 0.22.1

Arbor 0.22.1 improves memory management and repairs navigation state on phones.

Memory management

• Gives memory text the full card width and moves controls into a separate metadata row.
• Adds enabled/disabled filtering, category filtering, sorting, selection, and bulk enable/disable/delete.
• Adds confirmation before destructive single-item, selected-item, or disabled-item cleanup.
• Shows whether a memory was saved manually or from a chat, plus its update time.
• Prevents a small memory library from being injected wholesale into unrelated chats.
• Keeps a small set of profile/preference memories available as baseline context and prioritizes relevant or same-chat memories.
• Merges exact duplicates even when their categories differ, while fuzzy duplicate matching remains category-scoped.

Navigation and scrolling

• New Settings destinations start at the top; Back restores the prior page position.
• Settings and chat titles are derived from the actual restored scroll position, preventing expanded titles over scrolled content.
• Opening Settings from the drawer starts at the Settings root instead of reviving a stale nested page.
• Chat switching restores each chat's own saved position without inheriting another chat's app-bar state.

Predictive Back

• An open drawer now owns the complete predictive-back gesture.
• Back progress directly closes the drawer, cancellation restores it, and page navigation cannot steal the gesture halfway through.

Arbor 0.22.0

Arbor 0.22.0

Version code: 164

This release ships the adaptive Home-widget renderer and substantially upgrades Arbor's cross-chat memory management.

Memory management

• Replaces the flat newest-first memory injection with relevance-aware selection based on the active conversation.
• Enforces strict memory item and character budgets so a large memory library cannot crowd chat history out of the model context.
• Preserves a small recent-memory fallback when no strong topical match exists.
• Adds punctuation-, case-, and whitespace-insensitive duplicate detection while avoiding unsafe merges of negated preferences.
• Adds native memory search and edit tools, and expands memory listing with query, disabled-item, and result-limit controls.
• Makes memory saves report whether an item was created, updated, or merged.
• Adds search, inline editing, enable-all, disable-all, and delete-disabled controls to Memory settings.
• Keeps disabled memories out of ordinary model context and search results unless explicitly requested.

Home widgets

• Replaces equal-share row and column sizing with intrinsic, content-aware allocation.
• Allows readable adaptive fitting down to an 11sp compact-widget floor instead of rejecting every layout below 13sp.
• Makes compact list and metric nodes use their available space efficiently.
• Tests compact, standard, and expanded launcher canvases instead of the artificial 84dp compiler canvas.
• Reduces launcher chrome so generated content receives more usable height.
• Accepts common generated style aliases such as fontWeight and size, then stores canonical Arbor schema.
• Adds regression coverage for six-row prayer lists and genuine three-row prayer layouts.

Verification

The release workflow runs third-party license verification, offline license generation, release unit tests, release lint, optimized APK and AAB assembly, instrumentation APK assembly, and APK signature verification before publishing the GitHub release.

Arbor 0.21.0

Arbor 0.21.0

This release repairs long-running agent work and adds first-class local memory while relaxing arbitrary generated-widget UI limits.

Long-running work and Continue

• Streaming model requests no longer fail merely because a reasoning model emits no SSE bytes for two or three minutes.
• Manual Continue waits for the previous WorkManager instance to finish cancelling before marking the message as streaming, eliminating the cancellation race which made Continue appear to do nothing.
• Worker replacement cancellation no longer overwrites a newly resumed message with INTERRUPTED.
• Automatic output continuation rises from three to twelve segments.
• Normal tool workflows can use up to 64 tool rounds and Deep Research up to 128, replacing the previous 8/24-round ceiling which prematurely finalized widget-building sessions.

Memory

• Adds encrypted cross-chat memories stored in the SQLCipher Room database.
• Adds native memory_save, memory_list, and memory_forget tools with visible Working events.
• Injects enabled memories as bounded user-owned reference data, never as instructions.
• Adds Memory settings with global enable, automatic-memory policy, manual add, per-item enable/disable, and delete controls.
• Includes memories and memory policy in app-settings backups.
• Automatic memory excludes transient details and requires explicit consent for sensitive information.

Generated widgets

• Raises arbitrary schema ceilings while retaining bounded resource and security limits.
• Accepts input-style nodes on Home widgets as state readouts/actions instead of rejecting the entire program.
• Removes the six-row renderer truncation.
• Adds bounded multiline text wrapping and ellipsis instead of flattening every line into one clipped row.
• Exposes input actions in the Home-widget action strip.

Verification

Third-party license verification, offline license generation, release unit tests, release lint, and the optimized release APK build passed before publication.

Arbor 0.20.36

Arbor 0.20.36

This release repairs the widget compiler network layer falsely reporting EOFException for every normal HTTPS JSON response, including Open-Meteo.

Fixed

• Replaces the exact-length readUtf8(1 MB + 1) call which required every response to contain more than 1 MB and therefore threw EOFException for ordinary short API bodies.
• Reads response bodies incrementally until real EOF while retaining the strict 1 MB safety ceiling.
• Probes one byte beyond the ceiling so chunked and unknown-length oversized responses are still rejected without unbounded buffering.
• Treats genuine I/O failures, including premature EOF and TLS truncation, as transient during compiler preflight when complete offline fallbacks exist.
• Adds regression tests for short JSON, exact-limit, and oversized response bodies.

Verification

Release unit tests, release lint, the optimized release APK build, and the widget HTTP body regression suite passed before publication.

Arbor 0.20.35

Arbor 0.20.35

This release repairs repeated Response paused loops after a provider returns an output-length finish reason.

Fixed

• Arbor now automatically requests up to three continuation segments when each segment makes progress.
• A zero-progress length loop is stopped and reported as a provider error instead of offering an endless Continue cycle.
• Manual Resume uses the limits currently shown in Chat configuration, so raising the output ceiling now affects an already-paused response.
• DeepSeek prefix completion receives the exact saved assistant prefix. Internal reasoning/tool metadata is no longer appended after the visible answer and mistaken for text to continue.

The provider/model and prompt identity remain pinned to the original request; only user-adjustable runtime limits are refreshed on Resume.

Arbor 0.20.34

Arbor 0.20.34

• Fixes the remaining DeepSeek DSML leak after Arbor exhausted the tool-execution budget and entered its tools-disabled final synthesis turn.
• Keeps a separate protocol-firewall allowlist after callable tools are removed, so stale compile_widget, web_search, and web_fetch DSML can still be recognized without making those tools callable again.
• Quarantines the complete finalization response until it is classified; protocol text is rolled back before it can reach the message or title.
• Retries once with an explicit no-tools finalization instruction. Repeated protocol output is rejected cleanly and is never shown or executed.
• Adds an end-to-end transport regression test reproducing the prayer-time web_fetch DSML sequence from the device screenshot.

Arbor 0.20.33

Arbor 0.20.33

• Quarantines DeepSeek tool-turn content and reasoning until Arbor can classify the completed response, preventing raw DSML from ever being committed to chat.
• Parses DSML emitted through either content or reasoning_content, including HTML-escaped brackets, Unicode pipe glyphs, zero-width format characters, and whitespace-separated fences.
• Detects unparseable DSML-shaped output and retries with the strict structured-tool correction prompt; after the retry, unsafe protocol text is rejected rather than displayed.
• Expands DeepSeek detection across provider id, provider name, base URL, model id, and model display name.
• Adds regression tests reproducing the on-device prayer-time web_fetch failure in the reasoning channel and escaped Unicode fence variants.