feat(core): summarizer streaming/dynamic-context + whisper model fallback + pro bump [WIP/divergent]

[dishit-wednesday]

Stacked onto feat/locket-pro (PR #433). This session's core-side work:

• transcriptSummarizer: dynamic context sizing (scales to the active backend), token streaming (onToken across local/LiteRT/remote), abort()
• litert: expose getContextTokens for backend-aware chunk sizing
• whisperStore: fall back to another on-disk model when the active one is deleted (+ tests)
• bump pro submodule to the insights-redesign / continuous-smart tip

NOTE (WIP): branch cut Jul 9, behind feat/locket-pro by ~455 commits (it merged dev). GitHub's PR diff (merge-base) shows only these changes; needs rebase/cherry-pick onto the current tip before merge. Draft for review of the intended changes.

---

Summary by Gitar

• Summarizer Upgrades:
  • Refactored TranscriptSummarizerService to use dynamic, backend-aware context sizing based on INPUT_CONTEXT_FRACTION.
  • Added abort() support to TranscriptSummarizerService to interrupt active local, LiteRT, or remote generations.
  • Added grammar and repeatPenalty parameters to generateSummaryText and summarizeOne for improved output control.
• Whisper & CoreML:
  • Implemented automatic iOS CoreML encoder backfilling and integration to accelerate Whisper inference on the Neural Engine.
  • Updated whisperStore to automatically fall back to alternative models if the active one is deleted.
  • Restricted WhisperService to single-instance file transcription to prevent native context race conditions.
• Dev & Tools:
  • Added DevGrammarModal and DevGrammarStore to test GBNF/LLGuidance constraints behind __DEV__.
  • Implemented a global Toast component for imperative, non-blocking UI notifications.
  • Added System.gc() and finalization drains in LiteRT to mitigate potential native SIGSEGV memory crashes.

_{This will update automatically on new commits.}

Summary by CodeRabbit

• New Features
  • Added iOS CoreML acceleration support for compatible Whisper models, with automatic encoder downloads.
  • Added a developer-only grammar testing interface for constrained chat responses.
  • Added safer constrained-generation fallback so chat can continue normally if a custom constraint fails.
• Improvements
  • Improved Whisper model management by automatically selecting another available model when the active model is deleted or missing.
  • Improved generation controls and transcript summarization reliability.
• Documentation
  • Added planning and handoff documentation for recording, transcription, and insights features.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 4cc0020b-dd96-4afd-b0f9-3009c233a4a5

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

The change adds development-only grammar and constrained-decoding controls, iOS Whisper CoreML encoder support, recording insight tests, model fallback reconciliation, platform-specific GPU behavior, and extensive recorder, intelligence, and transcript-quality planning documents.

Changes

On-device generation and transcription

Layer / File(s)	Summary
Grammar harness and constrained decoding src/components/DevGrammarModal.tsx, src/stores/devInferenceStore.ts, src/services/devInference.ts, src/services/llm.ts, src/services/llmToolGeneration.ts, src/services/litert.ts, android/app/src/main/java/.../LiteRTModule.kt	Adds a persisted DEV grammar modal, llama grammar overrides with fallback retries, and Android LiteRT constrained-decoding wiring.
Context-aware summarization src/services/transcriptSummarizer.ts	Uses backend context sizes, forwards grammar and repetition settings, and adds abort handling.
Whisper CoreML lifecycle src/services/whisperModels.ts, src/services/whisperService.ts, __tests__/unit/services/whisperService.test.ts	Adds typed CoreML encoder metadata, iOS encoder installation, asset-gated loading, backfill behavior, and diagnostics.
Model and platform state src/stores/whisperStore.ts, src/services/llmHelpers.ts, __tests__/unit/stores/whisperStore.test.ts, __tests__/unit/services/llmHelpers.test.ts	Reconciles active Whisper models from disk and limits GPU-layer caps to Android.
Recording insight validation __tests__/unit/locket/*	Adds coverage for scoring, extractive parsing, action-item refinement, due dates, backend readiness, merging, and error states.

Estimated code review effort: 4 (Complex) | ~60 minutes

🚥 Pre-merge checks | ✅ 3 | ❌ 2

❌ Failed checks (2 warnings)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 66.67% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.
Description check	⚠️ Warning	The description gives a summary but omits the required template sections for type, screenshots, checklist, related issues, and additional notes.	Rewrite the PR description using the repository template and fill in the missing sections, especially Type of Change, Checklist, Related Issues, and screenshots if applicable.

✅ Passed checks (3 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly summarizes the main changes: summarizer upgrades, Whisper fallback, and a pro submodule bump.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/vad-gated-recorder

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[dishit-wednesday]

/gemini review

[dishit-wednesday]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[coderabbitai]

Actionable comments posted: 9

Note

Due to the large number of review comments, Critical, Major severity comments were prioritized as inline comments.

🟡 Minor comments (24)

src/screens/ChatScreen/ChatScreenComponents.tsx-89-90 (1)

89-90: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

devActive doesn't account for LiteRT constraints.

The selector only checks s.grammar.trim().length > 0 (GBNF), but the store also supports litertConstraintString for LiteRT. If a dev enables the override with only a LiteRT constraint set, the header icon won't light up, making it unclear the override is active.

🛡️ Proposed fix

-  const devActive = useDevInferenceStore((s) => s.enabled && s.grammar.trim().length > 0);
+  const devActive = useDevInferenceStore(
+    (s) => s.enabled && (s.grammar.trim().length > 0 || s.litertConstraintString.trim().length > 0),
+  );

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/screens/ChatScreen/ChatScreenComponents.tsx` around lines 89 - 90, Update
the devActive selector in ChatScreenComponents to consider both configured
constraint sources: keep requiring s.enabled, but treat a non-empty trimmed
grammar or a non-empty trimmed litertConstraintString as active. Preserve the
inactive state when neither constraint is configured.

docs/plans/locket-continuous-smart-recorder-handoff.md-168-168 (1)

168-168: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Use approved brand vocabulary.

“Robustness” is listed as a forbidden term in the repository’s documentation style. Replace it with “whole-day reliability” or another concrete description.

As per coding guidelines, documentation must follow the brand standard and avoid forbidden words.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/locket-continuous-smart-recorder-handoff.md` at line 168, Update
the “Whole-day robustness” heading in the documented plan to use approved brand
vocabulary, replacing “robustness” with “whole-day reliability” or another
concrete description while preserving the heading’s meaning.

Source: Coding guidelines

docs/plans/locket-continuous-smart-recorder-handoff.md-136-137 (1)

136-137: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add blank lines around the constants table.

The table immediately follows the heading, triggering Markdownlint MD058.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/locket-continuous-smart-recorder-handoff.md` around lines 136 -
137, Add a blank line between the “Decided constants” heading and its table, and
ensure the table has the required surrounding blank lines to satisfy
Markdownlint MD058.

Source: Linters/SAST tools

docs/plans/locket-insights-handoff.md-29-37 (1)

29-37: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add languages to all fenced code blocks.

Use text for the pipeline/configuration blocks so Markdownlint MD040 passes and rendering remains explicit.

Also applies to: 176-180

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/locket-insights-handoff.md` around lines 29 - 37, Update the
fenced code blocks in the handoff document, including the blocks around the
pipeline and the referenced later section, to specify text as the language
identifier. Preserve each block’s contents and formatting while ensuring every
fence has an explicit language.

Source: Linters/SAST tools

docs/plans/locket-insights-screen-plan.md-47-53 (1)

47-53: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add languages to the fenced documentation blocks.

Use text for the loose output format and router pseudocode so Markdownlint MD040 passes.

Also applies to: 121-129

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/locket-insights-screen-plan.md` around lines 47 - 53, Add the text
language identifier to the fenced documentation blocks showing the loose output
format and router pseudocode, including the corresponding block at the other
referenced section, so every fence has an explicit language for Markdownlint
MD040.

Source: Linters/SAST tools

docs/plans/locket-insights-handoff.md-1-1 (1)

1-1: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Remove em dashes from the documentation copy.

Examples include the title, “TL;DR — current state,” “pure heuristics — no model,” and “— llama.rn only.” Replace them with colons, semicolons, or standard hyphens.

As per coding guidelines, documentation copy must not use em dashes.

