Skip to main content
Version 1.6.0

Finch 1.6.0

Global right panel and mini program navigation stack, host-rendered Diff with a revamped changes panel, unified Twemoji rendering, and a broader model ecosystem with xAI, DeepSeek vision models, and GLM 5.3.

Back to changelog

1.6.0 Changelog

Span: v1.5.4 → v1.6.0 (552 commits)

New

Right panel & mini program system

  • Global right Panel: a new global right-side panel workspace with responsive width strategies; panels survive across sessions, limited to 20 instances across browser/Webview.
  • Session-level Webview Panel Host: panels can attach to a specific session and restore with it; supports file preview and a built-in browser.
  • Mini program navigation stack: mini programs can open child mini programs, built-in file previews, and browsers (AppView three-level navigation stack) with multi-level breadcrumb back; parent mini program pages stay alive during child navigation instead of refreshing on every return.
  • New mini program APIs: ctx.ui.openDiff (host-rendered diff), ctx.ui.delivery (deliverables), ctx.spaces (space list), dynamic toolbar updates (setToolbar/updateToolbarItem), window.finch.ui.toast/confirm, and the finch-file local file protocol exposed to Webviews.
  • Terminology unification: "mini tools" are now consistently called "mini programs" across the app UI and copy.
  • Webview experience: eliminated the light-mode flash on first render — the first frame stays hidden until theme injection completes; AppView child levels now hook into window forward/back history.

Diff / changes panel

  • Host-rendered Diff service: supports comparing two file paths or multi-file Git refs, with patience anchors + an LCS fallback algorithm for better large-file diff performance and precision.
  • Changes panel: the diff panel is renamed "Changes", integrates write files, optimizes the line-number gutter with modification counts; deleted files are marked with strikethrough + a tertiary color; large-file diff performance is improved.
  • Annotation integration: the changes panel and file preview support quick selection-based annotations that auto-generate annotation tokens with paths and line numbers.

Messaging & interaction

  • Twemoji rendering everywhere: 473 common emoji SVGs are bundled locally (no CDN, works offline) and rendered consistently across AI replies, user messages, sidebar titles, and ask cards, with a system-font fallback outside the subset.
  • Quote replies: every AI reply gets a quote button — lightweight quotes no longer inline the full text; hover cards show "Replying to: xxx" with a jump to the original reply.
  • Annotation enhancements: annotations support a reminder field (model prompt only, not displayed), hover cards truncate overly long content, and previews no longer leak internal tokens.
  • New conversation capabilities: the Session.read tool reads LLM reply turns from a session in a structured way; a "Open in new window" option joins the session menu; child-session cards stay visible with running-state indicators.
  • Pending message drawer: shows parsed human-readable text without leaking annotation payloads.

Model ecosystem

  • xAI: native device-code login (OAuth) for xAI with curated Grok models built in.
  • DeepSeek: new vision model (deepseek-v4-flash-vision-exp, 1M context); vision-capable models get a vision badge in the model list.
  • GLM 5.3: new model entry; restricted models (GLM/Kimi/DeepSeek) only show low/high/max thinking levels.
  • Custom models: filter and batch-select by ID; configure vision capability; one-click add via the finch://models/add deep link.
  • Model probing: models.dev local cache corrects auto-detected model names and context lengths; models flagged as non-text output are filtered out; a new 372K context-length preset.
  • OpenCode Zen Go: built-in OpenAI-compatible provider with 12 models.
  • Model picker UX: without an explicit model, the last model used in a real conversation is preferred; hovering a usage chart legend highlights the matching bar.

Other

  • Traditional Chinese: new zh-HK language support with proofread main UI copy.
  • SQLite app storage: new SQLite application storage infrastructure (app.db), powering usage caching and more.
  • MCP 2.0: client upgraded to the official @modelcontextprotocol/client 2.0 with protocol-version negotiation fallback.

Improvements

  • Performance: throttled agentStore/panelStore persistence writes (main-thread load down ~40% during streaming); throttled usage snapshots while running; async chunked quicksearch no longer stutters; progressive rendering and first-frame cost optimizations for long sessions; low-frequency state animations now render as discrete jumps.
  • Stability: in-place rendering of resumed running turns eliminates Timeline tearing; streaming output freezes when reading above the bottom to avoid jitter; refined running states (empty stop treated as network error, reconnect/overload shown directly, interrupted turns marked incomplete).
  • Experience: right-panel buttons show shortcuts in tooltips; browser default empty-state guidance; max panel width raised to 2560; faster settings-window load.

Fixes

  • Multi-window: fixed user-message sync across windows, unread-state races, follow-up messages not rendering in sync, and Dock badge unread sync.
  • Permissions/Interaction: fixed historical permission cards flashing when switching sessions; pending backend permission requests are now cleared when toggling acceptCalls; interaction card response and recovery.
  • Rendering: fixed blank history after session restore, abnormal timeline dates, image attachments/base64 failing to render, Chinese-boundary URL rendering issues, and Windows paths rendered as blocked.
  • Input: fixed IME Enter accidentally submitting in the rename input; drag-and-drop of directories now inserts @ mentions.
  • Stability: fixed message-queue races during auto-compaction, repeated compaction prompts, 524 long backoff not auto-reconnecting, and empty timelines when restoring window sessions.
  • Grep: fixed content-mode context line read failures (relative paths resolved against the wrong cwd).
  • Models: fixed GLM text-model image requests, OAuth model connection tests, temporary refresh failures deleting saved credentials, and Provider-enable races after login.
  • OAuth: fixed callback-page localization, popup window ownership, cancelled authorization keeping the termination page, and icon version-parameter mismatches being rejected.
  • Other: fixed rename IME misfires, hover sidebar menus closing accidentally, settings sub-page jumps, MCP client 1.3.x extraction failures, mini program prerelease-version compatibility checks, and directory left-ellipsis segment ordering.