[dzianisv]

Fixes #194
Fixes #156

What

Re-implementation (from scratch, per the Stage 1 context pack) of the Android keyboard fix — no candidate PR merged as-is:

• Base mechanism from fix(android): keep composer synchronized with IME #201 (head 0bb9af0): live IME height via Keyboard.metrics() + keyboardDidShow/keyboardDidHide listeners, applied as paddingBottom, explicitly reset to 0 on hide (src/lib/keyboard-offset.ts + test, session-screen wiring).
• (b) coverage from fix(android): keep IME prompts usable #208 (head 95b001d): bounded multiline composer (composerMaxHeight from live IME geometry, internal scroll via scrollEnabled) + scrollable QuestionPrompt with pinned footer (src/lib/session-layout.ts + test, composer cap, QuestionPrompt.tsx).

API-35 decision (the mandatory design point)

JS padding applies only below API 35; the native IME-insets listener owns API 35+. MainActivity.kt / plugins/withKeyboardInsets.js install a native listener gated on SDK_INT >= VANILLA_ICE_CREAM that adds the IME height to the root view's bottom padding. The new jsKeyboardPaddingApplies(platform, apiLevel) helper gates the JS padding to Platform.Version < 35, so exactly one layer applies the IME height on every API level — no double-padding. Chosen over removing/extending the native plugin because the native fix is the working, shipped behavior on 15+; the JS gate is minimal and reversible.

Deliberate non-ports

• fix(android): keep IME prompts usable #208's static insets.top keyboardVerticalOffset is not ported (Stage 1: it measured 56 dp short on a Pixel 11 Pro, Android: chat input hidden behind keyboard #194). session-layout.ts owns only the composer cap + question scroll props; the offset lives in keyboard-offset.ts. Its two offset tests were dropped from the ported test file (the iOS case is already covered by keyboard-offset.test.ts).
• Not taken: fix(keyboard): restore adjustResize for edge-to-edge displays (v0.4.16) #206 (stale v0.4.16 release commit), fix(chat): keyboard cover input + blank card #203 (new dependency react-native-keyboard-controller — owner call needed, not granted), fix(android): prevent keyboard overlap across input screens #193's adjustPan.
• iOS unchanged: behavior="padding", offset 90.

Diff

6 files, +217/−14 (vs main e8444931):

• app/session/[id].tsx (+50/−12): keyboard listeners, JS padding, composer cap
• src/components/chat/QuestionPrompt.tsx (+9/−7): scrollable body, pinned footer
• src/lib/keyboard-offset.ts (new) + src/lib/keyboard-offset.test.ts (new, 10 tests)
• src/lib/session-layout.ts (new) + src/lib/session-layout.test.ts (new, 5 tests)

Connection add/edit screens need no change: already KAV-wrapped on main (merged #216).

Gate evidence (on head 484e227, this VM — no Android SDK/device)

• npm run check:versions — exit 0
• npm run typecheck — exit 0
• npm test (full suite) — 395 pass, 0 fail (baseline on main: 380 pass — +15 tests, test files 48 → 50; 2 ported offset tests from fix(android): keep IME prompts usable #208 dropped, see above)
• No new dependencies; applicationId untouched (cc.agentlabs.opencode).

Open residual (Stage 5)

Physical-device proof (open / multiline / dismiss) was not run — this VM has no Android SDK or device. Needs a real Android phone before merge, including the API-35 vs pre-35 check that exactly one layer applies the IME height.

[dzianisv]

Stage 4 review — PASS

Independent review of head 484e227d305ae76bfaca1be94f2d34a4965ed81d (fix/keyboard-ime-194, 6 files +217/−14). I read the full diff, src/lib/keyboard-offset.ts, src/lib/session-layout.ts, app/session/[id].tsx (in situ), QuestionPrompt.tsx, MainActivity.kt, plugins/withKeyboardInsets.js, the context pack (issue #194 comment 6855205733), and issues #194/#156. No validated defects found; no FAIL findings.