Also applies to: 13-18, 66-75, 115-123

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/locket-insights-handoff.md` at line 1, Remove every em dash from
the documentation copy in the handoff document, including the title and the
referenced “TL;DR,” “pure heuristics,” and “llama.rn only” text. Replace each
with an appropriate colon, semicolon, or standard hyphen while preserving the
original meaning.

Source: Coding guidelines

src/services/litert.ts-142-155 (1)

142-155: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add tests for the new LiteRT paths src/services/litert.ts:142-155,507-511

• getContextTokens() has no direct coverage.
• The __DEV__-guarded setConstrainedDecoding branch in resetConversation() also needs a dev-mode test.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/litert.ts` around lines 142 - 155, The new LiteRT paths lack
test coverage. Add tests for getContextTokens() and for resetConversation() with
__DEV__ enabled, covering the setConstrainedDecoding branch and its expected
invocation when the dev constraint is armed, while preserving existing behavior
for disabled or empty constraints.

Source: Coding guidelines

docs/plans/continuous-smart-recorder-plan.md-27-29 (1)

27-29: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add a language to the fenced block.

Use ```ts for the TypeScript snippet so the document passes MD040.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/continuous-smart-recorder-plan.md` around lines 27 - 29, Update
the fenced code block containing the RecorderMode type declaration to specify
the TypeScript language with a ts fence, ensuring the documentation satisfies
MD040.

Source: Linters/SAST tools

docs/plans/locket-commit-plan.md-30-36 (1)

30-36: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add blank lines around the table.

Separate the preceding text from the table and the following text to satisfy MD058.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/locket-commit-plan.md` around lines 30 - 36, Add blank lines
immediately before and after the “Do NOT commit” table in the plan document,
keeping the table contents unchanged so it satisfies the Markdown table-spacing
rule MD058.

Source: Linters/SAST tools

docs/plans/insights-page-design-brief.md-9-153 (1)

9-153: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add a language to the embedded prompt fence.

Use ```text so the document passes MD040.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/insights-page-design-brief.md` around lines 9 - 153, Update the
embedded prompt fence in the design brief to specify the text language by
changing the opening fence to ```text, while leaving the prompt content and
closing fence unchanged.

Source: Linters/SAST tools

docs/plans/audio-recorder-notes.md-1-1 (1)

1-1: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Replace the em dash in the title.

The documentation copy standard prohibits em dashes; use a colon or hyphen instead.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/audio-recorder-notes.md` at line 1, Replace the em dash in the
“Audio / recorder work — current state” title with a permitted colon or hyphen,
preserving the rest of the title unchanged.

Source: Coding guidelines

docs/plans/auto-detect-recording-plan.md-71-71 (1)

71-71: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add a language to the pseudocode fence.

