vrr/qwen-code
2
0
Fork 0
mirror of https://github.com/QwenLM/qwen-code.git synced 2026-04-29 04:00:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 28
qwen-code/design/qwen-code-electron-desktop-implementation-plan.md

50 KiB
Raw Blame History

Qwen Code Electron Desktop Implementation Plan

This plan tracks the incremental MVP implementation for the Electron desktop client described in docs/design/qwen-code-electron-desktop/qwen-code-electron-desktop-architecture.md. The architecture document remains the source of truth; this file records execution order, verification, decisions, and remaining work.

Ground Rules

  • Use Electron only; do not introduce Tauri.
  • Keep Electron main thin: windows, native IPC, local server lifecycle, and ACP process lifecycle.
  • Reuse Qwen Code ACP, core configuration/auth/session/permission behavior, and shared web UI surfaces where practical.
  • Renderer must use nodeIntegration: false, context isolation, and a preload whitelist.
  • The local server must bind only 127.0.0.1, use a random token, and reject unauthorized requests.
  • Every completed slice must leave targeted verification and a conventional commit.

Codex Alignment Progress

Completed Slice: Dense Assistant File Reference Overflow

Status: completed in iteration 12.

Goal: harden assistant prose file-reference rendering for realistic, dense responses with repeated references, line/column suffixes, uncommon source file extensions, and more references than can comfortably fit in the message card.

User-visible value: assistant responses stay compact and readable in the conversation-first workbench while still exposing useful file chips for opening referenced files. Repeated paths do not add visual noise, and overflow is explicit instead of silently dropping references.

Expected files:

  • packages/desktop/src/renderer/components/layout/ChatThread.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/src/renderer/styles.css
  • packages/desktop/src/main/acp/createE2eAcpClient.ts
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/assistant-file-reference-overflow.md
  • design/qwen-code-electron-desktop-implementation-plan.md

Acceptance criteria:

  • Assistant prose deduplicates repeated file references while preserving the first visible label.
  • References with :line:column suffixes open the file path without the line suffix and keep the visible line/column label.
  • Common desktop/code references such as .mdx, .mts, .cts, .vue, .svelte, .astro, Dockerfile, Makefile, .env, .gitignore, and .npmrc can render as chips when they appear in assistant prose.
  • More than six references render the first six chips plus a compact overflow indicator with an accessible label.
  • Long chips wrap/truncate within the assistant message at normal and compact widths without horizontal page overflow or composer overlap.

Verification:

  • Unit/component test command: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, open the fake Git project, send a prompt, approve the fake command request, wait for the dense assistant response, assert deduped chips, line/column chips, overflow count, and contained chip geometry, then continue the existing copy/retry/review/settings/terminal smoke path.
  • E2E assertions: assistant file chips include README.md:1, packages/desktop/src/renderer/App.tsx:12:5, .env.example, Dockerfile, and an overflow indicator; duplicate README.md:1 references render once; every chip stays inside the assistant message/timeline; document scroll width does not exceed the viewport; console errors/failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, dense assistant reference JSON, assistant action JSON, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for prototype-constrained compact chip density and overflow treatment; electron-desktop-dev for renderer changes and real Electron CDP verification; brainstorming applied by choosing the smallest continuation of the recorded rich-conversation backlog from repo artifacts and home.jpg without pausing the autonomous loop.

Notes and decisions:

  • The prototype shows file/change context inline with the conversation, so this slice keeps file chips inside assistant messages rather than moving dense references into a separate drawer.
  • Overflow uses a quiet text chip so the message remains readable and does not become a file browser.
  • The fake ACP response includes deterministic dense references so the CDP harness can verify real Electron layout and dedupe behavior.
  • The first focused component test exposed a line/column stripping bug where path.ts:12:5 opened path.ts:12. The final implementation strips the full :line:column suffix for open-file callbacks while preserving the visible chip label.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx passed with 10 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed after launch through real Electron over CDP.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T18-17-10-902Z/.