Acceptance criteria

• (a) Composer above keyboard on open — Android applies the IME-reported height as paddingBottom via keyboardDidShow/Keyboard.metrics() (app/session/[id].tsx:99-130, 665-681). This is fix(android): keep composer synchronized with IME #201's mechanism, which passed a full open/close cycle on a physical Pixel 11 Pro (Android: chat input hidden behind keyboard #194 comments); fix(android): keep chat composer above the keyboard (#194, #156) #236's session code is that mechanism plus keyboardY tracking and the API-35 gate.
• (b) Multiline typing — composer maxHeight cap (composerMaxHeight, 88–120 dp, font-scale aware) + scrollEnabled on the TextInput; QuestionPrompt body is a ScrollView with pinned footer. The QuestionPrompt.tsx patch is byte-identical to fix(android): keep IME prompts usable #208's, which passed QA on a physical Pixel 8 Pro.
• (c) Returns to bottom on dismiss — keyboardDidHide explicitly resets height to 0 and keyboardY to null; Android KAV behavior={undefined} + offset 0, so no stale padding path remains (the RN 0.81 stale-offset bug cited in the code is avoided by construction).
• (d) Connection add/edit — unchanged by this PR, KAV-wrapped scrollable forms (app/connection/add.tsx, app/connection/[id].tsx, via fix: handle PermissionV2 events and keyboard avoidance on connection screens #216).
• (e) No iOS regression — iOS path is byte-identical in behavior: behavior="padding", offset 90 (keyboardVerticalOffset returns IOS_KEYBOARD_VERTICAL_OFFSET for iOS regardless of insets; tested).

API-35 guard — exactly one layer applies IME height on every API level (verified)

• MainActivity.kt:28 / plugins/withKeyboardInsets.js (wired in app.json:51): native IME-insets listener gated on SDK_INT >= VANILLA_ICE_CREAM (API 35).
• jsKeyboardPaddingApplies() (src/lib/keyboard-offset.ts:28-32): true only for Android with numeric Platform.Version < 35; false on iOS, on API 35+, and on non-numeric versions (fail-safe off, so no double-apply). JS padding is applied only through this gate ([id].tsx:123-126).

Test-count delta — verified by my own runs

• Baseline main e8444931: 380 pass / 0 fail, 48 test files (worktree run).
• PR head 484e227d: 395 pass / 0 fail, 50 test files; typecheck exit 0, check:versions exit 0 (0.4.17/44).
• Delta +15 tests: fix(android): keep composer synchronized with IME #201's 6 tests ported verbatim + 4 new API-35 guard tests (keyboard-offset.test.ts, 10 total); fix(android): keep IME prompts usable #208's 5 composer/question tests ported verbatim + its iOS-offset test moved to keyboard-offset.test.ts. The one dropped test asserted fix(android): keep IME prompts usable #208's static insets.top Android offset — correctly not ported (measured 56 dp / 106 px short on Pixel 11 Pro, Android: chat input hidden behind keyboard #194 comments). No mock-based tests; all 15 map to (a)/(b)/(c)/(e).

No scope creep

No applicationId change, no new dependencies, no manifest/gradle/package.json changes — the diff touches only the 6 listed files.

Merge gate — NOT green yet, do not merge on this verdict alone

Branch protection on main (strict) requires exactly: "Typecheck and unit tests" ✅ pass, "test" ✅ pass, "build" ⏳ pending. mergeable: true, mergeable_state: blocked (required build pending). CUA is not a required check: cua-smoke.yml is absent (deleted in e8444931); the checkout AGENTS.md "MANDATORY CUA before merge" text is stale — the actual gate is the three checks above.
Stage 5 gates before merge: (1) build green on this head; (2) physical-device proof (open/multiline/dismiss on a real Android phone, API <35 and 35+), which no run in this review could do — no Android SDK/device on this VM.

Filed notes (not failures, out of this PR's composer + connection-forms scope)

• app/(tabs)/index.tsx:599 (New Session modal) and :819 (Rename modal, autoFocus TextInput) still use KeyboardAvoidingView behavior="height" on Android — same KAV-on-Android class, untouched by this PR; candidates for a follow-up task.
• Connection forms on API <35 still rely on KAV behavior="padding" (offset 0) — mitigated by their scrollable ScrollView forms; worth a device check alongside (2) above.
• One correction to the context pack: it says "2 dropped tests" — only one test's assertions were abandoned (the static Android offset test); the iOS-offset test was moved, not dropped.