Use ```text so the document passes MD040.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/auto-detect-recording-plan.md` at line 71, Update the pseudocode
code fence in auto-detect-recording-plan.md to specify the text language as
```text, preserving the enclosed content unchanged.

Source: Linters/SAST tools

docs/plans/continuous-smart-recorder-plan.md-1-1 (1)

1-1: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Replace the em dash in the title.

The documentation copy standard prohibits em dashes.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/continuous-smart-recorder-plan.md` at line 1, Replace the em dash
in the “Continuous-Smart Recorder Mode — Implementation Plan” title with an
allowed non-em-dash separator, preserving the title wording.

Source: Coding guidelines

docs/plans/audio-recorder-notes.md-24-32 (1)

24-32: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add a language to the fenced block.

Use ```text or another appropriate language so the document passes MD040.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/audio-recorder-notes.md` around lines 24 - 32, Update the fenced
code block containing the commit list in the audio recorder notes to include an
explicit language identifier, preferably text, immediately after the opening
fence; leave the listed commits unchanged.

Source: Linters/SAST tools

docs/plans/locket-commit-plan.md-21-27 (1)

21-27: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Add a language to the fenced block.

Use ```gitignore so the document passes MD040.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/locket-commit-plan.md` around lines 21 - 27, Update the fenced
code block in the locket commit plan to specify the gitignore language, changing
the opening fence to gitignore while preserving its existing ignore patterns.

Source: Linters/SAST tools

docs/plans/auto-detect-recording-plan.md-1-1 (1)

1-1: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Replace the em dash in the title.

The documentation copy standard prohibits em dashes.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/auto-detect-recording-plan.md` at line 1, Replace the em dash in
the “Auto-detect (VAD-gated) recording” title with an approved non-em-dash
separator, preserving the title wording and meaning.

Source: Coding guidelines

docs/plans/insights-page-design-brief.md-1-1 (1)

1-1: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Replace the em dash in the title.

The documentation copy standard prohibits em dashes.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/insights-page-design-brief.md` at line 1, Update the “Insights
Page — Design Brief” heading to replace the em dash with an approved non-em-dash
separator, preserving the title wording.

Source: Coding guidelines

docs/plans/locket-commit-plan.md-10-10 (1)

10-10: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

List the required commit-time checks explicitly.

“As per coding guidelines, staged Kotlin/Gradle changes must run compileDebugKotlin, lintDebug, and npm run test:android, while staged JavaScript/TypeScript changes must run staged-file eslint, tsc --noEmit, and npm test.”

The current shorthand is too vague to enforce those gates reliably.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/locket-commit-plan.md` at line 10, Update the commit-time
quality-gates statement in the plan to explicitly require compileDebugKotlin,
lintDebug, and npm run test:android for staged Kotlin/Gradle changes, and
staged-file eslint, tsc --noEmit, and npm test for staged JavaScript/TypeScript
changes; retain the instruction to pause and report native hook failures instead
of using --no-verify without approval.

Source: Coding guidelines

docs/plans/combined-intelligence-layer-plan.md-116-127 (1)

116-127: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Em dashes used in the "Important flags" section.

Several bullets use em dashes (e.g., "biggest cost saver;" / "turn it on for" / "extra work if iOS matters" / "must add device" / "ingest into meetings"), which the docs brand-tone/copy standard disallows. As per coding guidelines, docs/text changes "must follow docs/brand_tone_voice.md and its copy standards (proof-first, no exclamation marks, no em dashes, no curly quotes, etc.)."

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/combined-intelligence-layer-plan.md` around lines 116 - 127,
Update the “Important flags” bullets in the combined intelligence layer plan to
remove all em dashes and replace them with approved punctuation or sentence
structure, while preserving each bullet’s meaning and existing proof-first tone.
Ensure the edited text follows the copy standards in docs/brand_tone_voice.md,
including no em dashes.

Source: Coding guidelines

docs/plans/transcript-quality-experiment.md-82-82 (1)

82-82: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Em dash in decode-params bullet.

"--no-context) — the loop killer, test first" uses an em dash, against the docs brand-tone standard of no em dashes. As per coding guidelines.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/transcript-quality-experiment.md` at line 82, Update the
condition_on_previous_text bullet in transcript-quality-experiment.md to replace
the em dash between “--no-context” and “the loop killer” with approved
punctuation, preserving the existing meaning and wording.

Source: Coding guidelines

docs/plans/transcript-quality-findings.md-260-262 (1)

260-262: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Stale diarization conclusion in the DELIVERABLES summary contradicts the doc's own later findings.

Item 5 says diarization needs a FIXED cluster count (auto-count fails 23 vs 2, cost ~10× transcription). But the body's later "Diarization status" section (Lines 149-177) reports auto-count SOLVED via NME-SC, and the "Cost correction" (Lines 186-190) revises the cost to ~⅓ of transcription, not 10×. This is the actionable summary readers will follow — leaving it stale risks shipping the older FastClustering/fixed-count approach.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/transcript-quality-findings.md` around lines 260 - 262, The
diarization summary in item 5 is stale and contradicts the later “Diarization
status” and “Cost correction” findings. Update item 5 to describe NME-SC
auto-count as solved and reflect the revised cost of approximately one-third of
transcription, removing the fixed-cluster-count and outdated 10× cost claims.

docs/plans/transcript-quality-findings.md-11-11 (1)

11-11: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Em dashes throughout violate the brand tone guideline.

This doc (and its sibling handoff doc) use em dashes ("—") extensively, e.g. Lines 11, 26, 93, 106. As per coding guidelines, "Any change to website copy, essays, docs text, UI strings, or marketing content must follow docs/brand_tone_voice.md and its copy standards (proof-first, no exclamation marks, no em dashes, no forbidden words, no curly quotes, etc.)."

Also applies to: 26-26, 93-93, 106-106

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/transcript-quality-findings.md` at line 11, Replace all em dashes
in docs/plans/transcript-quality-findings.md and its sibling handoff document
with brand-compliant punctuation or wording, including the occurrences near the
referenced lines. Preserve the original meaning and ensure no em dashes remain
in either document.

Source: Coding guidelines

docs/plans/session-handoff-transcription-intelligence.md-26-27 (1)

26-27: 📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Em dashes throughout violate the brand tone guideline.

Same pattern as the sibling findings doc — e.g. Lines 26, 27, 65. As per coding guidelines, docs text "must follow docs/brand_tone_voice.md and its copy standards (proof-first, no exclamation marks, no em dashes, no forbidden words, no curly quotes, etc.)."

Also applies to: 65-65

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/session-handoff-transcription-intelligence.md` around lines 26 -
27, Update the prose in the session handoff transcription intelligence plan,
especially the bullets “On-device / privacy by architecture, not by promise” and
“Honest limitations over hype,” to remove em dashes and replace them with
brand-compliant punctuation or sentence structure. Apply the same change to the
text around the additionally referenced line, preserving the meaning and all
other brand-tone requirements.

Source: Coding guidelines

docs/plans/session-handoff-transcription-intelligence.md-207-208 (1)

207-208: 🔒 Security & Privacy | 🟡 Minor | ⚡ Quick win

Redact the personal email and device UDID from this planning doc. They are unnecessary in tracked repo text and expose personal identifiers.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/session-handoff-transcription-intelligence.md` around lines 207 -
208, Update the planning document’s “Commit co-author” and iOS device build
instructions to remove the personal email address and device UDID while
preserving the surrounding attribution and build guidance in anonymized form.

🧹 Nitpick comments (9)

src/components/DevGrammarModal.tsx (1)

176-186: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Extract the nested ternary placeholder for readability.

The LiteRT constraint placeholder uses a nested ternary that's hard to scan at a glance. Extracting it to a variable or helper would improve clarity.

♻️ Optional refactor

-              placeholder={litertType === 'json_schema' ? '{"type":"object","properties":{...}}' : litertType === 'regex' ? '(TITLE: .+\\n)+' : 'start: ...'}
+              placeholder={
+                litertType === 'json_schema'
+                  ? '{"type":"object","properties":{...}}'
+                  : litertType === 'regex'
+                    ? '(TITLE: .+\\n)+'
+                    : 'start: ...'
+              }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/components/DevGrammarModal.tsx` around lines 176 - 186, In
DevGrammarModal, extract the nested ternary used by the TextInput placeholder
into a clearly named variable or helper based on litertType, then pass that
