fix: include selected text in follow-up context

[H-Chris233]

User description

Summary

• include the freshly captured selection in every Selection Ask turn instead of only the first turn
• wrap selected text in the existing <selected_text> untrusted-content envelope
• distinguish an empty Linux selection from all supported selection readers being unavailable
• show a localized, non-blocking prompt to install wl-paste, xclip, or xsel when none is installed
• keep QA message rendering compatible with both the new envelope and the legacy message format

Root cause

The QA coordinator captured a new selection for every recording, but message assembly injected it only when the conversation history was empty. Follow-up selections were therefore discarded before the provider request was built on both Windows and Linux.

The Linux reader also collapsed command-not-found, reader failure, and an empty selection into the same None result, so the UI could not tell users that a selection reader needed to be installed.

Validation

• cargo test --lib qa_ -- --nocapture — 22 passed
• cargo test --lib selection::linux_selection::tests -- --nocapture — 2 passed
• npm.cmd run build — passed
• git diff --check — passed

---

PR Type

Bug fix, Enhancement

---

Description

• Include selected text in every QA turn, not just first

• Wrap selected text in XML envelope for security

• Distinguish empty Linux selection from missing tools

• Show localized warning when no Linux selection tool installed

• Guard stale sessions with session_id to prevent race conditions

---

Diagram Walkthrough

flowchart LR
  User["User action"] --> Capture["capture_qa_turn_selection"]
  Capture --> Check{"session_id valid?"}
  Check -->|Yes| Compose["compose_qa_user_content"]
  Compose --> Answer["answer_qa_question_text"]
  Check -->|No| Discard["discard turn"]
  Check -->|No| Warning["show linux_selection_warning?"]

Loading

File Walkthrough

	Relevant files
Enhancement	9 files coordinator.rs Update imports and error monitor                                                  +19/-11  qa_session.rs Rewrite turn logic with session guarding                                  +596/-444 resources.rs Add session-scoped resource management                                      +28/-2    selection.rs Add SelectionCaptureOutcome mock                                                  +12/-0    selection.rs Add capture_with_status and Linux tool detection                  +132/-35 types.rs Add selection_text field to QaChatMessage                                +3/-0      qaMessage.ts Implement message parsing and session event filtering        +60/-0    types.ts Add session_id and selection_warning to types                        +6/-0      QaPanel.tsx Integrate session guarding and selection warning UI            +36/-10
Bug fix	1 files qa.rs Refresh session id on open/close                                                  +25/-13
Tests	3 files llm_gemini.rs Update tests with selection_text                                                  +3/-0      polish.rs Update test with selection_text                                                    +1/-0      qaMessage.test.ts Test helper functions for selection warnings                          +83/-0
I18n	5 files en.ts Add linuxSelectionToolsMissing translation                              +1/-0      ja.ts Add linuxSelectionToolsMissing translation                              +1/-0      ko.ts Add linuxSelectionToolsMissing translation                              +1/-0      zh-CN.ts Add linuxSelectionToolsMissing translation                              +1/-0      zh-TW.ts Add linuxSelectionToolsMissing translation                              +1/-0

---

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit dd8d7ee)

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ No major issues detected

[appergb]

Logical verdict: REQUEST_CHANGES (submitted as a COMMENTED review).

I reviewed exact head 24ad458ce930d2f51fde3df5d4f1ff4fcc60f76d and also synthesized/tested its conflict-free merge with current beta eb099c4f411af546628408e30482d23b2fcc6bf8 (merge tree 74d7321a9fbb952130ef8506c3aa0899875854c8). The Linux reader status split itself is sound: only three NotFound attempts produce ToolsUnavailable; any installed reader that produces no usable text yields NoSelection, and the first text result wins. Voice-to-voice selection refresh is also correct. The following issues block approval:

1. [P1] Typed follow-ups reuse the previous turn's selection. submit_qa_text_question only calls capture_selection_with_status() when state.selection.is_none() (qa_session.rs:105-109). A successful answer reads that value to compose the user message (qa_session.rs:432-440) but finishes by changing only state.phase (qa_session.rs:574); it does not consume or clear the selection. Therefore, after the first typed answer, a second typed submission skips capture and wraps the old selection again. The same stale reuse occurs for voice -> typed: begin_qa_session correctly clears and captures every voice turn (qa_session.rs:598,632-640), but its selection remains after a successful answer, so the next typed turn sees Some and reuses it. open_qa_panel clearing selection once (qa.rs:82-90) and the empty/cancel cleanup at qa_session.rs:1290-1306 do not cover the normal successful multi-turn path. The new qa_followup_includes_new_selection_in_model_context test only calls compose_qa_user_content directly and never constructs the two-turn state transition, so it passes while the advertised typed lifecycle remains broken. Please make selection turn-scoped (recapture/replace it for every typed turn, with the platform focus handling needed to capture the external selection, or consume it after composition as appropriate) and add stateful typed -> typed and voice -> typed regression tests that prove selection B, not A, reaches the second provider request.

2. [P2] The Linux install warning survives desktop close/reopen and is not announced accessibly. QaPanel only updates selectionWarning for an idle event when the field is present (QaPanel.tsx:159-166). Opening the panel emits an idle payload without selection_warning (qa.rs:105-114). On desktop, closing emits chat-panel:closing and hides the persistent webview (lib.rs:2322-2340); qa:dismiss, the only listener that clears this React state (QaPanel.tsx:218-227), is emitted only on Android. Reproduction from this lifecycle: trigger linux_selection_tools_missing, close with X/Esc/hotkey, then reopen; the old warning remains even though the new panel state has not reported missing tools (and can remain after tools are installed). Please clear the warning deterministically on every reset/open/desktop close, or include an explicit selection_warning: null in reset events, and add a lifecycle/component regression test. The dynamically inserted warning is also a plain div (QaPanel.tsx:397-400) with no role="status"/aria-live, so screen-reader users are not notified; please give the non-blocking update appropriate live-region semantics.

3. [P2] Sanitized envelope text is double-escaped in the selection quote. The backend deliberately rewrites an injected boundary such as </selected_text> to </selected_text> before storing the user message (qa_session.rs:11-20; polish.rs:1683-1709). splitSelectionUser returns that encoded capture verbatim (QaPanel.tsx:567-573), and React then escapes its ampersand. For A &lt;/selected_text> B & C < D, the rendered HTML is A </selected_text> B & C < D, so the user sees literal </selected_text> instead of the selected </selected_text>. Raw < and & remain safely text-escaped by React, and the legacy # 选区原文 form parses safely, but the new canonical format loses display fidelity for exactly the boundary-like content the sanitizer changes. There is no frontend parser/rendering test for canonical, escaped, or legacy input. Please preserve a display-safe copy or use an unambiguous reversible encoding/decoding contract (a naive global < replacement would corrupt an originally literal <) and add those tests.

Validation evidence on the current-beta synthetic merge:

• Conflict-free git merge-tree --write-tree; git diff --check passed.
• Full Rust library suite: 707 passed, 0 failed; focused qa_: 21 passed; Linux reader tests: 2 passed.
• npm run build passed (Android IPC boundary, TypeScript, Vite); QA markdown test and all eight package contract scripts passed, including current beta's macOS speech-usage-description contract.
• npm audit: 0 vulnerabilities. cargo audit: exit 0 with 18 allowed warnings. Gitleaks on the exact PR commit: no leaks.
• GitNexus compare reports 11 files / 45 changed symbols, 9 affected execution flows, high risk, which matches the lifecycle-sensitive scope above.
• All five remote checks on exact head are green: Android cargo check, Linux checks, macOS checks, Windows checks, and pr_agent_job. Those jobs ran before beta advanced, so the local current-beta merge-tree run above supplies the updated regression evidence.

Please keep the PR at exact head or call out a new head for re-review after addressing the three lifecycle/display gaps.

[H-Chris233]

收到，新改进将在不久推出

[github-actions]

Persistent review updated to latest commit f718515

[github-actions]

Persistent review updated to latest commit 07d10a2

[github-actions]

Persistent review updated to latest commit 9e13f79

[github-actions]

Persistent review updated to latest commit 3c9af35

[github-actions]

Persistent review updated to latest commit a121919

[github-actions]

Persistent review updated to latest commit 3b4da1b

[github-actions]

Persistent review updated to latest commit 3e02a6f

[github-actions]

Persistent review updated to latest commit dd8d7ee

[github-actions]

Persistent review updated to latest commit dd8d7ee