Next work:

  • Add a compact-viewport CDP pass or Browser bounds control for the dense conversation state so long assistant/file chips are also asserted near the lower supported desktop width.
  • Continue rich conversation fidelity by adding clearer assistant action feedback for clipboard/open-file failures and by keeping multiple assistant messages dense at compact widths.

Completed Slice: Assistant Message Actions and File Reference Chips

Status: completed in iteration 11.

Goal: add compact assistant message actions and clickable file-reference chips inside the conversation timeline.

User-visible value: after an assistant response, users can copy the response, reuse the last prompt, jump into changed-file review, and open referenced files without leaving the workbench or reading protocol/debug output.

Expected files:

  • packages/desktop/src/renderer/App.tsx
  • packages/desktop/src/renderer/components/layout/ChatThread.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/src/renderer/styles.css
  • packages/desktop/src/main/acp/createE2eAcpClient.ts
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/assistant-message-actions.md
  • design/qwen-code-electron-desktop-implementation-plan.md

Acceptance criteria:

  • Assistant messages render a compact action row with Copy, Retry last prompt, and Open Changes when changed files exist.
  • Retry last prompt is safe: it restores the previous user prompt into the composer instead of auto-sending a new agent request.
  • File references in assistant prose, such as README.md:1, render as compact chips with an accessible open action.
  • Copy and open actions use existing desktop-safe preload/browser APIs and do not expose ACP/session IDs in the main timeline.
  • The message card stays within the conversation column and does not overlap the composer in real Electron.

Verification:

  • Unit/component test command: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, open the fake Git project, send a prompt, approve the fake command request, wait for the assistant response, assert the assistant action row and file chip, copy the response, retry the last prompt into the composer, clear the retry draft, then continue the existing review/settings/ terminal smoke path.
  • E2E assertions: assistant message action row is present, the file chip shows README.md:1, Copy produces visible feedback, Retry restores the original prompt without auto-sending, Open Changes remains contextual, assistant geometry stays inside the timeline above the composer, and console errors/ failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, assistant action JSON, retry composer JSON, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for prototype-constrained compact inline actions and file chip density; electron-desktop-dev for renderer changes and real Electron CDP verification; brainstorming applied by choosing the narrow conversation-first option from the repo plan and immutable prototype without pausing the autonomous Ralph loop.

Notes and decisions:

  • The prototype shows response actions and changed-file controls in the reading flow, so the action row stays under assistant messages instead of becoming a toolbar or drawer.
  • Retry is intentionally non-destructive and does not auto-send; it drafts the last user prompt in the composer so users can inspect or edit before sending.
  • File reference chips reuse the existing project-relative open-file path and remain bounded so long paths cannot stretch the timeline.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx passed with 9 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed after launch through real Electron over CDP.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T18-10-35-606Z/.

Next work:

  • Continue rich conversation primitives by adding clearer assistant feedback states for copy/retry failures and by supporting multiple dense assistant messages at compact viewport widths.
  • Add a follow-up fake ACP scenario with longer assistant prose and several repeated file references to harden chip extraction, dedupe, and overflow.

Completed Slice: Rich Tool-Call Activity Cards

Status: completed in iteration 10.

Goal: make completed and in-progress tool calls read as useful task activity inside the conversation instead of a sparse tool row.

User-visible value: users can see what the agent did, what command/input was used, which files were referenced, and whether the tool completed or failed without reading ACP IDs or opening diagnostics.

Expected files:

  • packages/desktop/src/main/acp/createE2eAcpClient.ts
  • packages/desktop/src/renderer/components/layout/ChatThread.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/src/renderer/stores/chatStore.test.ts
  • packages/desktop/src/renderer/styles.css
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/rich-tool-call-activity-cards.md
  • design/qwen-code-electron-desktop-implementation-plan.md

Acceptance criteria:

  • Tool calls render as compact inline conversation activity cards with kind, title, status, and stable data-testid hooks.
  • Tool cards show a bounded command/input preview when safe user-facing input is present.
  • Completed or failed tool cards show a bounded output/result summary without exposing request/session IDs.
  • File locations render as compact chips with path and optional line number.
  • The previous generic .chat-tool row no longer appears for tool activity.
  • Cards stay within the timeline and do not overlap the composer in real Electron.

