fix(desktop): harden Notch chat and PTT reliability

[Git-on-my-level]

Summary

• preserve the user-resized Notch chat frame when macOS changes Spaces
• keep streaming output from reclaiming the viewport after a reader scrolls away from the live edge
• restore the canonical frame for non-chat Notch surfaces after a Space switch
• retain completed and failed tool rows as durable chat progress in compact, collapsed-by-default groups whose step count updates while work continues, while keeping the structured agent card as the sole spawned-agent entrypoint
• use a shared thin composer shell and transcript fade in regular and Notch chat, reduce the regular-chat header to one compact control row, and add a persistent Home action there
• pass a deduplicated, compact skill catalog to chat, keep CLAUDE.md reference-only, and search overflow skills only when a specialized workflow is relevant
• give active tools a 90-second no-progress watchdog that preserves the tool-stall terminal reason; use the 60-second generic watchdog only for a silent bridge with no active tool
• correlate the desktop, agent, and chat route with an opaque request ID, and retry terminal journal projection once when needed
• keep Push-to-Talk on its native voice path when screen capture is temporarily unavailable immediately after Screen Recording is granted
• prevent long named-bundle titles from clipping the Screen Recording drag helper, and coalesce idle PTT context refreshes so streaming chat does not churn the warm voice socket

Product invariants

• INV-AGENT-*
• INV-AUTH-1
• INV-CHAT-1
• INV-VOICE-1
• INV-INT-1

Verification

• desktop/macos/scripts/agent-logic-harness.sh — passed (Swift lifecycle, agent runtime, and pi extension suites)
• xcrun swift test --package-path Desktop --filter ChatSkillCatalogTests — passed (2 tests)
• xcrun swift test --package-path Desktop --filter AgentPillLifecycleTests/testFloatingAgentToolCallsUseCompactOneLinePresentation — passed
• npm run build and node --experimental-strip-types scripts/generate-tool-surfaces.mjs --check in desktop/macos/agent — passed
• npx vitest run tests/node-tools.test.ts tests/omi-tool-manifest.test.ts — passed (20 tests)
• npx --yes tsx --test index.test.ts in desktop/macos/pi-mono-extension — passed
• make preflight and the bounded pre-push gate — passed
• git diff --check — passed
• rebuilt /Applications/omi-tool-stall-reliability.app from ae2e79f; health check confirmed development backend and agent protocol 2, and automation confirmed Chat's Home button returns to the dashboard
• xcrun swift test --package-path Desktop --filter ScreenRecordingPermissionPolicyTests — passed (15 tests)
• xcrun swift test --package-path Desktop --filter RealtimeHubSessionHandoffPolicyTests — passed (12 tests)
• fast-rebuilt /Applications/omi-tool-stall-reliability.app from f9d5a81; health check confirmed the development backend and runtime protocol 2

Failure-Class: none

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8fa35e6f41

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".