value to placeholder while preserving the existing json_schema, regex, and
default placeholder strings.

docs/plans/locket-insights-handoff.md (1)

106-106: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Document the UI icon instead of an emoji.

The ⚡ notation conflicts with the repository’s no-emoji UI rule. Specify the intended vector-icon name so the implementation does not copy the glyph.

As per coding guidelines, UI icons must use the shared icon library and UI text must not contain emojis.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/locket-insights-handoff.md` at line 106, Update the
LocketTodayScreen.tsx entry in the handoff table to replace the ⚡ emoji notation
with the intended shared vector-icon name, preserving the described hub-entry
behavior and avoiding emoji UI text.

Source: Coding guidelines

src/services/devInference.ts (2)

27-27: 🚀 Performance & Scalability | 🔵 Trivial | ⚡ Quick win

Avoid the backtracking-prone grammar scan.

Sonar flags this regex as potentially super-linear. Scan individual lines instead so pasted grammars cannot cause avoidable JS-thread work.

Proposed fix

-  if (!/(^|\n)\s*root\s*::=/.test(g)) return 'no "root" rule';
+  const hasRootRule = g.split(/\r?\n/).some(line => /^\s*root\s*::=/.test(line));
+  if (!hasRootRule) return 'no "root" rule';

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/devInference.ts` at line 27, Replace the backtracking-prone
regex in the grammar validation flow with a line-by-line scan of g, checking
each line for a root rule after optional whitespace and confirming the rule
delimiter. Preserve the existing 'no "root" rule' return when no line matches,
and avoid applying a multiline regex to the entire grammar.

Source: Linters/SAST tools

---

31-76: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Split applyDevGrammarOverrides to satisfy the complexity gate.

Sonar reports cognitive complexity 19, above the configured limit of 15. Extract grammar handling, scalar overrides, and logging/error cleanup into focused helpers.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/devInference.ts` around lines 31 - 76, Reduce the cognitive
complexity of applyDevGrammarOverrides by extracting focused helpers for grammar
validation/application, scalar overrides, and final logging/error cleanup. Keep
applyDevGrammarOverrides responsible for the enabled-state check and
orchestration, preserving existing tool removal, invalid-grammar fallback,
temperature/prefix/maxWords updates, logging, stale-error clearing, and return
behavior.

Source: Linters/SAST tools

docs/plans/transcript-quality-experiment.md (1)

10-10: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Add a language hint to the main instructional fenced block (MD040).

📝 Suggested fix

-```
+```text
 You are a researcher solving a hard, open problem: ...

Also applies to: 165-165

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/transcript-quality-experiment.md` at line 10, Add a language
identifier such as text to the main instructional fenced code block in
transcript-quality-experiment.md, including the corresponding fenced block at
the other reported location, so both satisfy MD040 while preserving their
content.

Source: Linters/SAST tools

src/services/llm.ts (1)

20-20: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Missing test coverage for new dev-grammar/retry logic.

The new grammar-override, error-recording, and single-retry-without-grammar paths in generateResponse and generateWithMaxTokens (including the reset bug above) have no accompanying unit tests in this cohort (only llmHelpers.test.ts, whisperService.test.ts, whisperStore.test.ts are present). As per coding guidelines, "Coverage for new code must include both core and pro/, and new code should reach 100% statements, branches, functions, and lines with all branches exercised."

Want me to draft unit tests for LLMService.generateResponse/generateWithMaxTokens covering the grammar-apply, grammar-fail-retry, and non-grammar-fail-rethrow branches?

Also applies to: 235-264, 304-345

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/llm.ts` at line 20, Add unit tests for
LLMService.generateResponse and generateWithMaxTokens covering grammar
application, grammar-override failure with one retry without grammar, error
recording, reset behavior, and non-grammar failure rethrow paths. Provide
equivalent coverage in both core and pro test suites, exercising all new
branches, functions, statements, and lines.

Source: Coding guidelines

docs/plans/locket-llm-insights-plan.md (1)

53-56: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Add a language hint to fenced code blocks (MD040).

Three fenced blocks (pipeline diagrams / data-model snippet) have no language specified, flagged by markdownlint.

📝 Suggested fix

-```
+```text
 conversation/meeting ends (or time-of-day digest) -> local notification

(apply the same text language tag to the blocks at Lines 80 and 93.)

Also applies to: 80-86, 93-101

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/locket-llm-insights-plan.md` around lines 53 - 56, Specify the
text language on the three fenced code blocks in the plan document: the pipeline
diagram near the conversation/meeting flow and the blocks near the referenced
data-model sections. Add the text fence tag while preserving each block’s
existing content.

Source: Linters/SAST tools

src/services/transcriptSummarizer.ts (1)

185-204: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Missing test coverage for abort() and dynamic context sizing.

No test file for TranscriptSummarizerService (covering resolveContextTokens, dynamic INPUT_CONTEXT_FRACTION chunking, grammar/repeatPenalty wiring, or abort()) is included in this cohort. As per coding guidelines, "Coverage for new code must include both core and pro/, and new code should reach 100% statements, branches, functions, and lines with all branches exercised," and per the same guideline, tests here "must cover approved behavior changes ... and assert outcomes/invariants rather than merely asserting that a gate or mock was called."

Happy to draft tests for resolveContextTokens branch coverage (litert/remote/local) and an abort()-mid-map-loop scenario if useful.

Also applies to: 225-329

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/transcriptSummarizer.ts` around lines 185 - 204, Add
comprehensive tests for TranscriptSummarizerService covering
resolveContextTokens across litert, remote, and local branches; dynamic
INPUT_CONTEXT_FRACTION chunk sizing; grammar and repeatPenalty propagation; and
abort() during an active map loop, asserting resulting behavior and invariants
rather than only mock calls. Provide equivalent coverage in both core and pro
test suites, including all relevant statements, branches, functions, and lines.

Source: Coding guidelines

src/services/whisperService.ts (1)

367-391: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Design is sound; minor duplication risk flagged by the code's own comment.

Asset-presence gating (not a user toggle) is well-reasoned and documented. One nit: the .bin → -encoder.mlmodelc derivation is duplicated between coreMLPathFor (Line 100-102) and the inline regex at Line 375 — the comment at Line 98-99 already flags the "keep in lockstep" risk. Consider having coreMLPathFor delegate to a shared path-only helper and reuse it here.

♻️ Optional consolidation

-  private coreMLPathFor(modelId: string): string {
-    return this.getModelPath(modelId).replace(/\.bin$/i, '-encoder.mlmodelc');
-  }
+  private coreMLPathForFile(binPath: string): string {
+    return binPath.replace(/\.bin$/i, '-encoder.mlmodelc');
+  }
+  private coreMLPathFor(modelId: string): string {
+    return this.coreMLPathForFile(this.getModelPath(modelId));
+  }

-      const coreMLPath = modelPath.replace(/\.bin$/i, '-encoder.mlmodelc');
+      const coreMLPath = this.coreMLPathForFile(modelPath);

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/whisperService.ts` around lines 367 - 391, Consolidate CoreML
encoder path derivation by reusing the existing coreMLPathFor helper or a shared
path-only helper in the model-loading block around useCoreML, instead of
duplicating the inline “.bin” to “-encoder.mlmodelc” regex. Preserve the current
iOS-only asset-presence gating and backfill behavior.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@docs/plans/auto-detect-recording-plan.md`:
- Around line 76-93: Update the recording state flow so the IDLE-to-CAPTURING
path flushes only preroll data before writing the current buffer once, removing
the duplicate write caused by flushing the current buffer. In the max-duration
branch, require state == CAPTURING before rolling so a silence-triggered close
cannot reopen a file in the same iteration.

In `@docs/plans/continuous-smart-recorder-plan.md`:
- Line 64: Update the continuous smart recorder plan’s roll behavior to never
silently discard finalized all-noise recordings. Replace the discard requirement
with preserving the recording as an unprocessed item, or retaining a recoverable
original before excluding it from the main feed, while keeping the existing 0:00
guard.
- Around line 104-110: Resolve the contradiction in section “The ‘nothing lost’
guarantee” by renaming the guarantee and revising its contract to acknowledge
that crashes may lose only the last few unflushed seconds. Keep the existing
backup, duration gate, rolling, and crash-safety details, but do not claim that
no audio can ever be lost unless durable flushing before capture acknowledgment
is implemented.
- Around line 46-53: Update the periodic Silero checkpoint flow in section 2b so
the AAC restore snapshot from ensureCompressedBackup reflects the complete
current open file immediately before any prune. Refresh the backup after each
append or prune, or keep 2b disabled until finalization; do not allow
restoration to omit audio recorded after an earlier backup.

In `@src/services/devInference.ts`:
- Around line 84-88: Update noteDevGrammarError to restore the original tools
and tool_choice metadata removed by applyDevGrammarOverrides before
generateWithToolsImpl retries. Preserve the existing grammar cleanup, error