Verification:

  • Unit/component test command: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/stores/chatStore.test.ts src/renderer/components/layout/WorkspacePage.test.tsx
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, open the fake Git project, send from the composer, approve the fake command request, then assert the resolved tool activity card includes command title, status, command preview, output summary, and file chips before continuing the existing review/settings/terminal smoke path.
  • E2E assertions: activity card is present after approval, uses compact geometry inside the chat timeline, contains README.md:1, does not render the raw tool call ID or session ID, no legacy .chat-tool node remains, and console errors/failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, rich tool-call JSON, conversation summary JSON, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for prototype-constrained compact activity-card hierarchy and file chip density; electron-desktop-dev for renderer changes and real Electron CDP verification; brainstorming applied by deriving the slice from the repo plan and immutable prototype instead of asking ordinary product questions during the autonomous loop.

Notes and decisions:

  • The prototype keeps agent activity in the reading flow, so this slice replaces the generic tool row with an inline card rather than adding another panel.
  • The card intentionally surfaces title/kind/status, bounded input/output, and file locations only. ACP request IDs, session IDs, and transport details stay out of the main conversation.
  • The fake ACP path will emit deterministic location/output data so the CDP harness can assert a real user-visible resolved tool card.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/stores/chatStore.test.ts src/renderer/components/layout/WorkspacePage.test.tsx passed with 13 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed after launch through real Electron over CDP.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T17-57-31-788Z/.

Next work:

  • Continue rich conversation primitives by adding assistant message action rows for copy/retry/open changed files and by turning file references in assistant prose into compact open/reveal chips.
  • Tighten tool-card density at compact viewport widths after adding a second fake ACP scenario with multiple file references and longer command output.

Completed Slice: Inline Command Approval Cards

Status: completed in iteration 9.

Goal: make command approvals and ask-user prompts part of the conversation timeline instead of a detached permission strip or protocol-like event row.

User-visible value: users see what command/action needs attention in the same reading flow as the agent plan, tool activity, and changed-files summary. The main conversation can answer "what needs me now?" without exposing ACP request plumbing.

Expected files:

  • packages/desktop/src/renderer/components/layout/ChatThread.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/src/renderer/stores/chatStore.ts
  • packages/desktop/src/renderer/stores/chatStore.test.ts
  • packages/desktop/src/renderer/styles.css
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/inline-command-approval-cards.md
  • design/qwen-code-electron-desktop-implementation-plan.md

Acceptance criteria:

  • Pending command permissions render as compact inline conversation cards with command/tool title, optional command input, status, and approval/deny actions.
  • Pending ask-user questions render inline with question text, options, and Cancel/Submit actions.
  • The old permission strip is no longer rendered as a separate surface between the timeline and composer.
  • Permission and ask-user server messages no longer append generic Permission requested or Question requested event rows to the timeline.
  • Approval controls keep stable accessible labels and continue to send the same permission response.
  • The composer remains docked and usable while a pending approval card is visible; changed-files summary still appears after the request resolves.

Verification:

  • Unit/component test command: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/stores/chatStore.test.ts src/renderer/components/layout/WorkspacePage.test.tsx
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, open the fake Git project, send from the composer, assert the pending command approval appears as an inline conversation card with the fake command title/input and no separate permission strip, approve it, assert the card resolves away and the changed-files summary appears, then continue the existing review, settings, and terminal smoke path.
  • E2E assertions: inline approval card is present before approval, has compact geometry within the chat timeline, exposes approval/deny actions, does not render protocol request events, and console errors/failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, inline approval JSON, conversation summary JSON, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for prototype-constrained inline card density, action hierarchy, and conversation-first placement; electron-desktop-dev for renderer changes and real Electron CDP verification.

Notes and decisions:

  • The prototype keeps approvals and task state in the reading flow, so this slice removes the separate permission strip instead of duplicating the same action in two places.
  • The backing permission response contract remains unchanged; only renderer placement and noise filtering change.
  • The inline card intentionally shows only the tool title/kind/status and a string or command preview from tool input; request IDs and session IDs stay out of the main conversation.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/stores/chatStore.test.ts src/renderer/components/layout/WorkspacePage.test.tsx passed with 11 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed after launch through real Electron over CDP.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T17-47-26-492Z/.

