feat(workbench): converge v2 agent runtime and activity UX

[Qiyuanqiii]

Summary

This draft converges the current Traverse Board v2 work into one reviewable handoff:

• rewrites the desktop workbench in web/src/v2 with the project liquid-glass treatment
• adds Provider presets/custom Provider configuration, model switching, credential references, and Responses API runtime support
• connects Thread/Run recovery, permission/model epoch transitions, streaming model events, web evidence, MCP, and Full CDP lifecycle/actions
• exposes safe, typed activity details for commands, MCP, search/fetch, file reads/edits, verification, and browser actions
• persists real Agent attribution and replaces public generic facts/raw payloads with thread_activity_detail.v2
• fixes the Inspector route handoff so the full Harness Inspector no longer collides with legacy query routing
• adds UI reference baselines and visual-diff tooling without including the external reference repository

Validation completed

• Go application, store, HTTP, protocol registry, and focused race tests
• protocol registry synchronization check
• TypeScript typecheck
• Web test suite: 96 files / 550 tests
• production frontend build
• Windows Wails build and computer-use smoke of Settings, Models, custom Provider form, Inspector, and full Harness Inspector

All build/test caches and D:ClaudeProjectscodex-ui-extract remain outside Git.

Known follow-up work

• Production specialist Agents still need their own enabled tool loop; the new attribution protocol already supports their IDs.
• Live command output currently uses bounded polling/output frames rather than a push stream.
• The final handoff did not spend or expose the user's real Provider credentials, so external DeepSeek/MCP/search E2E must be repeated by a maintainer with an authorized test credential.
• File Diff rendering is deliberately bounded and still lacks full artifact pagination.
• Windows 10/11 matrix, reproducibility recheck, Authenticode signing, and Store certification remain release gates.
• This branch started from an older product baseline and may require deliberate conflict resolution against current main; do not replace durable IDs or weaken the new closed activity protocol during convergence.

Release candidate

Unsigned handoff assets are published in the prerelease:
https://github.com/CWNU-Open-Source-Community/Traverse-Board/releases/tag/v0.1.0-pr195-preview

• TraverseBoard.exe is the recommended direct-run artifact.
• PrayuDesktop.msix is an unsigned Development MSIX retaining the historical compatibility identity. It is only for developer sideload/lifecycle testing, not a Microsoft Store product.
• Both artifacts bind commit 8df0df9.

[Qiyuanqiii]

Handoff release and CI note

Preview assets:
https://github.com/CWNU-Open-Source-Community/Traverse-Board/releases/tag/v0.1.0-pr195-preview

The red Release dependency and license boundaries check is the expected exact-release provenance gate, not a code/test failure. The central CI workflow only produces qualifying push runs on main, so a v-prefixed tag pointing at an unmerged PR commit cannot have the required successful main-push provenance.

PR-scoped CI for this exact SHA passed, including Go control plane, TypeScript, Windows/macOS desktop shells, real Edge UI evidence, three-platform LSP, and the Windows dual-product candidate. After convergence and merge, maintainers should create the stable tag from the successful main SHA and let the protected release workflow publish signed and attested artifacts.

Local and GitHub asset digests agree:

• TraverseBoard.exe: ceec0bc4843ef0ddce943188e93bb0bfdd501165576f64b2d7ef197470a860a9
• PrayuDesktop.msix: 5c97812052e233b2f7e988f5eca4af141ef904688c444af1093dc54a33d244b0