fx changelog: what's new each month
Every stable fx release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 4 months covered; the current month updates daily.
November 2026 · so far
1 release: 0.0.17
November brought subagent support to fx, allowing the main agent to spawn and manage multiple specialized agents that work in parallel and report results back. Users can now switch between agents, customize each with different models and settings, and organize complex tasks across a nested agent hierarchy.
Highlights
- Subagents work alongside the main agent, with the parent automatically reusing idle ones for related tasks.
- Switch to any subagent to watch, message, or answer its prompts using the agent list.
- Each subagent has independent configuration: model, effort level, Fast mode, and permission settings.
- Agent profiles let you define tools, MCP servers, models, and prompts that persist in ~/.fx/agents or .fx/agents.
- The main agent detects when a subagent's command stalls and gets notified after 2 minutes of silence.
New
- Subagents that work in parallel with the main agent and report results back
- Ability to switch to any subagent to watch, message, or answer its prompts
- Agent list interface accessible with Down or Ctrl+X to see, search, switch, and close agents
- Per-agent configuration for model, effort level, Fast mode, and permission mode
- Nested subagent support up to two levels deep
- Agent profiles in ~/.fx/agents or .fx/agents to persist agent configuration
- fx --profile command to start a session in a named agent profile
- Timeout detection when a subagent's command produces no output for 2 minutes
Improved
- Main agent now sees and can reuse idle subagents each turn for related work
- Subagents can be customized with different tools and MCP servers independently
October 2026
4 releases: 0.0.13 → 0.0.16
October brought major performance improvements, durability features for production deployments, and expanded model support. fx is now significantly faster and more memory-efficient while supporting multiple agents and recovery from interruptions.
- New libfx checkpoints cannot be restored by versions older than 0.0.13; existing checkpoints remain compatible.
Highlights
- Durable libfx sessions can survive crashes and redeploys, with automatic recovery for marked-idempotent tools.
- Multiple named agents can run in one app with separate models, instructions, and sessions.
- Peak memory usage dropped 69.5% and turn completion sped up 3.3× in image-heavy workloads.
- libfx package downloads shrunk 60–61.6%, with native platform-specific binaries replacing universal packages.
- New Ultrafast mode available on supported OpenAI models for faster inference at higher cost.
New
- Ultrafast inference option on OpenAI models through AI Gateway, supported in subagents and ACP sessions.
- Multiple named libfx agents per app, each with independent model, instructions, tools, and sessions.
- wakeHandler() HTTP handler for managing queued prompts and recovery across all libfx agents.
- libfx support for ChatGPT and Grok subscriptions with built-in sign-in helpers.
- Custom OpenAI-compatible endpoint support in libfx, including local model servers.
- Clients can reconnect to libfx reply streams from their last received event across different servers.
- Cross-process steering and cancellation of running libfx sessions or resumption of interrupted turns.
- Image deduplication: fx saves images once per session and reuses them across turns and forks.
Improved
- Terminals launched through the shell tool now stop when fx exits instead of persisting.
- Slack integration simplified: /mcp add slack configures, signs in, and connects in one command.
- libfx sessions can override their agent's model, reasoning effort, and instructions per-session.
- Text session resume time improved approximately 18% over previous version.
- Durable libfx sessions skip rereading unchanged log entries, reducing storage calls by 36% in benchmarks.
- Model catalog refreshes in the background; newly released models appear without restarting fx.
- Long agent turns use 50–59% less CPU with 1,000 tool steps.
- Prompt history loads faster by skipping entries from other workspaces.
Fixed
- Invalid Base64 image input now fails immediately from prompt() instead of later in the turn.
- Durable libfx sessions no longer copy prompt, steering, and tool images into every checkpoint.
- Tool result timestamps are preserved when turns are rebuilt or resumed.
- Caching behavior improved: 41% fewer uncached input tokens across model families; 87.5% fewer prefix-breaking requests.
September 2026
5 releases: 0.0.8 → 0.0.12
September brought significant API changes, a streamlined interface with fewer commands and tools, and major performance improvements. The month focused on simplification and speed, with session startup times dropping dramatically.
- createFxAgent() API changed: now takes apiKey and optional model directly; Agent configuration through env variables no longer accepted.
- createFxAgent() exposes prompt, checkpoint, and close instead of session, model, and config.
- The memory tool has been removed (existing saved memories are preserved).
- The public --record startup flag has been removed.
Highlights
- Session startup is now 500+ times faster, dropping from 95 seconds to 0.17 seconds.
- The shell tool was redesigned with 75% fewer actions, and subagent commands reduced by 67%.
- You can now configure custom OpenAI Chat Completions connections and route between multiple providers.
