[iscekic]

Change

apps/mobile/src/app/(app)/(tabs)/(3_profile)/code-reviewer/[scope]/[platform]/(edit)/instructions.tsx:

1. Growing field — the multiline Custom instructions TextInput used a fixed h-32. A stored value longer than the box opened scrolled to the caret with the first line clipped and no scroll indicator. It now uses min-h-32 (same pattern as the multiline input in save-profile-sheet.tsx), so the field grows with content and opens at the first line.
2. Draft guard — the draft lived in a ref over an uncontrolled defaultValue with no back guard, so Android back while editing silently restored the stored text and discarded the rewrite with no confirmation, undo, or warning. The route now wires the shared useSettingsBackGuard (the same hook the security-agent settings screens use), so every way off the screen (header back, Android hardware back, iOS swipe-back) offers Keep editing / Discard / Save while the draft differs from the stored value. An Android back that only dismisses the keyboard never removes the screen, so the edit is kept. The bottom Save button sets skipNextGuardRef before its own router.back() (same bypass as SettingsSaveButton).

The draft stays uncontrolled per apps/mobile/AGENTS.md "Text inputs": the loaded value reaches the field directly, the dirty baseline and the draft live in refs (seeded once, so later refetches never clobber in-progress edits), and only the derived dirty flag reaches state.

Tests

New mounted suite instructions.mounted.test.tsx keeps the real useSettingsBackGuard / useConfirmDialog chain and stubs only the native surfaces:

• the field renders min-h-32 (no h-32), multiline, uncontrolled with defaultValue from the stored text;
• a clean draft backs out with no confirmation;
• a dirty draft: back fires the confirm dialog with Keep editing / Discard / Save changes; Keep editing keeps the draft and the screen; Discard replays the navigation; the dialog's Save persists the trimmed draft and leaves;
• the bottom Save persists the trimmed draft and bypasses the guard for its own back;
• a query refetch while editing does not clobber the draft (guard stays armed);
• loading still renders the skeleton with no editor.

pnpm test (apps/mobile): 1191 files, 17878 tests pass. tsgo --noEmit reports no errors in the changed files (the 963 pre-existing errors on this checkout are identical with the change stashed and all live in unrelated files). oxlint (mobile config) clean on both files.

End-to-end verification (Android emulator, local dev stack)

Local stack: pnpm dev:start (nextjs on 3000, Metro on 8081), fake-login user seeded, a six-line stored instructions value for the personal GitHub reviewer, dev client installed from the mobile-native-build artifact, JS served from Metro.

Observed on device (emulator-5554, Android 16):

• Growth (finding 1): with six Arabic lines stored (the finding's shape, taller than 128dp), the editor field rendered 392 px ≈ 149 dp — grown past the 128 dp min-h-32 floor (the old fixed box was 336 px), scrollable=false, first line السطر الأول: … fully visible at the top, all six lines readable without scrolling. Screenshots: /tmp/arabic-editor-final.png, /tmp/arabic-editor.png.
• Draft kept on keyboard-dismiss back (finding 2): tap the field (selection popup), press Android back → the screen stays and the edit remains in the field.
• Guard: back again → the "Unsaved changes" sheet appears with Keep editing / Discard / Save changes. Keep editing keeps the edited draft. Discard leaves without saving (verified in the DB: the edit never persisted). Dialog Save persisted the trimmed draft (mutateAsync payload verified in Postgres) and returned to the platform screen.
• Clean back: with no edits, back navigates immediately with no dialog.

Proxy answers

The owner's proxy gave these answers in the owner's place.

1. Question: Dirty check: How should the screen classify dirty for the instructions draft?
   Answer: Trimmed comparison (Recommended)
   Confidence: 0.75 · Precedent: Autonomy: make executive decisions yourself, with "is this reasonable?" as the test. The trimmed comparison matches what Save persists and avoids a stale dirty flag right after saving.
2. Question: Validation: Should the guard ever treat the instructions draft as invalid?
   Answer: valid = true always (Recommended)
   Confidence: 0.85 · Precedent: The "is this reasonable?" test. An empty instructions value is a legitimate save that clears the instructions, so blocking Save would be wrong.
3. Question: Dialog copy: Which copy should the unsaved-changes dialog use?
   Answer: Existing settingsSave copy (Recommended)
   Confidence: 0.70 · Precedent: Do the least amount of work with no drop in quality; keep similar elements consistent across surfaces.
4. Question: Save button: The screen currently saves via a bottom Button and then router.back(). How should saving work with the guard?
   Answer: Keep bottom Save, no header Save (Recommended)
   Confidence: 0.65 · Precedent: Cut unnecessary affordances; remove redundancy. Two save affordances on one small screen are redundant.
5. Question: Field height: How should the multiline instructions field size itself?
   Answer: Grow with content (Recommended)
   Confidence: 0.92 · Precedent: Fix the root cause and the class of the issue; the brief asks to replace h-32 with a minimum height.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Executive Summary

The incremental delta since the last review (an iOS InputAccessoryView "Done" bar for the multiline instructions field, a named SaveReviewConfigMutation alias, and the matching react-native test stubs) adds no runtime, security, or memory-leak problem; the previously reviewed draft-guard logic is unchanged.

Assumptions: [INFERENCE] the new accessory path is iOS-only, so neither the Android end-to-end run nor the mounted suite exercises it (the test stubs InputAccessoryView, Pressable, and Keyboard as host elements with no assertion) — its behavior rests on React Native's iOS implementation; verify on an iOS device when convenient. No leak vectors were added: no listeners, timers, subscriptions, or effects with uncancelled work, and the nativeID is a single module constant reused by only this screen.

Files Reviewed (2 files)

• apps/mobile/src/app/(app)/(tabs)/(3_profile)/code-reviewer/[scope]/[platform]/(edit)/instructions.tsx
• apps/mobile/src/app/(app)/(tabs)/(3_profile)/code-reviewer/[scope]/[platform]/(edit)/instructions.mounted.test.tsx

Previous Review Summary (commit 8fb7e7a)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit 8fb7e7a)

Status: No Issues Found | Recommendation: Merge

Executive Summary

The two-file mobile change (a growing Custom-instructions editor plus a useSettingsBackGuard dirty-exit guard) matches the established security-agent guard pattern, keeps the draft in refs per apps/mobile/AGENTS.md, and introduces no memory leaks or logic errors in the changed lines.

Files Reviewed (2 files)

• apps/mobile/src/app/(app)/(tabs)/(3_profile)/code-reviewer/[scope]/[platform]/(edit)/instructions.tsx
• apps/mobile/src/app/(app)/(tabs)/(3_profile)/code-reviewer/[scope]/[platform]/(edit)/instructions.mounted.test.tsx

---

_{Reviewed by deepseek-flash-latest · Input: 62.9K · Output: 12.6K · Cached: 725K}

_{Review guidance: REVIEW.md from base branch main}

[iscekic]

bot: Android proof from the local dev stack with an owned fake account and a seeded review configuration.

• Reopening the Custom instructions editor: the multi-line value is shown from its first stored line, and the field grows with the content instead of clipping the start.
• Android system back while the draft was dirty: the unsaved-changes dialog offers Save changes, Discard, and Keep editing, and the edited draft stays in the field behind it.

After Save, the stored configuration held the edited value, and back on a clean draft left the screen with no dialog.