Limrun Android: `fill` inserts at the cursor instead of replacing the field's text
メンテナーはふだん 1 日以内に返信
関連するプルリクエストがすでにマージされています。
- #3362 @thymikee による — マージ済み
評価
- 難易度
- 3/5
- 見積もり時間
- 1〜2日
- 初心者へのやさしさ
- 62/100
- issue の種類
- バグ
- 明瞭さ
- 明確に書かれている
- 活発さ
- 活発
- 技術スタック
- android, typescript
- 領域
- mobile
調査の方向性
Start in the Android fill path where agent-device returns after the provider-native provider.text call, since that call skips the adb clear, verify and retry steps (KEYCODE_MOVE_END and KEYCODE_DEL). The Limrun instance client's setText and pressKey calls are the relevant entry points. Done means fill on a field with existing text leaves exactly the new value, and a mismatch is reported rather than passing silently. Reproduce with the Settings search field commands in the issue.
索引モデルが issue の本文から書いたものです。
説明
Summary
On a Limrun Android instance, fill does not replace the existing contents of a text field. The new text is inserted wherever the tap leaves the cursor, so filling a field that already has a value produces a concatenation of the old and new text. The help text says fill <target> <text> replaces, and that is what happens on local emulators.
Environment
- agent-device: 0.21.15, and the same code path is present in 0.21.24 (latest at the time of writing)
- Provider: Limrun, Android instance (
agent-device connect limrun --platform android) - Keyboard on the instance:
com.android.inputmethod.latin(the agent-device test IME is not active) - App: the system Settings app (any
EditTextreproduces it)
Steps to reproduce
agent-device connect limrun --platform android --run-id fill-repro --session repro
agent-device open settings --platform android --session repro
agent-device snapshot -i --platform android --session repro # find "Search Settings"
agent-device click @e2 --platform android --session repro # opens the search field
agent-device fill @e7 "[email protected]" --platform android --session repro
agent-device fill @e7 "[email protected]" --platform android --session repro
agent-device snapshot -i --platform android --session repro
Expected
After the second fill, the field contains:
[email protected]
Actual
The second fill taps the field center, which puts the cursor inside the existing value, and inserts the text there:
[email protected][email protected]
Both fill calls exit successfully. A caller that does not re-read the field carries on with the wrong value.
Cause
On Android, fill first checks whether the device provider supplies its own text entry. Limrun does (provider.text → the Limrun instance client's setText({ x, y }, text)), so agent-device returns after that call. It never reaches the Android path that clears the field (KEYCODE_MOVE_END + KEYCODE_DEL), verifies the result and retries, or uses the test IME.
Calling the Limrun Android instance client directly on the same instance shows that setText never replaces:
| Call | Result |
|---|---|
setText({ x, y }, text) (what agent-device sends) |
inserted at the cursor |
setText({ selector: { focused: true } }, text) |
inserted at the cursor |
setText(undefined, text) |
inserted at the cursor |
setText({ selector: { focused: true } }, '') |
rejected: text is required |
So with the provider-native path there is currently no way to clear the field before typing.
Workaround that worked
Selecting all and deleting with hardware key events before setText, through the same Limrun instance client:
await client.pressKey('a', ['ctrl']);
await client.pressKey('del');
await client.setText(undefined, text);
This left the field holding exactly the requested text in 3 of 3 runs. The select and delete took about 350 ms together; setText took about 1 s.
Suggested fix
For a Limrun Android fill, clear the focused field before the provider-native setText (for example with the Ctrl+A / Delete key events above), then run the existing verification so a mismatch is reported instead of passing silently. Alternatively, route Limrun Android fill through the existing adb clear/verify/retry path or the test IME.
Possibly related (not reproduced)
In one run, a fill into a field that had just taken focus from another field lost its first character (jane… arrived as ane…). I could not reproduce it on a single field in 4 attempts, so it is mentioned only in case it shares the cause.
- 主要言語
- TypeScript
- スター
- 4.9k
- フォーク
- 328
- 平均マージ
- 11時間 49分
- マージ済み PR（30日）
- 585
環境構築
- Dockerfile・Docker Compose ファイルなし
- プルリクエストのテンプレートなし
- コントリビューションガイドを読む
はじめの一歩
- issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
- 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
callstack/agent-device のほかの issue
-
難易度 2/5 1〜3時間 初心者へのやさしさ 74/100
callstack/agent-device#3407 ·
メンテナーはふだん 1 日以内に返信
-
難易度 1/5 1時間未満 初心者へのやさしさ 88/100
callstack/agent-device#3353 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 68/100
callstack/agent-device#1869 ·
メンテナーはふだん 1 日以内に返信
-
Client stop paths treat EPERM as a confirmed exit and drop state for a live companion対応中かも @thymikee が今日担当しました。 オープン
難易度 4/5 3〜5日 初心者へのやさしさ 52/100
callstack/agent-device#3540 ·
メンテナーはふだん 1 日以内に返信
-
enhancement needs-triage
難易度 5/5 1週間以上 初心者へのやさしさ 35/100
callstack/agent-device#3531 ·
メンテナーはふだん 1 日以内に返信
callstack/agent-device の issue をすべて見る
似ている issue
-
難易度 1/5 1〜3時間 初心者へのやさしさ 70/100
siyuan-note/siyuan#20517 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 70/100
NaturalIntelligence/fast-xml-parser#890 · コメント 1 件 ·
メンテナーはふだん 3 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 62/100
zsviczian/obsidian-excalidraw-plugin#3000 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 68/100
pnpm/get.pnpm.io#68 ·
-
難易度 2/5 1〜3時間 初心者へのやさしさ 72/100
nuxt-modules/better-auth#578 · コメント 2 件 ·
メンテナーはふだん 1 日以内に返信