recording, and warning behavior, ensuring the fallback retains the tool request
fields.

In `@src/services/llmToolGeneration.ts`:
- Around line 157-161: Update the catch branch in the completion flow around
safeCompletion so fallback is attempted only when the error is confirmed to be a
dev-grammar rejection. Preserve noteDevGrammarError and the ungrammared retry
for that case, but immediately rethrow cancellation, model availability,
OOM/native, and other non-grammar failures; do not use devGrammarApplied alone
as the retry condition.
- Around line 154-161: Reset the streaming state before the fallback retry in
the catch block of the generateWithTools flow: recreate fullResponse,
collectedToolCalls, tokenCount, firstReceived, and toolCallFilter so the second
safeCompletion call starts clean and cannot duplicate partial output. Preserve
the existing retry behavior and dev-grammar error recording.

In `@src/services/whisperService.ts`:
- Around line 97-170: Prevent ensureCoreMLEncoder from moving or removing the
shared small.en encoder when processing small.en-tdrz. Extract reused encoders
into a per-model scratch location, or otherwise copy/retain the canonical source
directory while placing the target at the requested model path; preserve
ggml-small.en-encoder.mlmodelc for subsequent small.en loads.
- Around line 260-264: Update the iOS CoreML backfill in downloadModel so
ensureCoreMLEncoder(modelId) is launched fire-and-forget rather than awaited,
while retaining its non-fatal error handling and allowing downloadModel to
resolve immediately after the main model is validated.

---

Minor comments:
In `@docs/plans/audio-recorder-notes.md`:
- Line 1: Replace the em dash in the “Audio / recorder work — current state”
title with a permitted colon or hyphen, preserving the rest of the title
unchanged.
- Around line 24-32: Update the fenced code block containing the commit list in
the audio recorder notes to include an explicit language identifier, preferably
text, immediately after the opening fence; leave the listed commits unchanged.

In `@docs/plans/auto-detect-recording-plan.md`:
- Line 71: Update the pseudocode code fence in auto-detect-recording-plan.md to
specify the text language as ```text, preserving the enclosed content unchanged.
- Line 1: Replace the em dash in the “Auto-detect (VAD-gated) recording” title
with an approved non-em-dash separator, preserving the title wording and
meaning.

In `@docs/plans/combined-intelligence-layer-plan.md`:
- Around line 116-127: Update the “Important flags” bullets in the combined
intelligence layer plan to remove all em dashes and replace them with approved
punctuation or sentence structure, while preserving each bullet’s meaning and
existing proof-first tone. Ensure the edited text follows the copy standards in
docs/brand_tone_voice.md, including no em dashes.

In `@docs/plans/continuous-smart-recorder-plan.md`:
- Around line 27-29: Update the fenced code block containing the RecorderMode
type declaration to specify the TypeScript language with a ts fence, ensuring
the documentation satisfies MD040.
- Line 1: Replace the em dash in the “Continuous-Smart Recorder Mode —
Implementation Plan” title with an allowed non-em-dash separator, preserving the
title wording.

In `@docs/plans/insights-page-design-brief.md`:
- Around line 9-153: Update the embedded prompt fence in the design brief to
specify the text language by changing the opening fence to ```text, while
leaving the prompt content and closing fence unchanged.
- Line 1: Update the “Insights Page — Design Brief” heading to replace the em
dash with an approved non-em-dash separator, preserving the title wording.

In `@docs/plans/locket-commit-plan.md`:
- Around line 30-36: Add blank lines immediately before and after the “Do NOT
commit” table in the plan document, keeping the table contents unchanged so it
satisfies the Markdown table-spacing rule MD058.
- Around line 21-27: Update the fenced code block in the locket commit plan to
specify the gitignore language, changing the opening fence to gitignore while
preserving its existing ignore patterns.
- Line 10: Update the commit-time quality-gates statement in the plan to
explicitly require compileDebugKotlin, lintDebug, and npm run test:android for
staged Kotlin/Gradle changes, and staged-file eslint, tsc --noEmit, and npm test
for staged JavaScript/TypeScript changes; retain the instruction to pause and
report native hook failures instead of using --no-verify without approval.

In `@docs/plans/locket-continuous-smart-recorder-handoff.md`:
- Line 168: Update the “Whole-day robustness” heading in the documented plan to
use approved brand vocabulary, replacing “robustness” with “whole-day
reliability” or another concrete description while preserving the heading’s
meaning.
- Around line 136-137: Add a blank line between the “Decided constants” heading
and its table, and ensure the table has the required surrounding blank lines to
satisfy Markdownlint MD058.

In `@docs/plans/locket-insights-handoff.md`:
- Around line 29-37: Update the fenced code blocks in the handoff document,
including the blocks around the pipeline and the referenced later section, to
specify text as the language identifier. Preserve each block’s contents and
formatting while ensuring every fence has an explicit language.
- Line 1: Remove every em dash from the documentation copy in the handoff
document, including the title and the referenced “TL;DR,” “pure heuristics,” and
“llama.rn only” text. Replace each with an appropriate colon, semicolon, or
standard hyphen while preserving the original meaning.

In `@docs/plans/locket-insights-screen-plan.md`:
- Around line 47-53: Add the text language identifier to the fenced
documentation blocks showing the loose output format and router pseudocode,
including the corresponding block at the other referenced section, so every
fence has an explicit language for Markdownlint MD040.

In `@docs/plans/session-handoff-transcription-intelligence.md`:
- Around line 26-27: Update the prose in the session handoff transcription
intelligence plan, especially the bullets “On-device / privacy by architecture,
not by promise” and “Honest limitations over hype,” to remove em dashes and
replace them with brand-compliant punctuation or sentence structure. Apply the
same change to the text around the additionally referenced line, preserving the
meaning and all other brand-tone requirements.
- Around line 207-208: Update the planning document’s “Commit co-author” and iOS
device build instructions to remove the personal email address and device UDID
while preserving the surrounding attribution and build guidance in anonymized
form.

In `@docs/plans/transcript-quality-experiment.md`:
- Line 82: Update the condition_on_previous_text bullet in
transcript-quality-experiment.md to replace the em dash between “--no-context”
and “the loop killer” with approved punctuation, preserving the existing meaning
and wording.

In `@docs/plans/transcript-quality-findings.md`:
- Around line 260-262: The diarization summary in item 5 is stale and
contradicts the later “Diarization status” and “Cost correction” findings.
Update item 5 to describe NME-SC auto-count as solved and reflect the revised
cost of approximately one-third of transcription, removing the
fixed-cluster-count and outdated 10× cost claims.
- Line 11: Replace all em dashes in docs/plans/transcript-quality-findings.md
and its sibling handoff document with brand-compliant punctuation or wording,
including the occurrences near the referenced lines. Preserve the original
meaning and ensure no em dashes remain in either document.

In `@src/screens/ChatScreen/ChatScreenComponents.tsx`:
- Around line 89-90: Update the devActive selector in ChatScreenComponents to
consider both configured constraint sources: keep requiring s.enabled, but treat
a non-empty trimmed grammar or a non-empty trimmed litertConstraintString as
active. Preserve the inactive state when neither constraint is configured.

In `@src/services/litert.ts`:
- Around line 142-155: The new LiteRT paths lack test coverage. Add tests for
getContextTokens() and for resetConversation() with __DEV__ enabled, covering
the setConstrainedDecoding branch and its expected invocation when the dev
constraint is armed, while preserving existing behavior for disabled or empty
constraints.

---

Nitpick comments:
In `@docs/plans/locket-insights-handoff.md`:
- Line 106: Update the LocketTodayScreen.tsx entry in the handoff table to
replace the ⚡ emoji notation with the intended shared vector-icon name,
preserving the described hub-entry behavior and avoiding emoji UI text.

In `@docs/plans/locket-llm-insights-plan.md`:
- Around line 53-56: Specify the text language on the three fenced code blocks
in the plan document: the pipeline diagram near the conversation/meeting flow
and the blocks near the referenced data-model sections. Add the text fence tag
while preserving each block’s existing content.

In `@docs/plans/transcript-quality-experiment.md`:
- Line 10: Add a language identifier such as text to the main instructional
fenced code block in transcript-quality-experiment.md, including the
corresponding fenced block at the other reported location, so both satisfy MD040
while preserving their content.

In `@src/components/DevGrammarModal.tsx`:
- Around line 176-186: In DevGrammarModal, extract the nested ternary used by
the TextInput placeholder into a clearly named variable or helper based on
litertType, then pass that value to placeholder while preserving the existing
json_schema, regex, and default placeholder strings.

In `@src/services/devInference.ts`:
- Line 27: Replace the backtracking-prone regex in the grammar validation flow
with a line-by-line scan of g, checking each line for a root rule after optional
whitespace and confirming the rule delimiter. Preserve the existing 'no "root"
rule' return when no line matches, and avoid applying a multiline regex to the
entire grammar.
- Around line 31-76: Reduce the cognitive complexity of applyDevGrammarOverrides
by extracting focused helpers for grammar validation/application, scalar
