Repository navigation
Description
Umbrella #4368. This narrows the original proposal to type through the test IME; fill already independently verifies field content and must retain its single verification owner.
Source observations on main ddbfc85: TestInputMethodService drops input for a null InputConnection, rejects base64 over64000 characters, truncates text at32000 UTF-16 units and ignores commitText's boolean. ime-helper.ts checks broadcast process exit only; typeAndroidImeHelper can report success despite a dropped/truncated payload. No runtime reproduction is claimed yet.
Required outcome: Never silently truncate or label rejected/no-receiver/failed editor dispatch as successful type. Preserve permission gates and use typed outcomes/partial-dispatch evidence. Split normal type payloads within supported byte/UTF-16 limits on code-point boundaries. An ordered broadcast acknowledgement can prove receiver acceptance and attempted length/commitText return, not final field content: editors may transform text. Do not call acknowledgement length committed length. Do not add a second fill read-back.
Feasibility gate: One bounded owned-device experiment with a throwaway helper must show ordered result code/data through the actual exported permission-gated receiver, a no-receiver negative and privacy in broadcast/log output. AOSP API23 and main have matching ordered/result surfaces; no three-API prerequisite is justified without contrary evidence. Result data carries only codes and lengths, never user text. No runtime or log-privacy proof exists yet.
Before source implementation, adopt the exact producer-owned acknowledgement protocol, result/error/partial-dispatch semantics, old provider-helper handling and any protocol migration. Proposed v2 strict refusal is not yet adopted. If the acknowledgement surface cannot be established, stop and redesign at the existing confirmation owner; do not add an implicit fallback.
Completion: current trigger reproduction through type, verified negative cases, minimal helper/host owner change, independent correctness review, request-route and native tests for oversized/non-ASCII/no-editor and normal type/fill controls, meaningful regression/mutation detection, exact-head validation with #4050 fallback, and clear help/docs semantics. Use T3 device tools and returned CLI for app/device actions. Do not duplicate helper auth packaging ownership, weaken permission tests, merge/release/tag or maintain unrelated PRs.
Activity
apex-sentinel-staging commentedon Nov 13, 2026
Apex Sentinel · Issue review
Suggested label:
featureThe issue frames the work as an umbrella proposal to add truthful typed outcomes and an acknowledgement protocol for Android IME type behavior, rather than presenting a reproducible broken test; its detailed design and completion criteria are actionable enhancement scope and do not request a specific triage verdict.