- Decision
- ADR-191
- Date
- 2026-10-23
- Status
- ProposedProposed
- Part
- Passwords
- Mentions
- ADR-146, ADR-185
- Mentioned by
- ADR-192, ADR-207
1Decision
Show the logins of the Passwords app as chips in the suggestion strip of the on-screen keyboard (the second plan of ADR-185): 2026-10-22-keyboard-autofill-chips.md, source facts in inline-autofill-chip-path.md.
(1) LatinIME opts in with supportsInlineSuggestions and not with the touch-exploration variant (TalkBack gets the system list), asks for chips with the style of the design page, and shows them in a new row of the strip that has priority over the paste actions and the word suggestions and keeps the strip visible on a password field; the system draws a chip in ExtServices and the keyboard receives a surface only.
It is the patch package keyboard/autofill-chips, applied on top of keyboard/restyle. (2) Chips go only to the system keyboard; any other keyboard gets the system list (default deny). (3) Every dataset field has a null filter, so a chip or list entry goes at the first typed character and the system never matches typed text against the stored value; a typed or deleted character ends the offer for that field until the next field gets focus.
(4) The keyboard gives each chip an exact size (30 dp high, 260 dp alone, 158 dp when several share the strip), at most five chips, the row scrolls. (5) The switch "Suggestions above the keyboard" governs the automatic offer (chip and list), the switch of a login governs it while the vault is open, a manual request always fills, and after the unlock every account of the site is offered.
(6) Build guards fail on a keyboard source that reads a suggestion, on a network, frame-buffer or log permission in the keyboard, and on a dataset value without the null filter.
2Why and evidence
The owner agreed on 2026-10-22 that this plan is written; the chips on focus, the generic locked chip, accounts before the paste chip and the username only are ADR-185 and the design.
Points (2) to (6) are proposals that follow from what a keyboard could learn from a chip (that a site has an account, how many accounts, whether the vault is locked, the length of a name) and from the source facts of the record: the system hosts the chip in another process, the default filter is a prefix match of the stored value, and a size set by the keyboard hides the length of the text.
Prior art: development/samples/AutofillKeyboard and InlineFillService in the tree, androidx.autofill:autofill:1.4.0-alpha01 in prebuilts/sdk/current/androidx, the restyle design section 8 and ADR-146; no earlier record designs the strip row or its priority.
3Consequences
Written, landed and module-built on 2026-10-29 (ACH-001 to ACH-004); an image and every device row (ACH-005 to ACH-016) need the owner's yes and are not done. The switches of the Passwords app start to do something. ExtServices, which draws the chip, holds INTERNET in the source tree and sees the usernames of the chips; restricting it is a follow-up outside this decision. The chip is 30 dp high and the paste chip in the tree 28 dp (compared in ACH-009). A third-party keyboard never gets chips even if it supports them.
4Review when
A device shows that the chip does not render, takes focus, collapses the keyboard or is not drawn after the unlock; ExtServices can send chip text off the phone and the owner wants usernames kept from it (then generic chips only, or a network restriction); a change of the inline suggestion framework or of androidx.autofill; the owner wants chips for another keyboard or content-sized chips.
Read from docs/governance/decision-log.md, entry ADR-191.