overrides, and final logging/error cleanup. Keep applyDevGrammarOverrides
responsible for the enabled-state check and orchestration, preserving existing
tool removal, invalid-grammar fallback, temperature/prefix/maxWords updates,
logging, stale-error clearing, and return behavior.

In `@src/services/llm.ts`:
- Line 20: Add unit tests for LLMService.generateResponse and
generateWithMaxTokens covering grammar application, grammar-override failure
with one retry without grammar, error recording, reset behavior, and non-grammar
failure rethrow paths. Provide equivalent coverage in both core and pro test
suites, exercising all new branches, functions, statements, and lines.

In `@src/services/transcriptSummarizer.ts`:
- Around line 185-204: Add comprehensive tests for TranscriptSummarizerService
covering resolveContextTokens across litert, remote, and local branches; dynamic
INPUT_CONTEXT_FRACTION chunk sizing; grammar and repeatPenalty propagation; and
abort() during an active map loop, asserting resulting behavior and invariants
rather than only mock calls. Provide equivalent coverage in both core and pro
test suites, including all relevant statements, branches, functions, and lines.

In `@src/services/whisperService.ts`:
- Around line 367-391: Consolidate CoreML encoder path derivation by reusing the
existing coreMLPathFor helper or a shared path-only helper in the model-loading
block around useCoreML, instead of duplicating the inline “.bin” to
“-encoder.mlmodelc” regex. Preserve the current iOS-only asset-presence gating
and backfill behavior.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 092b8228-c4c2-4d23-a464-32b49093f7e0

📥 Commits

Reviewing files that changed from the base of the PR and between 87948b6 and 95dfadf.

⛔ Files ignored due to path filters (6)

• android/app/src/main/assets/ggml-hexagon/libggml-htp-v69.so is excluded by !**/*.so
• android/app/src/main/assets/ggml-hexagon/libggml-htp-v73.so is excluded by !**/*.so
• android/app/src/main/assets/ggml-hexagon/libggml-htp-v75.so is excluded by !**/*.so
• android/app/src/main/assets/ggml-hexagon/libggml-htp-v79.so is excluded by !**/*.so
• android/app/src/main/assets/ggml-hexagon/libggml-htp-v81.so is excluded by !**/*.so
• android/app/src/main/assets/whisper-vad/ggml-silero-v5.1.2.bin is excluded by !**/*.bin

📒 Files selected for processing (39)

• .gitignore
• __tests__/unit/locket/recordingImportance.test.ts
• __tests__/unit/locket/recordingInsights.test.ts
• __tests__/unit/locket/recordingInsightsExtractive.test.ts
• __tests__/unit/services/llmHelpers.test.ts
• __tests__/unit/services/whisperService.test.ts
• __tests__/unit/stores/whisperStore.test.ts
• android/app/src/main/java/ai/offgridmobile/litert/LiteRTModule.kt
• docs/plans/audio-recorder-notes.md
• docs/plans/auto-detect-recording-plan.md
• docs/plans/chat-grammar-test-harness-plan.md
• docs/plans/combined-intelligence-layer-plan.md
• docs/plans/continuous-smart-recorder-plan.md
• docs/plans/insights-page-design-brief.md
• docs/plans/locket-commit-plan.md
• docs/plans/locket-continuous-smart-recorder-handoff.md
• docs/plans/locket-insights-handoff-v2.md
• docs/plans/locket-insights-handoff.md
• docs/plans/locket-insights-product-design-handoff.md
• docs/plans/locket-insights-screen-plan.md
• docs/plans/locket-insights-transcript-quality-handoff.md
• docs/plans/locket-llm-insights-plan.md
• docs/plans/locket-recorder-handoff.md
• docs/plans/session-handoff-transcription-intelligence.md
• docs/plans/transcript-quality-experiment.md
• docs/plans/transcript-quality-findings.md
• pro
• src/components/DevGrammarModal.tsx
• src/screens/ChatScreen/ChatScreenComponents.tsx
• src/services/devInference.ts
• src/services/litert.ts
• src/services/llm.ts
• src/services/llmHelpers.ts
• src/services/llmToolGeneration.ts
• src/services/transcriptSummarizer.ts
• src/services/whisperModels.ts
• src/services/whisperService.ts
• src/stores/devInferenceStore.ts
• src/stores/whisperStore.ts

[coderabbitai]

🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Make the state transitions mutually exclusive.

preroll.push(buffer) followed by flush preroll already writes the current buffer, so write(buffer) duplicates it. Also, after closing on silence, the max-duration branch can immediately reopen a file because it is still evaluated in the same iteration. Flush only prior preroll data and guard the roll branch with state == CAPTURING.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/auto-detect-recording-plan.md` around lines 76 - 93, Update the
recording state flow so the IDLE-to-CAPTURING path flushes only preroll data
before writing the current buffer once, removing the duplicate write caused by
flushing the current buffer. In the max-duration branch, require state ==
CAPTURING before rolling so a silence-triggered close cannot reopen a file in
the same iteration.

[coderabbitai]

🗄️ Data Integrity & Integration | 🟠 Major | 🏗️ Heavy lift

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

# Map the relevant plan file and search for related terms across docs.
git ls-files 'docs/plans/continuous-smart-recorder-plan.md' 'docs/**/*.md' | sed -n '1,120p'

echo
echo '--- outline ---'
ast-grep outline docs/plans/continuous-smart-recorder-plan.md --view expanded || true

echo
echo '--- related search ---'
rg -n --hidden -S 'ensureCompressedBackup|checkpoint prune|roll-time post-pass|unflushed slice|finalized|backup|AAC backup|prune gap|continuous-smart-recorder' docs

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 10493

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

# Inspect the target section with line numbers.
sed -n '35,70p' docs/plans/continuous-smart-recorder-plan.md | cat -n

echo
echo '--- nearby references ---'
rg -n --hidden -S '2b\.|3 \+ 5|ensureCompressedBackup|open file|unflushed slice|prune|backup|append' docs/plans/continuous-smart-recorder-plan.md

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 7132

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

# Find implementation or references to ensureCompressedBackup and pruning logic.
rg -n --hidden -S 'function ensureCompressedBackup|ensureCompressedBackup\(|vadDetect|prune gap|checkpoint prune|roll-time post-pass|capture-continuous|trim-at-roll|AAC backup|compressed backup' .

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 6503

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

# Read the plan section and look for any language that says the backup is refreshed
# or tied to finalization/rollover.
sed -n '42,58p' docs/plans/continuous-smart-recorder-plan.md | cat -n

echo
rg -n --hidden -S 'refresh.*backup|backup.*refresh|restore snapshot|finalized|finalization|rollover|roll-time|append.*backup|ensureCompressedBackup' docs/plans docs

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 5535

---

Refresh the restore snapshot before pruning the open file. A backup taken before 2b will not include audio recorded afterward, so restoring it can drop later clips. Rebuild the AAC backup after each append/prune, or keep 2b disabled until the file is finalized.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/continuous-smart-recorder-plan.md` around lines 46 - 53, Update
the periodic Silero checkpoint flow in section 2b so the AAC restore snapshot
from ensureCompressedBackup reflects the complete current open file immediately
before any prune. Refresh the backup after each append or prune, or keep 2b
disabled until finalization; do not allow restoration to omit audio recorded
after an earlier backup.

[coderabbitai]

🗄️ Data Integrity & Integration | 🟠 Major | 🏗️ Heavy lift

Do not silently discard all-noise recordings.

Discarding a finalized file conflicts with the non-destructive rule in Line 167 and the project rule to never auto-delete user audio. Preserve it as an unprocessed recording, or retain a recoverable original before hiding it from the main feed.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/continuous-smart-recorder-plan.md` at line 64, Update the
continuous smart recorder plan’s roll behavior to never silently discard
finalized all-noise recordings. Replace the discard requirement with preserving
the recording as an unprocessed item, or retaining a recoverable original before
excluding it from the main feed, while keeping the existing 0:00 guard.

[coderabbitai]

🗄️ Data Integrity & Integration | 🟠 Major | 🏗️ Heavy lift

Resolve the “nothing lost” guarantee contradiction.

Line 110 explicitly allows crash loss of unflushed audio, so this section cannot promise that nothing is lost. Either implement durable crash-safe flushing before acknowledging capture, or rename the guarantee to state the bounded-loss behavior precisely.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@docs/plans/continuous-smart-recorder-plan.md` around lines 104 - 110, Resolve
the contradiction in section “The ‘nothing lost’ guarantee” by renaming the
guarantee and revising its contract to acknowledge that crashes may lose only
the last few unflushed seconds. Keep the existing backup, duration gate,
rolling, and crash-safety details, but do not claim that no audio can ever be
lost unless durable flushing before capture acknowledgment is implemented.

[coderabbitai]

🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Restore tool metadata on the ungrammared fallback.

applyDevGrammarOverrides removes tools and tool_choice, but noteDevGrammarError only deletes grammar. Therefore, generateWithToolsImpl retries without tool definitions, silently converting a tool request into plain-text generation. Preserve the original tool fields and restore them before retrying.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/devInference.ts` around lines 84 - 88, Update
noteDevGrammarError to restore the original tools and tool_choice metadata
removed by applyDevGrammarOverrides before generateWithToolsImpl retries.
Preserve the existing grammar cleanup, error recording, and warning behavior,
ensuring the fallback retains the tool request fields.

[coderabbitai]

🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

set -euo pipefail
git ls-files src/services/llmToolGeneration.ts
