1.7.2 Changelog

New

Long sessions & compaction

• Agent-initiated compaction: the Session tool gains a compact action, so a compaction can be started at a tool boundary while a run is in progress and execution continues from there—no need to wait until the context crosses the line.
• Visible compaction: the progress ring spins and finishes smoothly while compacting; the compaction boundary card now includes token counts and a summary, making it clear how much context sits on either side of the boundary.
• Space reclamation for archived sessions: archived sessions support storage statistics and space reclamation, so long-accumulated session data can be released safely.
• Long-session stress testing: a new long-session compaction stress-test CLI (run through npm run, with scenario suites, JSON output, and assertion levels) for local regression of long-session scenarios.

Mini Tool platform

• Immediate follow-up API: the mini program session API gains ctx.sessions.promoteTurn, which moves a mini program's own queued message to the front of the queue along the same cooperative-interruption path as the UI's "follow up now"; it is paired with ctx.api.supports() detection so mini programs can stay compatible with older hosts.
• Session titles: mini program session information gains a title field, so mini programs can display the session name directly.
• Toolbar layout: the mini program toolbar scrolls horizontally and gains an icon-only mode, so narrow windows no longer squeeze the input area.
• Background subtask entry: mini program views get the same background subtask menu as the main session, letting them jump straight into background sessions they created, reusing the same unread and running-state badges.

Models & providers

• Custom request headers: model providers can be configured with custom request headers, making it easier to connect gateways or self-hosted services that need extra authentication headers.
• Model catalog update: updates the built-in model catalog and authoritative catalog, adding GPT-6.1 Sol; a session's model choice is now pinned, so reopening a past session restores the model it used at the time.

Diagnostics

• Main-process network console: a new network console means proxy and request problems no longer require digging through logs; database performance diagnostic logs are added alongside it.

Improvements

• Faster long-session opening: branches and entries are merged into a single pi round trip; the first screen reads only the tail window of a branch, and earlier rounds are fetched on demand.
• Leaner session file compression: on-disk compression of session logs now streams line by line instead of reading large files into memory all at once.
• Large code blocks no longer block typing: chat syntax highlighting now runs asynchronously in slices.
• Image format in preview: the full-screen preview title bar derives an image format chip from the MIME type, file extension, and file magic number, and shows nothing when the format cannot be identified.
• Consistent quota reporting: subscription quotas are shown consistently in Settings, the model menu, and the tray, with clearer reset times and plan display; an empty card is no longer shown when quota has not been fetched.
• Browse models while signed out: when signed out, all of Finch's available models are shown, narrowing to your own permissions after signing in.
• MCP presentation: refines the MCP service card and its capability list, aligns the height and empty state of the switch dialog, and collapses a fully empty connector list into a centered icon, a single line of guidance, and an add button.
• Mini program permission copy: permission declarations and incremental review copy are clearer, so a first-time grant is no longer mistaken for a "permission upgrade".
• Pin queued messages: queued messages can be pinned, with clearer action hints and queue-full messaging.
• Bash reclamation: Bash shells are reclaimed by run ownership, and a session-level force-kill entry point is added.
• Storage layering: configuration, rebuildable caches, and run state are split by purpose; legacy data such as model config versus model settings, session display state versus clearable caches, window size versus session restore state, automation tasks and their run records, log settings and mini program state, workspace context, and unread sessions and usage history all migrate automatically at startup, with a standalone CLI following suit. Files left over from the previous version are cleaned up according to the established list.

Fixes

Sessions, timeline & compaction

• Fixes tools being executed twice after reconnecting from a network hiccup: calls with identical parameters reuse the result that already succeeded before the reconnection, duplicated tool cards in the timeline collapse into one marked ×N, and side effects such as summoning multiple agents twice are eliminated.
• Fixes queue problems during compaction: follow-ups jumping the queue, queued messages arriving out of order, the queue being dispatched early while a compaction rerun is still in progress, and leaked state when compaction finishes.
• Fixes Unicode line separators (U+2028 / U+2029) causing dropped frames in session history IPC.
• Fixes a blank session timeline after a model timeout.
• Fixes link parsing in assistant replies, where bolded URLs and the markup immediately following them are no longer swallowed.
• Fixes image previews and relative links under Chinese paths: non-ASCII characters in Markdown links were double-encoded, so the main process could not find the file; relative links crossing between sessions are fixed as well.
• Fixes attachments that kept occupying the input box after a failed read, and draft attachments left behind after sending from Home.

Queue, follow-ups & input

• Fixes four issues in the follow-up path: bottom-stick scrolling briefly failing, the page bouncing back down when a reply completes, the pending-send button staying in a loading state after a follow-up is consumed, and submitting an edit accidentally sending queued messages.
• Adjusts the waiting policy for permission cards: in ask mode there is no longer a countdown that auto-declines, so approval can take as long as it needs; act mode keeps its two-minute countdown and timeout decline.

Models, accounts & mini programs

• Fixes custom models being misjudged as not supporting thinking, built-in providers falling out of sync with the remote catalog, local and remote providers being merged, messages failing to send after switching to a particular model, and OAuth sign-in from model settings.
• Fixes Finch Cloud account and subscription state falling out of sync, quota still showing after signing out, short-lived credentials being refreshed immediately, and quota differing between the tray and Settings.
• Fixes mini program subtasks showing unread after completion, archived sessions still showing unread, the panel fetching its list repeatedly, the full-screen page showing a duplicated hierarchy, a missing first breadcrumb level, and uninstalled bundled mini programs being reinstalled automatically.

Platform stability

• Fixes force-quitting a session restarting the pi runtime, the main process potentially launching twice in development, and the app failing to close after shutdown cleanup.
• Fixes incomplete file search results in large workspaces, the Workspace split still depending on old data files, the Space "more" state frequently failing, and other expanded spaces collapsing when one space is expanded.
• Fixes quotes and dollar signs in shell commands being expanded prematurely.