feat(desktop): render approvals inline in conversation

.qwen/e2e-tests/electron-desktop/inline-command-approval-cards.md

@@ -0,0 +1,51 @@
+# Inline Command Approval Cards
+
+- Slice date: 2026-04-26
+- Executable harness: `packages/desktop/scripts/e2e-cdp-smoke.mjs`
+- Command:
+  `cd packages/desktop && npm run e2e:cdp`
+- Result: pass
+- Artifact directory:
+  `.qwen/e2e-tests/electron-desktop/artifacts/2026-04-25T17-47-26-492Z/`
+
+## Scenario
+
+1. Launch the real Electron app with isolated HOME, runtime, user-data, and a
+   fake dirty Git workspace.
+2. Open the fake project through the desktop directory picker path.
+3. Send the first composer prompt without manually creating a thread.
+4. Wait for the fake ACP command permission request.
+5. Assert the request renders as an inline conversation card with the command
+   title, command preview, pending status, and approval/deny actions.
+6. Assert the old detached permission strip is absent and the body does not
+   show a generic `Permission requested` event.
+7. Approve once, assert the pending card resolves, then continue the existing
+   changed-files, review, settings, terminal, and final layout smoke path.
+
+## Assertions
+
+- The inline approval card is inside the chat timeline and stays above the
+  composer without overlap.
+- The card exposes `Approve Once`, `Approve for Thread`, and `Deny` actions.
+- The card includes `Run desktop E2E command`, `printf desktop-e2e`, and a
+  pending status.
+- `.permission-strip` is absent.
+- The conversation body does not contain `Permission requested`.
+- The changed-files summary appears after approval and no approval card remains.
+- Console errors: 0.
+- Failed local network requests: 0.
+
+## Artifacts
+
+- `inline-command-approval.json`
+- `inline-command-approval.png`
+- `conversation-changes-summary.json`
+- `completed-workspace.png`
+- `electron.log`
+- `summary.json`
+
+## Known Uncovered Risk
+
+The harness covers deterministic fake ACP command approval with a string command
+input. It does not yet validate live ACP approvals with structured tool input,
+ask-user free-form answer capture, or long command wrapping at compact widths.