wc -l src/services/llmToolGeneration.ts
sed -n '1,260p' src/services/llmToolGeneration.ts
printf '\n--- SEARCH ---\n'
rg -n "fullResponse|collectedToolCalls|tokenCount|firstReceived|toolCallFilter|safeCompletion|generateWithTools" src/services/llmToolGeneration.ts

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 10589

---

🏁 Script executed:

set -euo pipefail
sed -n '1,260p' src/services/llmHelpers.ts
printf '\n--- DEV INFERENCE ---\n'
sed -n '1,260p' src/services/devInference.ts
printf '\n--- SEARCH PARTIAL/RETRY ---\n'
rg -n "partial|retry|safeCompletion|onCompletionData|tool_calls|firstReceived|tokenCount|fullResponse|toolCallFilter" src/services

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 35941

---

🏁 Script executed:

set -euo pipefail
sed -n '1,260p' src/services/llmSafetyChecks.ts
printf '\n--- LLM RETRY SECTION ---\n'
sed -n '300,350p' src/services/llm.ts

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 9109

---

🌐 Web query:

llama.rn completion callback throws after streaming partial tokens grammar error

💡 Result:

In recent versions of llama.cpp, errors in completion callbacks during streaming—or specifically, exceptions thrown during the final parse—are frequently linked to regressions in the chat parser's PEG implementation [1][2][3]. The issue typically manifests in the following way: 1. During streaming, the server parses output with the is_partial flag set to true, which allows the parser to fall back to partial AST extraction when it encounters tokens that do not fully match the grammar [2][3]. 2. When the generation finishes (or reaches max_tokens), the server performs a final parse on the accumulated text with is_partial set to false [1][2]. 3. In recent builds, this final parse can fail if the parser encounters unexpected input, causing the server to throw a std::runtime_error ("Failed to parse input at pos...") instead of utilizing the partial fallback used during streaming [1][2][3]. This results in an HTTP 500 error, discarding the valid content already sent to the client [1][2]. Technical context and mitigation: - Regression Source: This behavior has been identified in issues following major refactors to the chat parser and PEG implementation (e.g., commit 566059a26) [1][2]. - Error Handling: Earlier issues regarding grammar error reporting in streaming mode have been partially addressed by wrapping sampling functions (like common_sampler_accept) in try/catch blocks [4]. However, the specific "final parse" regression has been the subject of active bug reports and pull requests aimed at ensuring the parser behaves consistently (i.e., gracefully) in non-partial mode [2][3][5]. - Troubleshooting: If you are encountering this, ensure you are using the latest version of llama.cpp, as maintainers have been actively merging patches to improve parser leniency and error handling for partial outputs [3][5]. For development purposes, some users have implemented local patches to catch these specific exceptions and treat them as valid completions rather than server errors [2]. If the crash is due to specific model tokens (e.g., "Unexpected empty grammar stack"), it may also indicate a vocabulary-grammar mismatch specific to certain model variants [6].

Citations:

• 1: Misc. bug: Autoparser throws on final parse after streaming succeeds ggml-org/llama.cpp#20814
• 2: chat: new parser should not crash inference ggml-org/llama.cpp#20708
• 3: Fix chat parser regressions: inference crashes/frozen; output backtracked ggml-org/llama.cpp#20660
• 4: fix: GLM 4.5 streaming tool-call parsing + grammar error handling ggml-org/llama.cpp#19612
• 5: common : gracefully handle incomplete output ggml-org/llama.cpp#20191
• 6: Eval bug: Grammar sampler crash: "Unexpected empty grammar stack" on Gemma 4 <unused> tokens ggml-org/llama.cpp#23677

---

Reset stream state before retrying. The fallback path reuses fullResponse, collectedToolCalls, tokenCount, firstReceived, and toolCallFilter; if the first attempt emits any tokens or tool calls before failing, the second pass can append stale data and duplicate output. Recreate those buffers for the retry, or only retry before any callback fires.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/llmToolGeneration.ts` around lines 154 - 161, Reset the
streaming state before the fallback retry in the catch block of the
generateWithTools flow: recreate fullResponse, collectedToolCalls, tokenCount,
firstReceived, and toolCallFilter so the second safeCompletion call starts clean
and cannot duplicate partial output. Preserve the existing retry behavior and
dev-grammar error recording.

[coderabbitai]

🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

Do not retry every completion failure as a grammar failure.

This branch retries after any rejected completion whenever a grammar was applied. That includes cancellation, model unloads, OOM/native failures, and other availability errors, so an abort can immediately start generation again. Restrict fallback to a confirmed grammar-rejection case and rethrow cancellation and infrastructure failures.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/llmToolGeneration.ts` around lines 157 - 161, Update the catch
branch in the completion flow around safeCompletion so fallback is attempted
only when the error is confirmed to be a dev-grammar rejection. Preserve
noteDevGrammarError and the ungrammared retry for that case, but immediately
rethrow cancellation, model availability, OOM/native, and other non-grammar
failures; do not use devGrammarApplied alone as the retry condition.

[coderabbitai]

🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '\n== file list ==\n'
git ls-files 'src/services/whisperService.ts' 'src/services/whisperModels.ts' | cat

printf '\n== outlines ==\n'
ast-grep outline src/services/whisperService.ts --view expanded || true
ast-grep outline src/services/whisperModels.ts --view expanded || true

printf '\n== relevant line ranges ==\n'
nl -ba src/services/whisperService.ts | sed -n '80,190p'
printf '\n--- whisperModels ---\n'
nl -ba src/services/whisperModels.ts | sed -n '1,220p'

printf '\n== url reuse checks ==\n'
python3 - <<'PY'
from pathlib import Path
import re, json

text = Path('src/services/whisperModels.ts').read_text()
# naive parse of model entries to inspect ids and coreMLUrl fields
for m in re.finditer(r"id:\s*'([^']+)'.{0,300}?coreMLUrl:\s*'([^']+)'", text, re.S):
    print(m.group(1), '=>', m.group(2))
PY

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 3870

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '\n== whisperModels relevant lines ==\n'
sed -n '1,120p' src/services/whisperModels.ts | cat -n

printf '\n== coreMLUrl occurrences ==\n'
python3 - <<'PY'
from pathlib import Path
import re

text = Path('src/services/whisperModels.ts').read_text()
for line in text.splitlines():
    if 'coreMLUrl' in line or "id: 'small.en" in line or "id: 'small.en-tdrz'" in line:
        print(line)
PY

printf '\n== path derivation check ==\n'
python3 - <<'PY'
def core_ml_path_for(model_id):
    return f"/models/ggml-{model_id}.bin".replace(".bin", "-encoder.mlmodelc")

for mid in ["small.en", "small.en-tdrz"]:
    print(mid, "->", core_ml_path_for(mid))

core_url = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en-encoder.mlmodelc.zip"
extracted = core_url.split('/')[-1].removesuffix('.zip')
print("extracted dir from shared URL:", extracted)
PY

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 7488

---

Avoid moving the shared CoreML encoder out from under small.en src/services/whisperService.ts:117-160
small.en-tdrz reuses small.en’s CoreML zip, but ensureCoreMLEncoder() unzips it under the shared top-level dir and then renames that dir to the tdrz path. If small.en is already installed, this removes its canonical ggml-small.en-encoder.mlmodelc directory, and the next small.en load falls back to CPU. Extract into a per-model scratch dir, or skip the move when the source encoder is shared.

🧰 Tools

🪛 GitHub Check: SonarCloud Code Analysis

[warning] 162-162: Refactor this code to not use nested template literals.

See more on https://sonarcloud.io/project/issues?id=off-grid-ai_off-grid-ai-mobile&issues=AZ9X3wdPdq34Lg8Wn6fk&open=AZ9X3wdPdq34Lg8Wn6fk&pullRequest=532

---

[warning] 162-162: Refactor this code to not use nested template literals.

See more on https://sonarcloud.io/project/issues?id=off-grid-ai_off-grid-ai-mobile&issues=AZ9X3wdPdq34Lg8Wn6fj&open=AZ9X3wdPdq34Lg8Wn6fj&pullRequest=532

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/whisperService.ts` around lines 97 - 170, Prevent
ensureCoreMLEncoder from moving or removing the shared small.en encoder when
processing small.en-tdrz. Extract reused encoders into a per-model scratch
location, or otherwise copy/retain the canonical source directory while placing
the target at the requested model path; preserve ggml-small.en-encoder.mlmodelc
for subsequent small.en loads.

[coderabbitai]

🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

# Map the service and locate the relevant symbols
ast-grep outline src/services/whisperService.ts --view expanded || true

printf '\n--- relevant matches ---\n'
rg -n "ensureCoreMLEncoder|downloadModel|downloadFile|Platform.OS === 'ios'|fire-and-forget|backfill" src/services/whisperService.ts

printf '\n--- surrounding lines around the awaited call ---\n'
sed -n '230,300p' src/services/whisperService.ts

printf '\n--- surrounding lines around the fire-and-forget backfill mention ---\n'
sed -n '360,420p' src/services/whisperService.ts

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 10727

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

# Inspect the CoreML helper and its timeout/error handling
rg -n "function ensureCoreMLEncoder|ensureCoreMLEncoder\\(|downloadFile\\(|setTimeout|timeout|RNFS\\.downloadFile" src/services/whisperService.ts src -g '!**/node_modules/**'

printf '\n--- CoreML helper region ---\n'
sed -n '1,220p' src/services/whisperService.ts

printf '\n--- remainder if helper is lower in file ---\n'
sed -n '220,520p' src/services/whisperService.ts

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 41463

