What's new in Flashbag.

Improved

• Chat on iPhone no longer zooms in when you tap the message box.
• The top bar and the message box stay in place instead of sliding behind Safari's toolbars.
• Long agent and task lists now scroll on phones.

New Small screens

• Flashbag now works on phones and narrow windows, making remote access from a phone usable.
• The window can now be resized down to 320 by 480.

Improved

• Task runs now show a cost for models whose CLI reports none, including every GPT model.
• Gordon can now retry a skill change that partially failed.

New Agents and Groups

• Organize agents into collapsible, reorderable Groups and subgroups across Agents, Chat, and task agent pickers.
• Create and edit agents from the Agents page, including their name, personal or business context, role, emoji, color, and Group; newly created agents open Chat and begin context-aware guided setup.
• Agent deletion now uses two confirmations, shows exactly which identity, memory, transcript, session, and organization data will be removed, preserves saved tasks and agent outputs, and refuses deletion during an active task run.

Improved

• All available referenced files now get download cards even when they cannot be previewed, and small text file previews continue loading beyond the fourth referenced file.
• Browser uploads now explain when they are unavailable because the destination volume is at least 90% full.
• Usage cost tracking now recognizes Claude Sonnet 5 and GPT-5.6 models.
• Gordon now reports when a skill change succeeds but its CLI links fail to refresh, including guidance to reconcile them.

New Add files from anywhere

• Use the new + control in Chat, Gordon, and Tasks to select multiple files or folders already on the Mac.
• Remote access browser sessions can browse files on the Mac or upload up to 20 files from the current device, with visible progress and the option to resume by reselecting a file after a refresh or interrupted connection.
• Drag and drop files on chat now works with remote access browser sessions.

New Visual file previews

• Files referenced in messages now appear below them: images and text can be previewed, PDFs get file cards, and originals can be opened on the Mac or downloaded from a remote access browser session.
• Messages load at most 20 distinct file references and disclose any remainder.
• Works for both your messages and the agents.

Improved

• Model pickers now show which model Auto will actually use, even when another model is selected.
• Type / in Chat to find and insert available skills.
• You can now press Esc on the keyboard to stop a turn. If a skills popup is currently open Esc closes the popup first, then a second Esc stops a turn.
• Chat now holds its position more reliably while replies stream, older messages load, and you switch between conversations.
• Claude Opus 5 is now available and is the default for new agents.

New Experimental Remote access

• Use Flashbag from another Tailscale-connected computer through password-protected browser access.
• Browser sessions sync messages across open clients, and can be revoked from the Mac.
• Remote access can only be managed directly in the Mac app. Some features (like the new attach files or folders from Finder) are not available via remote access yet.

Chat

• Attach files or folders from Finder to your next chat message, including queued messages.
• Queue one message while an agent replies, then edit it or stop the current reply to send it immediately.
• Sticky day separators and improved scrolling keep new replies in view while preserving your reading position.

Improved

• Choose layout density and activity-detail level in Display settings.
• Tasks now flag missing or inaccessible input sources and pause runs until they are repaired.

• Added GPT-5.6 Sol, Terra, and Luna plus Claude Sonnet 5, updated defaults and Luna or Sonnet 5 handling helper work based on which providers are available.
• Reasoning controls now show only levels supported by the selected model, add None and Ultra, warn when Ultra may rapidly increase usage through subagents, and default new agents to Extra High.

• Task times now use 24-hour local-time formatting consistently, weekly schedules respect the new "Week starts on" setting, and editing a task without changing its recurrence no longer shifts its next scheduled run.
• Agents now pick up task replies posted into their chat on the next continued turn, so they can respond with the same task output the user already sees.
• Agents can now inspect saved task runs through MCP, including run status, final responses, errors, output links, usage, and timestamps.
• Agents are now aware of how Flashbag stores dates and times, and your local time and can convert between them.

New Tasks

• Create reusable automations for any agent, run them on demand, schedule them for later, and review every run from the new Tasks page.
• Schedule tasks by interval, weekday, or monthly day, with optional catch-up when Flashbag was closed during a run window.
• Watch task runs live, inspect the agent's work, open outputs, and send final results back into the agent's main chat when useful.
• Manage tasks from chat too: agents can create and maintain their own tasks, while Gordon can manage and diagnose tasks across agents.

Improved

• Context preload is clearer and more consistent across agents and tasks, with aligned message/fact timelines and better identity estimates.
• Chat and task output now render more cleanly, including headings, paragraphs, lists, tables, code, and Finder-opening file links.
• Dashboard refresh now recognizes Codex sessions with revoked refresh tokens as signed out.
• New agents default to Claude Opus 4.8 whenever Claude is available, with Codex helpers still used when available.

• Chat header token stats are more compact and no longer get cut off in narrow windows. Each stat's breakdown now opens in a hover card (with a "?" hint) instead of crowding the header.
• Chat messages and thinking steps now render Markdown.
• Chat now warns when a conversation is nearing its automatic session refresh.
• A failed turn now stays in its correct chronological place instead of jumping below newer messages after a reload.
• Turn stats now show durations of a minute or more as "4m 15s" rather than raw seconds.

• All-new chat turn view: thinking and tool steps stream in live as a flow strip, with per-turn stats and smoother scrolling.
• Major tune-up of the built-in system prompts: agents follow their instructions more reliably, significant memory and fact improvements.
• New token & cost tracking: a per-agent usage ledger feeds costs in the chat header (cache spend broken out) and in/out/cached breakdowns on every turn.
• New dashboard usage chart: tokens or cost per day, filter by agent, per-segment hover values, agents keep their own colors.
• Sturdier sessions: they survive app restarts, rotate to a fresh session before the context window overflows, and failures show the real CLI error.
• Stopping a reply keeps everything produced so far, and the session stays alive so you can steer and continue.
• Long chats stay fast: history is virtualized so big sessions no longer bog down the UI.
• Gordon got a knowledge base (memory, skills, agents, session lifecycle) so his answers are more accurate.
• Closing the window now hides the app. Click the Dock icon to bring it right back (Cmd+Q still quits).
• Setup re-checks Claude/Codex sign-in on every launch and right after installs.
• Sidebar busy dot clears after a stopped turn, and "last online" updates live while chatting.
• Messages containing markdown headings now display correctly in transcripts.
• Accurate token and cost figures on Codex: numbers no longer balloon as a conversation grows.
• Token gauges no longer overcount turns that use tools.
• The Bootstrap stat now shows where startup tokens go: identity, memory, and the CLI.
• New agents default to GPT-5.5 (Opus 4.8 when only Claude is installed).
• Claude Fable 5 is available in the model picker.
• Gordon greets once per day instead of on every visit.
• "Started fresh" markers show up reliably, and model switches take effect immediately.
• Chat page remembers which agent you were viewing (and your position if you scrolled up).
• Agents page remembers which agent you were viewing.

• Gordon's reasoning effort capped at Extra High (Max tends to overthink on Claude).

• Claude Opus 4.8 is now available in the model picker and is the default for new agents.
• New per-agent reasoning-effort control below the main model (Low to Max; new agents default to High).
• Levels a model can't honor map to its nearest available level, so switching models never breaks.
• Gordon always runs at the highest available level.

• Helper model now tracks the live role assignment instead of going stale mid-session.
• Claude/Codex always run from ~/.flashbag/cwd so the working directory is predictable.

• Light/dark theme switch now actually applies across the whole app.
• Gordon's avatar is amber, matching his accent color.
• Agent picker label is tighter.
• Fixed a hang when reading your shell environment.

• Faster startup: sign-in checks for Claude and Codex now run in parallel.
• Codex works outside trusted Git folders again.
• Agents launched by Flashbag now see the same PATH and tools as your own terminal.

• Linux build.
• Chat feels much better: streaming is smoother, new in-flight indicator, scroll no longer jumps.
• New agents come with their helper CLIs already wired up.
• Gordon picks his own model automatically.
• Onboarding remembers what it found between restarts (no more redoing the same checks).

• Fixed a crash that could hit during import.

• Gordon now handles agent identity during chat-log import end-to-end.

New Gordon

• Gordon arrives — a built-in assistant who sets up your agents, writes your skills, and answers questions about what your agents produced.
• Agents now have a type, so the assistant behaves differently from the agents you create yourself.

New The designed interface

• The first designed interface replaces the bare one that existed only for testing.

New Output folder

• Files your agents create now land in a dedicated output folder instead of being dropped in Downloads.

Improved

• Flashbag now runs on Claude alone. Until now it only worked if both Claude and Codex were installed.
• Skills are now linked correctly when an agent is created.
• The dashboard and the agents page show the real number of facts instead of zero.
• The activity dot now stays visible when an agent finishes instead of flashing and disappearing.
• Chat stops scrolling itself once you scroll by hand.
• Claude and Codex sign-in status refreshes when the app starts.
• The app no longer sends you straight to Gordon on launch.

New More than one agent

• Create, edit, and delete agents in the app. Each one gets its own folder, its own memory, and its own skills.
• Agents run side by side — set one working, switch to another, and a dot tells you when each is done.

Improved

• The debug view was rebuilt: calls appear in the order they happened, and the internal searches that were previously invisible now show up with their own cost.
• The interactive prompts built into Claude are switched off, so a reply can no longer stall waiting on a question you were never shown.
• How long a reply may take before it gives up is now a setting.

• Fixed Codex failing on every call — it was being started without a flag it needs.
• Fixed Flashbag not finding Claude or Codex when they were installed somewhere other than the usual place.

New The first build

• Chat with your agent from a desktop app instead of a terminal.
• Memory: your agent pulls facts out of the conversation as you talk and remembers them afterwards.
• Claude or Codex does the thinking, whichever you have installed.
• A debug panel on every message showing the tokens spent, the time taken, and each call that was made underneath.
• A dashboard that detects which command-line tools are installed and whether they are signed in.