- fx can install and manage Slack workspace bots directly from the CLI.
- Custom TUI themes and vision model image reading were added to expand customization.
New
- capabilitysearch replaces skillsearch for finding available tools.
- Subagents can now use their own model and reasoning effort settings independently.
- Send feedback to running named subagents without interrupting their current tool.
- Model picker (ctrl+p) lets you switch models while writing, preserving your draft.
- Automatic recovery: fx now retries and continues on its own without manual /continue.
- Custom TUI themes can be loaded from ~/.fx/themes/<name>.json, including VS Code themes.
- readfile now attaches PNG, JPEG, GIF, and WebP images to supported vision models.
- Interactive flags (--provider, --model, --effort, --fast) for one-off runs without changing preferences.
Improved
- Enter now steers active turns instead of queuing follow-ups; ctrl+enter does the same.
- New conversations automatically get a short title from the first prompt; turn this off in settings.
- Shell failures now give the model clearer recovery guidance.
- Resume recovery restores shell activity and scrollback more reliably for long conversations.
- Resolved questions and answers render cleanly across terminal widths.
- MCP errors have stronger secret protection.
- Oversized PNGs shrink before sending without modifying the original file.
- Structured replies now use GitHub-flavored Markdown with wrapping tables and clickable URLs.
Fixed
- fx --resume last now completes in 0.5 seconds instead of 33 seconds.
- /trace now provides better diagnostics for session-title issues.
- System notices have cleaner visual treatment.
- Detached processes such as tmux, ssh-agent, and browser drivers now stay alive between shell calls.
- /mcp list now opens the interactive Servers menu.
- Long sessions show the most improvement in speed gains.
August 2026
7 releases: 0.0.1 → 0.0.7
August brought major improvements to session management, permission handling, and provider integration for fx. The month focused on reliability, workspace features, and streamlined command discovery.
- Removed /appearance, /input, and /maxxing commands; fx now uses the same layout everywhere.
- Removed /provider slash command; provider selection moved to /setup and top-level fx provider.
- terminal.exec now requires timeoutms between 1 millisecond and 10 minutes; use terminal.start for long-running work.
- Removed sandbox configuration and related status fields and commands; host command execution now uses ordinary subprocesses.
Highlights
- Session resume command lets you pick up from your latest workspace session or a specific saved session ID.
- New subscription login flows for Codex and Grok providers unlock authenticated models and features like images and automatic reviews.
- Permission system now lets bounded read operations and safe edits run automatically while keeping sensitive actions gated for human approval.
- Terminal execution unified across captured commands and interactive sessions with proper shell profile loading.
- Active-turn steering with Ctrl+Enter lets you redirect fx's work at the next model boundary without queuing a follow-up.
New
- fx session resume command to continue the latest or a specific workspace session.
- Headless permission prompts with --prompt-permissions for JSON and quiet runs that need Y/N approval.
- Terminal attention bells to flag waiting panes in terminal multiplexers when fx pauses for input.
- Workspace status line option to show active path and Git branch in the UI.
- Project MCP configuration through .mcp.json for workspace-specific servers.
- Top-level fx mcp command with add, list, remove, auth, logout, and trust subcommands.
- Capability discovery to search skills and MCP tools together before loading a match.
- Collapsed tool calls view showing one summary per tool-call group with full calls in the transcript.
Improved
- MCP server awareness shows the agent available aliases, status, and tool counts for better capability finding.
- Auto mode recovery lets the agent revise its plan after denied or timed-out reviews instead of stalling.
- Trusted auto mode actions allow bounded reads and safe Git operations to proceed without extra review.
- Command discovery ranks exact, prefix, and substring slash-command matches with highlighted help.
- Session cache handling continues writes and keeps listings current during contention from other processes.
- Memory clearing now reports real deletion failures instead of falsely claiming success.
- Notification sounds use clearer 48 kHz AAC cues with proper volume differences between actions.
- Transcript scrollback preserves retained rows in native scrollback across pruning and resize.
Fixed
- Release validation checks hardened to prevent false failures in asynchronous terminal and Gateway readiness.
- Wrapped diff layout keeps file-diff rows aligned with gutters across Inline, Review, and Full detail views.
- Inline picker layout keeps transcript and composer adjacent instead of leaving a blank band when closing pickers.
- Skill creation shows invalid names inline and keeps the session usable instead of breaking the interface.
- Model catalog rejects malformed responses with a nonzero exit instead of treating them as empty.
- Background URLs refuse /background open for stopped or stale tasks to prevent unrelated processes from opening after port reuse.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. fx: pricing, features and alternatives · All changelogs