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

# Read only the relevant implementation to see whether the CoreML fetch has its own timeout / early return.
rg -n "ensureCoreMLEncoder|downloadFile|timeout|connectionTimeout|readTimeout|catch\\(\\) \\{\\}" src/services/whisperService.ts

printf '\n--- core implementation slice ---\n'
sed -n '300,460p' src/services/whisperService.ts

Repository: off-grid-ai/off-grid-ai-mobile

Length of output: 8035

---

Don't block downloadModel() on the CoreML backfill.
ensureCoreMLEncoder() calls RNFS.downloadFile() with no timeout, so a stalled iOS encoder fetch can keep the model-download promise open even after the .bin is validated. Make this fire-and-forget like the load-time backfill.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/whisperService.ts` around lines 260 - 264, Update the iOS CoreML
backfill in downloadModel so ensureCoreMLEncoder(modelId) is launched
fire-and-forget rather than awaited, while retaining its non-fatal error
handling and allowing downloadModel to resolve immediately after the main model
is validated.

[dishit-wednesday]

Re: deleteModel leaves deleted model in presentModelIds — fixed in fd8df7e. deleteModel now recomputes presentModelIds from disk after deleting the active model (dropping the just-deleted one), matching deleteModelById.

[gitar-bot]

⚠️ Quality: New Toast component and summarizer fallback logic ship without tests

Per the project's testing standard (100% coverage on new code, both unit and integration, plus a fails-before/passes-after regression test), the following new/changed logic has no dedicated tests:

• src/components/Toast.tsx (entirely new: show/hide store, nonce-based timer restart, fade-in/out lifecycle) has no test file.
• The new remote-preferred selection (isRemoteActive, resolveContextTokens) and the remote-before-streaming fallback in generateSummaryText (emittedAny gate, fall-through to LiteRT/local) are only exercised indirectly via a mocked summarizer in __tests__/unit/locket/recordingInsights.test.ts, not unit-tested directly.

Add unit tests covering: (1) remote preferred even when a local model is loaded; (2) remote failure BEFORE any token → falls through to on-device; (3) remote failure AFTER a token → error surfaced (no double-write); (4) Toast timer reset on repeated showToast with identical text (nonce). Include a regression test for the abort/remote gap above.

_{Was this helpful? React with 👍 / 👎}

[dishit-wednesday]

Acknowledged — deferring to a follow-up. Toast and the summarizer fallback are earlier-branch code; adding their tests is out of scope for this review-fix pass, tracked separately.

[gitar-bot]

Understood — tracking Toast and summarizer fallback test coverage as a follow-up rather than blocking this PR.

[gitar-bot]

Code Review ⚠️ Changes requested 8 resolved / 9 findings

Refactors summarization and Whisper management with dynamic context sizing and CoreML support, but the new Toast component and summarizer fallback logic require test coverage to meet project standards.

⚠️ Quality: New Toast component and summarizer fallback logic ship without tests

📄 src/components/Toast.tsx:1-15 📄 src/services/transcriptSummarizer.ts:98-100 📄 src/services/transcriptSummarizer.ts:110-122 📄 src/services/transcriptSummarizer.ts:142-156

Per the project's testing standard (100% coverage on new code, both unit and integration, plus a fails-before/passes-after regression test), the following new/changed logic has no dedicated tests:

• src/components/Toast.tsx (entirely new: show/hide store, nonce-based timer restart, fade-in/out lifecycle) has no test file.
• The new remote-preferred selection (isRemoteActive, resolveContextTokens) and the remote-before-streaming fallback in generateSummaryText (emittedAny gate, fall-through to LiteRT/local) are only exercised indirectly via a mocked summarizer in __tests__/unit/locket/recordingInsights.test.ts, not unit-tested directly.

Add unit tests covering: (1) remote preferred even when a local model is loaded; (2) remote failure BEFORE any token → falls through to on-device; (3) remote failure AFTER a token → error surfaced (no double-write); (4) Toast timer reset on repeated showToast with identical text (nonce). Include a regression test for the abort/remote gap above.

✅ 8 resolved

✅ Bug: Ungrammared retry re-uses stale streaming accumulators

📄 src/services/llm.ts:243-257 📄 src/services/llmToolGeneration.ts:140-154
When a dev-grammar completion fails and is retried ungrammared, the streaming state is not reset before the second ctx.completion call. In generateResponse (llm.ts), fullContent, fullReasoningContent, streamedContentSoFar, and streamedReasoningSoFar carry over from the failed attempt; in generateWithToolsImpl (llmToolGeneration.ts), fullResponse and collectedToolCalls carry over. If the first attempt streamed any tokens before throwing, the retry appends to (and re-emits via onStream) the earlier partial output, producing duplicated/garbled content and potentially duplicated tool calls. Note generateWithMaxTokens already handles this correctly by resetting fullResponse = '' before its fallback (llm.ts:339). Reset the accumulators in the catch block before retrying (and clear the ToolCallTokenFilter buffer in the tools path).

✅ Bug: deleteModel leaves deleted model in presentModelIds

📄 src/stores/whisperStore.ts:186-200
The deleteModel action unloads and deletes the active model, then sets downloadedModelId to a fallback and isModelLoaded: false, but it never updates presentModelIds. The just-deleted model id therefore stays in presentModelIds, so any UI driven off that list (Download Manager / model picker) will keep showing a model whose file is gone on disk. The sibling deleteModelById (line 224-227) and refreshPresentModels (line 248/257) both correctly recompute the present list via listDownloadedModels(); deleteModel should do the same to stay consistent.

✅ Quality: Accidentally committed scratch file android/temp.txt (1532 lines)

📄 android/temp.txt:1-15
android/temp.txt is a new 1532-line file that contains a raw AI-assistant chat/planning transcript (design discussion about VAD/continuous recording, embedded shell command output, etc.), not source code. It is placed under android/ where it is not covered by any .gitignore rule added in this PR, and clearly was committed by mistake. It bloats the repo, will confuse reviewers/tooling, and has nothing to do with the feature. Remove it (and consider adding an ignore rule for such scratch files).

✅ Quality: Stray design note redesing.md committed with broken image refs

📄 redesing.md:1-15 📄 .gitignore:91-95
redesing.md is a new top-level markdown file (note the misspelled filename 'redesing') containing an insights-redesign planning note. It references local images image-1.png..image-4.png that are not part of the repo (and the .gitignore in this same PR explicitly ignores /image-*.png), so the images will never render. This looks like a scratch/handoff doc committed to the repo root by accident. If it is intended documentation, move it under docs/plans/ with a corrected filename and either commit or remove the image references; otherwise delete it.

✅ Quality: Verbose DIAG-UAF logging left in the LiteRT hot path

📄 android/app/src/main/java/ai/offgridmobile/litert/LiteRTModule.kt:102-106 📄 android/app/src/main/java/ai/offgridmobile/litert/LiteRTModule.kt:331-332 📄 android/app/src/main/java/ai/offgridmobile/litert/LiteRTModule.kt:453 📄 android/app/src/main/java/ai/offgridmobile/litert/LiteRTModule.kt:602 📄 android/app/src/main/java/ai/offgridmobile/litert/LiteRTModule.kt:611
Several Log.i(TAG, "DIAG-UAF ...") diagnostic lines were added on conversation create/send/close, plus the convSeq/currentConvSeq bookkeeping. The commit message and code comments explicitly say to revert the DIAG-UAF logging once the fix is confirmed. These fire on every send (hot path) and leak conversation identity/thread info to logs. Before merge, either remove them or move them behind the project's debug-log gate (src/utils/debugLogFile.ts device-log path) so production builds don't emit them on every generation.

...and 3 more resolved from earlier reviews

🤖 Prompt for agents

Code Review: Refactors summarization and Whisper management with dynamic context sizing and CoreML support, but the new Toast component and summarizer fallback logic require test coverage to meet project standards.

1. ⚠️ Quality: New Toast component and summarizer fallback logic ship without tests
   Files: src/components/Toast.tsx:1-15, src/services/transcriptSummarizer.ts:98-100, src/services/transcriptSummarizer.ts:110-122, src/services/transcriptSummarizer.ts:142-156

   Per the project's testing standard (100% coverage on new code, both unit and integration, plus a fails-before/passes-after regression test), the following new/changed logic has no dedicated tests:
   
   - `src/components/Toast.tsx` (entirely new: show/hide store, nonce-based timer restart, fade-in/out lifecycle) has no test file.
   - The new remote-preferred selection (`isRemoteActive`, `resolveContextTokens`) and the remote-before-streaming fallback in `generateSummaryText` (`emittedAny` gate, fall-through to LiteRT/local) are only exercised indirectly via a mocked summarizer in `__tests__/unit/locket/recordingInsights.test.ts`, not unit-tested directly.
   
   Add unit tests covering: (1) remote preferred even when a local model is loaded; (2) remote failure BEFORE any token → falls through to on-device; (3) remote failure AFTER a token → error surfaced (no double-write); (4) Toast timer reset on repeated `showToast` with identical text (nonce). Include a regression test for the abort/remote gap above.

Options

Auto-apply is off → Gitar will not commit updates to this branch.
Display: compact → Showing less information.

Comment with these commands to change the behavior for this request:

Auto-apply	Compact
gitar auto-apply:on	gitar display:verbose

_{Was this helpful? React with 👍 / 👎 | Gitar}

[sonarqubecloud]

Quality Gate passed

Issues
16 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud