[Epic] Windows Phase 9 — final parity push (APIService, self-compact loop, multi-assistant, browser, gaps)

[Aryansharma28]

Goal

Close the remaining gaps between the Windows Tauri port (`windows/`) and the macOS Swift app (`Sources/`). After Phase 9 lands, the only intentional divergence will be platform-specific shell (Liquid Glass, NSStatusItem tray, SwiftTerm, Amphetamine, LidStateDetector) and explicit out-of-scope items (ConversationMarkdownExporter, SessionMigrator).

Phases 0–8 — already shipped

• Phases 0–5 (PR [Windows parity] Phases 0–5 — full Windows port #98): build baseline, logging, native cmd.exe terminal, polished local board, GitHub depth, session orchestration + hooks, breadth/multi-assistant scaffolding, Mutagen sync
• Phase 7 (PR [Windows parity] Phase 7 — channels + kanban CLI #105): channels + DMs + `kanban` CLI
• Parts 1–4 follow-ups (PR feat(windows): Phase 7 follow-ups (Parts 1–4) #115): `KANBAN_CARD_ID` injection, channels-store cleanups, DM panel, chat notifications, drafts watcher, image upload, message actions (edit/delete/reactions/@mention), pinned cards, self-compact data shapes + drop guard
• Phase 8 (PRs feat(windows): image-paste pipeline for prompts #119/feat(windows): Process Manager modal (Ctrl+Shift+M) #120/feat(windows): Session History parity — transcript font + auto-load #121): image-paste pipeline, Process Manager modal, Session History parity (transcript font + auto-load)

Current main: `770c5e4`

Phase 9 sub-issues

Land each as its own PR off `main`. PRs should target `main` directly — there is no integration branch for Phase 9.

• [Phase 9] feat(windows): APIService entity + per-card endpoint binding #122 APIService entity + per-card endpoint binding (medium) — starts first; [Phase 9] feat(windows): multi-assistant — Codex + Gemini adapter suites #124 depends on this
• [Phase 9] feat(windows): self-compact generation loop + Claude Code statusline #123 Self-compact generation loop + Claude Code statusline (medium) — independent
• [Phase 9] feat(windows): multi-assistant — Codex + Gemini adapter suites #124 Multi-assistant: Codex + Gemini adapters (very large; split into 3 sub-PRs) — depends on [Phase 9] feat(windows): APIService entity + per-card endpoint binding #122
• [Phase 9] feat(windows): embedded browser tabs subsystem #125 Embedded browser tabs subsystem (very large) — independent
• [Phase 9] feat(windows): standalone parity gaps — lastOpenedAt, channel reorder, bulk kill, checkpoint affordance #126 Standalone gaps: `lastOpenedAt`, channel reorder, bulk Kill All Managed, Ctrl-held checkpoint cursor (small, one PR)

Recommended order

1. [Phase 9] feat(windows): standalone parity gaps — lastOpenedAt, channel reorder, bulk kill, checkpoint affordance #126 first — small mechanical wins, builds confidence with the codebase
2. [Phase 9] feat(windows): APIService entity + per-card endpoint binding #122 — small/medium, prerequisite for [Phase 9] feat(windows): multi-assistant — Codex + Gemini adapter suites #124
3. [Phase 9] feat(windows): self-compact generation loop + Claude Code statusline #123 in parallel with [Phase 9] feat(windows): APIService entity + per-card endpoint binding #122 — no dependency
4. [Phase 9] feat(windows): multi-assistant — Codex + Gemini adapter suites #124 after [Phase 9] feat(windows): APIService entity + per-card endpoint binding #122 lands
5. [Phase 9] feat(windows): embedded browser tabs subsystem #125 any time — independent, but largest single chunk

Working agreement

• macOS sources under `Sources/` are read-only reference — all changes land in `windows/`
• Each issue gets its own worktree branch off `origin/main`
• Verify loop: `cd windows && npm run build` AND `cd windows/src-tauri && cargo build` both green; `cargo test` for any new modules; no unit is "done" without evidence
• Conventional Commits — release-please uses these for changelog generation
• Don't skip pre-commit hooks (`--no-verify`); investigate and fix the underlying issue if a hook fails
• Out-of-scope items are listed in each sub-issue's "Out of scope" section — respect those deliberately

Out of scope for the whole epic

• Liquid Glass UI, NSStatusItem tray, SwiftTerm, Amphetamine, LidStateDetector (macOS-only stack)
• ConversationMarkdownExporter (copy-conversation-as-markdown) — intentional divergence
• SessionMigrator (cross-assistant migration)
• ImageSender's tmux clipboard-paste loop — Windows uses the markdown-ref fallback (Phase 8 covers this)
• `discoveredRepos` on Link (intentionally dropped — see `merge_ops.rs` rationale)
• Native `TmuxLink` struct (Windows uses WSL tmux; `manual_overrides.tmux_session` covers the override case)
• `MessageSource.External` enum case

Notes for agents

Each sub-issue body is self-contained — an agent picking one up has enough context from the issue alone. Read the issue, check the latest `main`, and start. Look at the file paths referenced in each issue's "macOS reference" + "Windows scope" sections — those are the breadcrumbs.

When in doubt about a divergence: prefer mirroring macOS (it's the reference) unless the issue body explicitly says otherwise.