dia-source-analysis

Analyze Dia Browser source

Dia (/Applications/Dia.app, The Browser Company, company.thebrowser.dia) is a WebKit-based browser (~1.4 GB) that shares ArcCore.framework with Arc. termio uses it as a reference for how a shipping product drives AI coding agents as sandboxed child processes (Dia shells to the real Claude Code CLI under Seatbelt) and for its agent-server / IPC architecture. This skill captures how to dig into it and what's already known, so analysis doesn't start from scratch each time.

Golden rules

• Read-only, on the user's own machine, for research. This is proprietary code shipped to the user. Use it to understand and learn, never to copy/redistribute verbatim or ship lifted code. State this caveat when sharing recovered source.
• Verify before trusting old findings. The "Established facts" below are version-stamped. Dia auto-updates (Sparkle); bundle layout, file hashes, runtime versions, and even whether source maps ship can all change. Always re-read info.json + CFBundleShortVersionString first and treat mismatches as "re-derive from scratch."
• Two recoverability tiers: the Bun/TS agent backend is near-fully recoverable (source maps); the native Swift app is symbols/structure only (no real source).

Established facts (as of Dia 1.34.2 — RE-VERIFY)

Agent-server info.json (1.34.2): version 1.0.0, buildDate 2026-06-10, commit 47de9fbe6c4, claudeCodeVersion 2.1.131. (Was 1.32.0 / commit 55d9d2e174b / buildDate 2026-05-20.)

• ⚠️ Source maps NO LONGER SHIP as of 1.34.2. extract-sourcemaps.py returns "No .map files found" — the Bun/TS agent-server backend is no longer recoverable as near-original TS (only logic compiled into the Mach-O remains). The native-Swift tier is unchanged: symbols + embedded source paths only. So "the big win" in §1 below is gone on current builds; keep the procedure for older installs / in case they return.
• Agent backend = Bun-compiled standalone Mach-O (agent-server, handler) + the real Claude Code CLI (claude, 206 MB, itself a Bun binary). No separate Node/Bun runtime is installed — each binary embeds Bun. They run as child processes under Seatbelt sandbox (sandbox-exec -f agent.sb / agent-claude-code.sb, params via -D DATA_DIR=…), talking to the native UI over SSE/IPC (transport/sse.ts, ipc-gateway.ts). Per-context workspace at data/contexts/{contextId} with resumable disk buffers (session/buffer.ts, JSONL). This is the part most relevant to termio — termio runs agents in real PTYs (.exec) instead, but the child-process-under-sandbox + resumable-buffer + IPC design is the directly comparable bit.
• Streaming smoothness comes from native rendering + session/update-batcher.ts: deltas are coalesced and flushed only on ≥256 B OR 250 ms idle OR completion — so the UI updates a few times/sec, not per token. The Bun agent-server is backend orchestration and does NOT touch UI perf.
• AI chat ("AssistantPanel") renders NATIVELY, not in a webview. Markdown parsed by cmark (swift-markdown), code highlighted by Highlightr (CodeAttributedString → NSAttributedString), math by SwiftMath (MTMathListDisplay, CoreText). UI is AppKit NSViewControllers (source paths Frameworks/BoostBrowser/Sources/AssistantPanel/*.swift).
• The AI input is the native AppKit module BoostCommandBar (Frameworks/BoostBrowser/ Sources/BoostCommandBar/*): a token/pill TokenTextView w/ SkillPillTokenViewProvider; Skills = SkillsV3; @-mentions (AtMentionKeywords: @Search/@Slack/@Gmail/@Notion) insert context/tool pills. On-device cmd_t_router (3-label intent) + skills classifier rank proactive suggestions.
• WKWebView (~45 refs) is for web pages + HTML "artifacts" (reports/slides via report-kit/slide-kit), NOT for chat bubbles.
• On-device ML (OnDeviceLoRAadaptors bundle, 152 M): one shared DistilBERT-base encoder (126 M, fp16) + three LoRA adapters + heads — cmd_t_router (input intent), skills (which skill to fire), sensitive_content (privacy gate). Run via MLX (mlx-swift).