Next work:

  • Continue rich conversation primitives by improving tool-call cards with file-reference chips, copy/retry/open actions, and clearer completed/failed command output summaries.
  • Run another prototype fidelity pass on message density and assistant action rows now that approvals, changed files, terminal, review, and settings have all moved into supporting surfaces.

Completed Slice: Settings Information Architecture

Status: completed in iteration 8.

Goal: make Settings read like product settings instead of a runtime debug panel by grouping account, model provider, permission, tools, terminal, appearance, and diagnostics controls.

User-visible value: users can find model/API key and permission controls without seeing server URLs, Node versions, ACP state, active session IDs, or other diagnostics in the default settings view. Advanced diagnostics remain available when needed.

Expected files:

  • packages/desktop/src/renderer/components/layout/SettingsPage.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/src/renderer/styles.css
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/settings-information-architecture.md
  • design/qwen-code-electron-desktop-implementation-plan.md

Acceptance criteria:

  • Settings defaults to product sections: Account, Model Providers, Permissions, Tools & MCP, Terminal, Appearance, and Advanced Diagnostics.
  • The default settings view does not visibly expose server URL, Node version, ACP status, health milliseconds, settings path, or active session IDs.
  • Model, Base URL, API key, OAuth, Save, and permission-mode controls remain reachable from the settings page.
  • API key state is shown as configured/missing without rendering saved secret values in the DOM.
  • Advanced Diagnostics can be opened explicitly and then shows runtime, session, and config diagnostic fields.

Verification:

  • Unit/component test command: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx src/renderer/stores/settingsStore.test.ts
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, open the fake Git project, complete the existing composer, review, and commit path, open Settings, assert product sections are visible while diagnostics are hidden, edit model/Base URL/API key, save, assert the saved model appears without secret leakage, open Advanced Diagnostics, and assert runtime diagnostics are available only there.
  • E2E assertions: settings replaces chat/review/terminal, default settings text excludes server URL, Node, ACP, active session ID, health ms, settings path, and the fake API key; Advanced Diagnostics renders the runtime diagnostics after an explicit click; console errors and failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, settings layout JSON, advanced diagnostics JSON, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for prototype-constrained product settings hierarchy and lower-noise surfaces; electron-desktop-dev for renderer changes and real Electron CDP verification.

Notes and decisions:

  • The prototype treats Settings as supporting product chrome rather than a main debug dashboard, so diagnostics move behind an explicit Advanced action.
  • This slice does not change settings persistence contracts; it reorganizes the renderer around the existing server/settings store APIs and keeps secrets out of rendered text.
  • Settings remains a full workbench page for now, consistent with the previous verified behavior; this slice focuses on information architecture inside the page rather than converting Settings to a modal or drawer.
  • The first CDP run reached Advanced Diagnostics but failed on a harness-only case-sensitive label assertion because diagnostic labels are rendered uppercase by CSS. The harness now asserts diagnostics case-insensitively.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx src/renderer/stores/settingsStore.test.ts passed with 8 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed after launch through real Electron over CDP.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T17-40-11-622Z/.

Next work:

  • Continue rich conversation primitives by rendering command approvals and tool activity inline in the timeline rather than only in the permission strip.
  • Tighten settings density and responsive behavior further after the next conversation-first fidelity pass.

Completed Slice: Conversation Changed-Files Summary and Protocol Noise Cleanup

Status: completed in iteration 7.

Goal: make the main conversation timeline feel like a product task flow by hiding ACP/session protocol noise and surfacing Git changes inline.

User-visible value: users should not see internal session IDs or protocol stop reasons in the main reading flow, and they can discover changed files from the conversation itself instead of starting from the topbar.

Expected files:

  • packages/desktop/src/renderer/components/layout/ChatThread.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/src/renderer/stores/chatStore.ts
  • packages/desktop/src/renderer/stores/chatStore.test.ts
  • packages/desktop/src/renderer/styles.css
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/conversation-changes-summary.md
  • design/qwen-code-electron-desktop-implementation-plan.md

Acceptance criteria:

  • The chat timeline no longer renders Connected to <session id> or Turn complete: <stop reason> event rows.
  • Connection state remains available in the compact header/topbar rather than as protocol prose in the timeline.
  • When the active project has Git changes, the conversation shows a compact changed-files summary with file names, staged/unstaged/untracked state, and addition/deletion totals.
  • The inline summary opens the review drawer while keeping the conversation mounted.
  • The summary hides itself when there are no changed files.

Verification:

  • Unit/component test commands: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/stores/chatStore.test.ts src/renderer/components/layout/WorkspacePage.test.tsx
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, open the fake Git project, create a composer-first thread, approve the fake command, assert protocol IDs and stop reasons are absent from the body text, assert the conversation changed-files summary is present, open review from that summary, then continue through discard cancel, stage, commit, settings, and terminal paths.
  • E2E assertions: the body text does not contain Connected to session-e2e, session-e2e-1, or Turn complete; the inline summary reports the fake dirty files and opens the review drawer; console errors and failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, conversation summary JSON, review layout JSON, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for prototype-constrained inline cards and conversation density; electron-desktop-dev for renderer changes and real Electron CDP verification.

Notes and decisions:

  • The renderer still tracks connection state in ChatState.connection and the compact header/topbar, but connected and message_complete protocol messages no longer create timeline rows.
  • The changed-files summary is derived from the active project Git diff instead of fake ACP payloads, so it appears whenever the review drawer would have meaningful content and disappears after commit/clean states.
  • The inline summary opens the existing review drawer rather than introducing a separate review surface, keeping the first viewport conversation-first and consistent with home.jpg.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/stores/chatStore.test.ts src/renderer/components/layout/WorkspacePage.test.tsx passed with 9 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed after launch through real Electron over CDP.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T17-28-04-569Z/.

Next work:

  • Continue rich conversation primitives by rendering command approvals and tool activity as inline cards instead of relying mostly on the permission strip.
  • Improve settings information architecture so runtime diagnostics move under Advanced and model/API key controls are reachable as product settings.

Completed Slice: Review Safety Terminology and Discard Confirmation

Status: completed in iteration 6.

Goal: make the review drawer use Git-safe language and require explicit confirmation before destructive discard operations.

User-visible value: users can review changed files without seeing ambiguous Accept/Revert controls, and high-risk discard actions cannot be triggered with one accidental click. This keeps review as a compact supporting surface while preserving the conversation-first workbench shown in home.jpg.

Expected files:

  • packages/desktop/src/renderer/components/layout/ReviewPanel.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/src/renderer/styles.css
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/review-safety.md
  • design/qwen-code-electron-desktop-implementation-plan.md

Acceptance criteria:

  • Review drawer controls use Stage/Discard terminology instead of Accept/Revert.
  • Staged hunks/files show a staged state, and staging controls are disabled when already staged.
  • Discard all/file/hunk actions open a confirmation UI that names the target, explains the local-change risk, and supports Cancel and confirmed discard.
  • Canceling a discard leaves the Git worktree and review counts unchanged.
  • Committing staged changes remains available from the same compact drawer.

Verification:

  • Unit/component test command: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, open the fake Git project, create a composer-first thread, approve the fake command, open Changes, verify Stage/Discard language, initiate Discard All, cancel it, assert the workspace still has changes, stage all changes, commit, and continue through settings and terminal paths.
  • E2E assertions: Accept/Revert labels are absent from the main review drawer; discard confirmation appears and can be canceled; modified/untracked counts remain after cancel; staged counts update after Stage All; console errors and failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, review layout JSON, discard confirmation JSON, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for prototype-constrained compact review actions, danger hierarchy, and confirmation wording; electron-desktop-dev for renderer changes and real Electron CDP verification.

Notes and decisions:

  • The underlying server endpoint remains named revert for now; this slice changes product-facing language to Discard while keeping the existing reviewed backend contract.
  • Confirmation stays inside the review drawer rather than using a native dialog so the CDP harness can assert the user path deterministically and the first viewport remains desktop-native.
  • The first CDP run exposed a harness-only assertion mismatch: review counts render as definition rows while the topbar renders the combined dirty count. The harness now asserts both surfaces through their actual UI shapes.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx passed with 6 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed after launch through real Electron over CDP.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T17-18-14-754Z/.

Next work:

  • Continue conversation timeline fidelity by hiding session/protocol IDs such as Connected to session-e2e-1 from the main user flow.
  • Add inline changed-file summary cards in the conversation that open the review drawer without forcing users to start from the topbar.

Completed Slice: Terminal Attach-to-Composer Workflow

Status: completed in iteration 5.

Goal: change terminal output follow-up from an immediate Send to AI action into an explicit attach-to-composer flow, so users can review and edit command output before deciding whether to send it to the agent.

User-visible value: terminal output becomes contextual material in the task composer rather than a hidden second send path that can unexpectedly trigger a new agent turn. This keeps the conversation-first workbench aligned with home.jpg while preserving the terminal as a supporting tool.

Expected files:

  • packages/desktop/src/renderer/App.tsx
  • packages/desktop/src/renderer/api/websocket.ts
  • packages/desktop/src/renderer/components/layout/WorkspacePage.tsx
  • packages/desktop/src/renderer/components/layout/TerminalDrawer.tsx
  • packages/desktop/src/renderer/components/layout/SidebarIcons.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/terminal-drawer.md
  • design/qwen-code-electron-desktop-implementation-plan.md

Acceptance criteria:

  • The expanded Terminal action is labeled as attaching output to the composer, not sending directly to AI.
  • Attaching terminal output appends a bounded terminal transcript to the existing composer text and shows a clear success notice.
  • The attach action works whenever terminal output exists, including before a thread is selected, and does not require or write to the session WebSocket.
  • The user must still click Send from the composer before a new agent turn is created.
  • Copy, clear, kill, run command, stdin, expand, and collapse behavior is unchanged.

Verification:

  • Unit/component test command: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, open the fake Git project, create a composer-first thread, approve the fake command, review and commit changes, expand Terminal, run stdout and stdin commands, attach the resulting output to the composer, assert no fake ACP follow-up happens until Send is clicked, then send the composer text and approve the fake command request.
  • E2E assertions: attach button is present and Send to AI is absent; composer contains the terminal transcript after attach; terminal notice confirms the attachment; the output stays editable in the composer; console errors and failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, terminal layout JSON, composer attach JSON, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for prototype-constrained terminal action wording and compact composer-centric hierarchy; electron-desktop-dev for renderer changes and real Electron CDP verification.

Notes and decisions:

  • The prototype keeps the composer as the task control center, so terminal output should land there for user review rather than bypassing it.
  • This slice intentionally preserves the transcript formatting and bounding logic from the existing send path, but changes the destination from WebSocket send to composer draft text.
  • The WebSocket helper no longer needs a separate terminal-output send method because the final send is the same explicit user-message path as any other composer submit.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx passed with 5 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed after launch through real Electron over CDP.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T17-08-17-022Z/.

Next work:

  • Improve review safety by replacing Accept/Revert terminology with Stage/Unstage/Discard and confirming destructive discard paths.
  • Continue prototype fidelity work in the conversation timeline by hiding protocol/session noise and adding inline changed-file summaries.

Completed Slice: Collapsed Terminal Status Strip Alignment

Status: completed in iteration 4.

Goal: collapse the terminal into a compact bottom status strip by default, so the first viewport keeps the conversation as the dominant surface while still making terminal access discoverable.

User-visible value: users see the active project, conversation, composer, and Git/review controls without the terminal permanently consuming a large block of height. Terminal commands remain available through an explicit expand/collapse control.

Expected files:

  • packages/desktop/src/renderer/components/layout/WorkspacePage.tsx
  • packages/desktop/src/renderer/components/layout/TerminalDrawer.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/src/renderer/styles.css
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/terminal-drawer.md
  • design/qwen-code-electron-desktop-implementation-plan.md

Acceptance criteria:

  • The default workbench renders a compact terminal strip rather than a full terminal drawer.
  • The strip shows project/status context and an accessible Expand Terminal control.
  • Expanding the terminal reveals command, stdin, output, copy, send, clear, and kill controls without replacing the conversation.
  • Collapsing the terminal after use hides the large output region and restores first-viewport conversation dominance.
  • Settings still replaces chat/review/terminal as before.
  • Existing terminal run, stdin, copy, kill, clear, and send-to-AI behavior keeps working.

Verification:

  • Unit/component test command: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, assert the first viewport terminal strip is collapsed, open the fake Git project, send from the composer, approve the fake command, review and commit changes, open settings, return to conversation, expand Terminal, run commands including stdin, send output to the fake ACP session, collapse Terminal again, and assert the final layout returns to a compact strip.
  • E2E assertions: initial and completed terminal heights stay compact; expanded terminal height stays supporting and docked; conversation remains wider and taller than terminal by default; console errors and failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, collapsed/expanded layout JSON, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for prototype-constrained bottom strip hierarchy, compact controls, and conversation-first density; electron-desktop-dev for renderer changes and real Electron CDP verification.

Notes and decisions:

  • This slice follows home.jpg over the older always-visible terminal panel: terminal remains a supporting workbench tool, not a permanent third major viewport region.
  • The terminal strip remains in the workbench rather than moving into settings or review, because running commands in the active project is part of the coding-agent loop.
  • The existing Send to AI behavior is preserved for this slice; changing that to attach output to the composer is still the next terminal workflow refinement.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx passed with 4 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed after launch through real Electron over CDP.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T17-00-08-461Z/.

Next work:

  • Rename terminal Send to AI into an attach-to-composer flow so command output does not unexpectedly trigger another agent turn.
  • Continue review safety work by replacing Accept/Revert terminology with Stage/Unstage/Discard and adding confirmations for destructive discard paths.

Completed Slice: Review Drawer and Compact Topbar Alignment

Status: completed in iteration 3.

Goal: make review a supporting drawer that opens beside the conversation, and replace the heavy topbar tabs with compact icon-led workbench actions.

User-visible value: the first viewport keeps the conversation as the main workspace while still exposing changed files, settings, Git refresh, and status from a slim topbar that better matches home.jpg.

Expected files:

  • packages/desktop/src/renderer/components/layout/WorkspacePage.tsx
  • packages/desktop/src/renderer/components/layout/TopBar.tsx
  • packages/desktop/src/renderer/components/layout/ReviewPanel.tsx
  • packages/desktop/src/renderer/components/layout/SidebarIcons.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/src/renderer/styles.css
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/review-drawer-topbar.md

Acceptance criteria:

  • Opening Changes renders ChatThread and ReviewPanel together; review no longer replaces the conversation.
  • The default first viewport has no review drawer, and the conversation spans the workbench.
  • Topbar action controls are compact icon buttons with accessible labels and tooltips; the previous Chat/Changes/Settings segmented text tabs are removed.
  • The topbar title remains the active thread/project identity instead of changing to Changes when review opens.
  • Settings still opens as a full workbench page and hides the terminal.
  • Existing review actions, comments, staging, and commit workflow keep working.

Verification:

  • Unit/component test command: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, open the fake Git project, send from the project composer, approve the fake command, open Changes from the compact topbar action, review and comment on the README diff while chat remains mounted, stage all changes, commit, return to chat, open settings, and run terminal paths.
  • E2E assertions: default layout has no review drawer; opening Changes creates a drawer without unmounting chat; drawer width stays supporting rather than dominant; topbar has compact action buttons; console errors and failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, layout JSON, DOM text, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for prototype-constrained topbar density and drawer hierarchy; electron-desktop-dev for renderer changes and real Electron CDP verification.

Notes and decisions:

  • This slice deliberately keeps Settings as a full page because that behavior was already implemented and verified; only review moves into the supporting drawer pattern.
  • The review drawer remains closed by default to preserve the first viewport emphasis from home.jpg; Git dirty count and the Changes action are the visible entry points.
  • frontend-design guidance is applied with the project prompt constraint that the prototype wins: compact utility controls, restrained borders, and no new decorative art direction.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx passed with 4 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed after launch through real Electron over CDP.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T16-51-33-353Z/.

Next work:

  • Improve review terminology and safety by replacing Accept/Revert with Stage/Unstage/Discard language and adding confirmations for destructive discard paths.
  • Collapse the terminal into a status strip by default so the first viewport gets closer to home.jpg.

Completed Slice: Composer-First Thread Creation Alignment

Status: completed in iteration 2.

Goal: let a user open a project and type immediately, without first learning that they must create or select a session.

User-visible value: the default path becomes Open project -> type request -> agent works; the composer explains the active project context and creates the backing desktop session on first send.

Expected files:

  • packages/desktop/src/renderer/App.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.tsx
  • packages/desktop/src/renderer/components/layout/ChatThread.tsx
  • packages/desktop/src/renderer/components/layout/WorkspacePage.test.tsx
  • packages/desktop/src/renderer/styles.css
  • packages/desktop/scripts/e2e-cdp-smoke.mjs
  • .qwen/e2e-tests/electron-desktop/composer-first-thread-creation.md

Acceptance criteria:

  • Composer is enabled whenever a project is active, even when no session is selected.
  • With no project, composer remains disabled and gives a clear disabled reason.
  • First send from a project with no selected session creates a desktop session, sends the message, clears the composer, and publishes the created thread.
  • Existing explicit New Thread behavior continues to work.
  • The composer visibly carries compact project/branch, permission, and model context so it reads as the task control center rather than a plain textarea.
  • Enter send and Shift+Enter newline behavior are preserved.

Verification:

  • Unit/component test command: cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx
  • Build/typecheck/lint commands: cd packages/desktop && npm run typecheck && npm run lint && npm run build
  • Real Electron harness: cd packages/desktop && npm run e2e:cdp
  • Harness path: packages/desktop/scripts/e2e-cdp-smoke.mjs
  • E2E scenario steps: launch real Electron with isolated HOME/runtime/user-data and fake ACP, open the fake Git project, type a prompt into the project-scoped composer without clicking New Thread, send it, approve the fake command request, and assert the created thread/message/response appear.
  • E2E assertions: first viewport landmarks stay present; composer is enabled after project open; no New Thread click is required; fake ACP response is received; console errors and failed local requests are absent.
  • Diagnostic artifacts: CDP screenshots, layout JSON, DOM text, Electron log, summary JSON under .qwen/e2e-tests/electron-desktop/artifacts/.
  • Required skills applied: frontend-design for composer layout/control hierarchy with the prototype as the strict visual contract; electron-desktop-dev for renderer changes and real Electron CDP verification.

Notes and decisions:

  • The prototype wins over earlier tab/dashboard guidance. This slice keeps the conversation as the default surface and upgrades the bottom composer without opening review, terminal, or settings by default.
  • Model and permission controls are compact context controls in the composer. They use existing session runtime state when available and safe fallback labels before a session exists; changing values still requires a live session until the server API supports project-level defaults.
  • Implementation changed first-send behavior so any active project with no active session creates a session on submit. The explicit New Thread button still creates a draft thread for users who want to start intentionally from the sidebar.
  • CDP smoke now sends the first prompt immediately after opening the fake project and before clicking Changes, proving the New Thread click is no longer required.

Verification results:

  • cd packages/desktop && SHELL=/bin/bash npx vitest run src/renderer/components/layout/WorkspacePage.test.tsx passed with 4 tests.
  • cd packages/desktop && npm run typecheck passed.
  • cd packages/desktop && npm run lint passed.
  • cd packages/desktop && npm run build passed.
  • cd packages/desktop && npm run e2e:cdp passed.
  • Passing artifacts: .qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T16-41-09-752Z/.

Next work:

  • Continue prototype fidelity by reducing topbar tab weight and moving review access toward compact icon/drawer behavior.
  • Follow-up model configuration work should make composer model/permission controls editable before a session exists by persisting project-level defaults, rather than only reflecting live session runtime state.