feat(cot): DeepSeek-R1 CoT accordion, multimodal EXIF downsampling, TokenStorage hardening & test suite by PrinceBad · Pull Request #3 · PrinceBad/Local-LLM-AI

PrinceBad · 2026-09-17T04:58:18Z

Summary

This PR hardens the offline Android client (com.example.auralocalai) across five key audit areas:

DeepSeek-R1 Chain-of-Thought (CoT) Accordion: Collapsible reasoning UI with cancellation lifecycle handling, multi-block parsing, and 3-tier model scoping.
Multimodal Memory & EXIF Matrix: Coarse-to-fine bitmap downsampling with 3-stream EXIF orientation/mirroring normalization and native bitmap recycling.
TokenStorage Hardening: Unified secure Keystore token persistence, eliminating plaintext SharedPreferences write-path bypasses and migration loops.
Startup ANR & Passive Download Policy: Backgrounded startup file migrations and prevented mid-chat model eviction upon download completion.
Full Unit Test & Build Verification: Added pure Kotlin JUnit test suite ParseThinkBlocksTest; verified 42/42 unit tests passing and clean debug APK packaging.

Key Changes

1. Reasoning CoT Accordion & Cancellation Lifecycle (`ChatScreen.kt`)

Mid-Think Cancellation: Decoupled <think> tag presence from the generation lifecycle. When a user cancels mid-thought before </think> arrives, the header immediately transitions from "Thinking in progress…" to "Thinking (stopped)" in muted slate, collapsing by default rather than remaining frozen in live streaming mode.
Multi-Block Parsing: Replaced single-block string scans with an iterative while-loop supporting arbitrary N thought blocks joined with \n\n---\n\n dividers, aggregating interleaved responses cleanly into mainContent.
3-Tier Model Scoping:
- Catalog Non-Reasoning Models (qwen2.5-coder-3b, qwen-1.5b, gemma4-e2b): Strictly disabled from accordion parsing to prevent false-positive collapses when outputting XML/HTML tags or discussing prompt engineering.
- Catalog Reasoning Models (deepseek-1.5b, qwq): Enabled by default.
- Custom / Unknown Imported Models: Gated strictly on confirmed prefix: message.content.trimStart().startsWith("<think>"). Ambiguous prefixes (e.g. < or <th) are treated as provisional plain text until confirmed, avoiding false-positive locks on <html> or code blocks.
Stateless Scroll Recycling: Avoided ephemeral remember(message.id) latches in ChatBubble. Reasoning classification evaluates against the immutable message.content, surviving off-screen item disposal and recomposition when users scroll during long generations.
Legacy History Decoupling: Saved messages lacking modelId evaluate content prefix independently of uiState.activeModelId, ensuring switching models does not alter past DeepSeek chat presentation.

2. Multimodal Image Downsampling & EXIF Correction (`LlmViewModel.kt`, `LlmInferenceEngine.kt`)

Coarse-to-Fine Downsample: Added decodeSampledBitmap(context, uri, maxDimension = 1024) using 3 independent stream passes:
1. Dimension probing (inJustDecodeBounds = true).
2. EXIF orientation detection via ExifInterface(stream).
3. Power-of-2 inSampleSize decode.
Unified Matrix Transform: Combines EXIF orientation (90°/180°/270°/flip/transpose) and fine scale resize in a single Bitmap.createBitmap pass, recycling intermediate allocations.
Native Memory Cleanup: Local vision bitmap is recycled in sendMessage's finally block. Decoupled from Coil UI image loading.

3. Keystore TokenStorage Hardening & Startup ANRs (`LlmViewModel.kt`)

Direct Keystore Persistence: Routed saveHfToken() and clearHfToken() asynchronously through tokenStorage.saveToken() and tokenStorage.clearToken() on Dispatchers.IO. Dropped raw app_settings plaintext writes.
Single Authoritative Reader: Removed plaintext read in init block 2, eliminating startup race conditions and retiring the accidental migrateTokenIfPresent() loop.
Startup ANR Prevention: Backgrounded migrateExistingModels() and refreshDownloadedModels() off the main thread into viewModelScope.launch(Dispatchers.IO).

4. Passive Download Completion Policy (`LlmViewModel.kt`)

Mid-chat download completions no longer forcibly evict the active loaded model. New models are auto-loaded only if modelState is ModelState.Unloaded || modelState is ModelState.Error; otherwise, a non-intrusive Toast is displayed ("$displayName downloaded. Ready in Model Settings.").

Test & Build Verification

1. Unit Tests (`./gradlew.bat testDebugUnitTest`)

Result: BUILD SUCCESSFUL in 30s
Total Tests: 42/42 passed (100% success rate)
- ParseThinkBlocksTest (9 tests):
  - testSingleClosedThinkBlock: PASSED
  - testUnclosedThinkBlock_MidGenerationOrCancelled: PASSED
  - testMultipleThinkBlocks_Interleaved: PASSED
  - testMultipleThinkBlocks_TrailingUnclosed: PASSED
  - testNoThinkBlock_StandardOrCodingOutput: PASSED
  - testThinkBlock_EmbeddedInCode: PASSED
  - testEmptyAndWhitespaceThinkBlocks: PASSED
  - testLeadingWhitespaceBeforeThink: PASSED
  - testModelPresetReasoningResolution: PASSED
- TokenStorageTest (5 tests): PASSED
- ModelSafetyValidatorTest (10 tests): PASSED
- ModelDownloaderTest (6 tests): PASSED
- ChatRepositoryTest (6 tests): PASSED
- TelemetryBenchmarkTest (6 tests): PASSED

2. Full Debug APK Build (`./gradlew.bat assembleDebug`)

Result: BUILD SUCCESSFUL in 19s (38 actionable tasks, 0 compiler/DEX errors).

Manual Verification Protocol

Test Scenario	Action	Expected Outcome
7A (Rear Camera)	Attach portrait rear photo to Gemma 4 chat	Image decodes upright with correct EXIF orientation; inference succeeds without OOM.
7B (Front Camera)	Attach selfie front camera photo	EXIF mirror/rotation is normalized; no upside-down inversion.
7C (Pre-rotated)	Attach 90°/270° pre-rotated gallery photo	Single-pass Matrix handles rotation + downsampling cleanly.
8A (Mid-Think Stop)	Send prompt to DeepSeek-R1; tap Stop mid-thought	Header switches to `"Thinking (stopped)"` in muted gray; animation stops cleanly.
8B (Coding Model)	Ask Qwen 2.5 Coder to write XML containing `<think>`	Output rendered as literal text/code; no accordion is created.
8C (Scroll Lifecycle)	Scroll up and down in `LazyColumn` during generation	Reasoning accordion retains folded state upon item disposal and recomposition.
8D (Legacy History)	Switch active model to Qwen; view past DeepSeek chat	Historical reasoning blocks remain formatted as `"Thinking Process"`.

…tool, and reconcile 12GB RAM tier for Gemma 4 E4B

…, 3-tier scoping, and test suite - Decouple unclosed tag parsing from generation lifecycle to show 'Thinking (stopped)' on cancellation - Add 3-tier reasoning model scoping (catalog non-reasoning, catalog reasoning, and custom imported prefix check) - Add live-streaming token prefix anticipation and remember(message.id) state latching - Support arbitrary multi-block think sequences joined with dividers - Add sampled bitmap decode with 3-stream EXIF matrix transform and vision memory recycling - Route Hugging Face tokens strictly via TokenStorage on Dispatchers.IO and prevent passive session eviction - Add ParseThinkBlocksTest JUnit test suite with 42/42 unit tests passing and verify assembleDebug

…ning detection to confirmed prefix - Replace ambiguous partial match latch with confirmed tag check (trimmed.startsWith('<think>')) - Prevent custom models outputting <html> or XML tags from locking permanently into reasoning mode - Remove ephemeral remember(message.id) latch from ChatBubble, avoiding state resets on LazyColumn item disposal - Retain confirmed startsWith('<think>') prefix gating on message.content across live streaming and scroll recycling

PrinceBad added 3 commits September 16, 2026 18:59

ci(actions): add GitHub Actions workflow, physical device diagnostic …

b924a9b

…tool, and reconcile 12GB RAM tier for Gemma 4 E4B

PrinceBad merged commit 5386614 into main Sep 17, 2026
1 check passed

Navigation Menu

feat(cot): DeepSeek-R1 CoT accordion, multimodal EXIF downsampling, TokenStorage hardening & test suite - #3

feat(cot): DeepSeek-R1 CoT accordion, multimodal EXIF downsampling, TokenStorage hardening & test suite#3
PrinceBad merged 3 commits into
mainfrom
feature/phase-5-ci-validation

PrinceBad commented Sep 17, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

PrinceBad commented Sep 17, 2026

Summary

Key Changes

1. Reasoning CoT Accordion & Cancellation Lifecycle (ChatScreen.kt)

2. Multimodal Image Downsampling & EXIF Correction (LlmViewModel.kt, LlmInferenceEngine.kt)

3. Keystore TokenStorage Hardening & Startup ANRs (LlmViewModel.kt)

4. Passive Download Completion Policy (LlmViewModel.kt)

Test & Build Verification

1. Unit Tests (./gradlew.bat testDebugUnitTest)

2. Full Debug APK Build (./gradlew.bat assembleDebug)

Manual Verification Protocol

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

1. Reasoning CoT Accordion & Cancellation Lifecycle (`ChatScreen.kt`)

2. Multimodal Image Downsampling & EXIF Correction (`LlmViewModel.kt`, `LlmInferenceEngine.kt`)

3. Keystore TokenStorage Hardening & Startup ANRs (`LlmViewModel.kt`)

4. Passive Download Completion Policy (`LlmViewModel.kt`)

1. Unit Tests (`./gradlew.bat testDebugUnitTest`)

2. Full Debug APK Build (`./gradlew.bat assembleDebug`)