Changelog

UI fix (build 36)

• Main window icon — the Silica app icon now appears in the title bar again on the main window.

Browser Agent Support (build 35)

• Smaller app download — Playwright/Chromium is no longer bundled (~50 MB updates instead of ~370 MB).
• One-time install — browser automation downloads from Settings → Advanced → Browser Agent Support (~350 MB) on first use.
• Honest install flow — agent browser tasks prompt to install when missing; declining fails honestly with no fake page proof.
• Playwright-backed proof — URL, title, readable text, links, and buttons with clean failure when proof is unavailable.
• No Screen Recording prompt — browser URL proof uses Playwright only, not automatic screen capture.

Gmail Drafts

• Natural compose requests — Gmail draft requests can open a compose draft through the browser automation path without sending anything.
• Clear sign-in blockers — if Google sign-in blocks Playwright compose verification, Silica explains the blocker and does not claim the draft was completed.

Agent Reliability

• More deterministic local workflows — web research, notes, file output, reminders, and recovery after tool failures are more reliable on small local agent tasks.
• Cleaner chat summaries — raw browser proof/debug dumps are hidden from normal chat summaries while verified evidence remains visible.

Browser Agent Support (build 35)

• Smaller app download — Playwright/Chromium is no longer bundled (~50 MB updates instead of ~370 MB).
• One-time install — browser automation downloads from Settings → Advanced → Browser Agent Support (~350 MB) on first use.
• Honest install flow — agent browser tasks prompt to install when missing; declining fails honestly with no fake page proof.
• Playwright-backed proof — URL, title, readable text, links, and buttons with clean failure when proof is unavailable.
• No Screen Recording prompt — browser URL proof uses Playwright only, not automatic screen capture.

Gmail Drafts

• Natural compose requests — Gmail draft requests can open a compose draft through the browser automation path without sending anything.
• Clear sign-in blockers — if Google sign-in blocks Playwright compose verification, Silica explains the blocker and does not claim the draft was completed.

Agent Reliability

• More deterministic local workflows — web research, notes, file output, reminders, and recovery after tool failures are more reliable on small local agent tasks.
• Cleaner chat summaries — raw browser proof/debug dumps are hidden from normal chat summaries while verified evidence remains visible.

Presentations

• Direct chat-to-slides generation — no separate outline LLM step or editor gate before HTML generation.
• Faithful slide content — titles, Key Concept, and bullet text render as written in chat.
• Per-slide styling — light or dark themes chosen by the model; no forced uniform dark shell.
• Optional visuals — real image URLs when they help; inline SVG for diagrams; text-only slides are fine.

Agent performance

• Faster deterministic paths for common file organization and simple create-file/folder tasks.

UI

• Silica blue palette rollout, Models Hub redesign, and broader UI polish across chat, settings, and workflows.

New & Improved

• Financial Analysis workspace — analyze spreadsheets directly in chat: growth rates, margins, variance, DCF, and benchmark comparisons with charts and exportable results.
• Financial Pack v2 — expanded sector and company benchmarks (including Healthcare), with in-app pack updates from Settings -> Knowledge Packs.
• Smarter spreadsheet handling — multi-file and multi-sheet support with active file selection, sheet dropdown, and clearer dataset disambiguation when attachments are ambiguous.
• Gemma 4 MTP acceleration — Gemma 4 26B/31B models now route through llama-server with draft-MTP for faster generation on supported Apple Silicon Macs (requires llama-server b9585+).
• Better Gemma 4 support overall — improved built-in Gemma 4 vision handling, higher Pro-tier context defaults on 24 GB machines, and clearer guidance to update Vision Support when needed.

Fixes

• Fixed follow-up prompts losing tables and charts on repeat financial analysis questions.
• Fixed Excel workbook extraction and prevented silent text-only fallback on compute turns.
• Improved benchmark routing so sector/ticker clarifications reflect the installed Financial Pack, not just bundled seed data.
• Fixed multi-series chart legend colors and split charts by numeric scale for clearer visualization.

New & Improved

• Initial release candidate providing stable local models mapping to core Apple Silicon architectures.
• Introduced the full Native Mac UI including dark-glass translucency and command palettes.
• Launched Voice and Capture utilizing native Whisper workflows completely without the cloud.

Fixes

• Fixed major memory retention leaks during persistent Agent workflows.
• Resolved the model selector sheet hard-freezing the main thread by deferring SwiftData external storage writes.
• Refined macOS Gatekeeper download and warning signatures.