Skip to content

v0.0.17

Latest

Choose a tag to compare

@github-actions github-actions released this 14 Nov 21:37
· 544 commits to main since this release
d1d4a88

Subagents now work alongside you: switch between agents and run several fx instances in one window. Vercel Token Pass is now first-class in fx. Tool definitions 48.0% smaller, and 82.0% of later input from the prompt cache, up from 61.3%.

New Features

  • Subagents work alongside the main agent and report back when they finish.
  • Switch to any subagent to watch it, message it, or answer its prompts.
  • A parent agent hears when its subagent's command has printed nothing for 2 minutes.
  • Open the agent list with Down or Ctrl+X to see, search, switch, and close agents.
  • Each subagent has its own model, effort, Fast mode, and permission mode, and subagents can nest two levels deep.
  • The main agent sees its subagents each turn and reuses idle ones for related work.
  • Agent profiles in ~/.fx/agents or .fx/agents set an agent's tools, MCP servers, model, and prompt.
  • fx --profile <name> starts a session in an agent profile.
  • fx --instances <n> and /launch run several fx instances in one window.
  • fx sessions --live lists open sessions and their subagents.
  • fx session <id>/<agent> reads a subagent's conversation, and --tail <n> shows only the last turns.
  • Vercel Token Pass models come first in the model picker.
  • /credits and fx credits show your Token Pass credit and its reset date.
  • Token Pass sessions use Token Pass models for titles, images, and auto mode review.
  • Login offers Token Pass to teams that don't have it yet.
  • The toolbox tool replaces skill, capability_search, and mcp_select_tool, so every request sends the same tools.
  • fx mcp call TOOL runs an MCP tool from a shell command.
  • Custom connections can set effort levels, request fields, headers, timeouts, and a title model.
  • Cache warming keeps Claude's prompt cache warm while you pause, and "cache_warming": false turns it off.
  • Reopening a session after a crash or a closed window sends the interrupted request again.
  • Plain fx shows Interrupted turn · /resume to continue when a turn was left unfinished.
  • fx config lists, reads, and changes settings.
  • fx ask --config key=value overrides a setting for one request.
  • fx completion bash, zsh, and fish print a shell completion script.
  • fx "<prompt>" starts a session with that prompt, and -r is short for --resume.
  • FX_HOME moves the profile folder from ~/.fx.
  • Ctrl+T cycles the reasoning effort.

Improvements

  • Toolbox raised prompt cache reads from 61.3% to 82.0% of later input in live tests across 9 model families.
  • Built-in tool definitions are 48.0% smaller than in 0.0.16, and a new session's first request is 32.4% smaller.
  • libfx's native addons are 10.1% to 12.9% smaller than in 0.0.16.
  • The Linux arm64 and x86_64 binaries are 36.6% and 31.1% smaller than in 0.0.16.
  • libfx requests now use AI Gateway's prompt cache.
  • A shell result for a command with no output is about 70 characters instead of about 340.
  • Skill descriptions in the toolbox menu stop at 200 characters, down from 1,024.
  • Earlier turns go back to the model unchanged, so the prompt cache keeps covering them.
  • An attached file's folder rules arrive as a note, so the prompt cache keeps matching.
  • Conversations with many images drop the oldest in blocks, so the prompt cache stays valid longer.
  • Amazon Nova models now use the prompt cache.
  • Long sessions stay responsive, with the last turn of a 2,400-tool-call session taking 3 s instead of 11 s.
  • Full detail in Ctrl+O shows each model call with its model, time, tokens, and call ID.
  • /trace writes JSON Lines instead of a Markdown report.
  • /trace includes subagents and keeps failed calls whole.
  • Questions from the model render as Markdown.
  • fx doctor now tests your credential and connects to each approved MCP server.
  • Requests keep retrying through network drops, gateway errors, and rate limits until they succeed or you press Esc.
  • Steering moves a running shell command to the background, so your message reaches the model within about a second.
  • Cancelling a turn also stops the commands that steering moved to the background.
  • A command still running in the background shows Still running instead of Ran.
  • A waiting shell command uses under 1% of a core instead of about 3%.
  • More gateway errors are retried, including 520, 524, and 529.
  • The status line shows why a request is retrying and when it tries again.
  • fx resends a request within about a second when the Mac changes networks, instead of after 45 seconds.
  • fx finds a dead connection within about 5 seconds after the network returns, instead of waiting up to 10 minutes.
  • On a dead network, fx ask and fx acp start in about 2 seconds instead of 20 to 30.
  • A brief Mac sleep no longer restarts a healthy answer.
  • Errors show in plain words instead of error names.
  • A rejected or filtered request waits for you to send it again with Enter instead of ending the turn.
  • Running out of credits offers to retry, switch models, or cancel.
  • An expired sign-in opens the sign-in and then sends the request again.
  • A restarted answer replaces the broken one instead of following it.
  • A reply cut off before any text is sent again once instead of ending the turn.
  • fx asks the model to continue once when it stops without a reply.
  • After a tool call it can't read, fx tells the model and asks it again at once, with tools available.
  • A compaction the provider refuses as too long is sent again as a smaller request.
  • A command that fails again after an edit goes back to the model instead of ending the turn.
  • Shell results tell the model when a command has printed nothing for 2 minutes.
  • edit_file errors show the lines the model most likely meant.
  • MCP arguments sent as text go to the server as the type its schema names.
  • Tool arguments a model writes as JSON text, such as question options and toolbox runs, are read like objects.
  • libfx sends host tools the argument types their schemas name.
  • Menus and pickers match letters in order, so /mdl finds /model.
  • The model picker labels Fast models and Fast mode apart and lists newer GPT versions first.
  • Ordered lists in replies align their numbers.
  • Tool group headers say what the group did, such as Edited a file, ran commands, instead of a tool-call count.
  • Skill discovery warnings no longer show at startup.
  • MCP config changes made outside a session load at the next prompt.
  • fx mcp add signs in to a new server that needs it, and --no-login skips that.
  • Mistyped commands, flags, and settings suggest the closest match.
  • Every command accepts flags as --flag=value.
  • fx pr, fx issue, and fx acp take the same run flags as fx ask.
  • fx turns off color when NO_COLOR is set or TERM=dumb.
  • The resume command prints on its own line at exit.
  • fx sessions and fx sessions --live print aligned tables.
  • fx retries a sign-out that the provider did not confirm.
  • A new libfx session skips 3 empty storage reads on its first prompt.
  • libfx starts preparing a prompt's run while the queue confirms the prompt.

Bug Fixes

  • Closing the window during a request no longer cancels it, so fx -c sends it again.
  • fx ask and ACP no longer wait 15 seconds for a stuck session title.
  • Esc and Ctrl-C stop a request at once during a stuck DNS lookup.
  • Cancelling a turn ends a running tty command.
  • A cancelled provider switch ends at once.
  • fx uses a login another fx just refreshed instead of asking you to sign in again.
  • Several fx processes that get a 401 together refresh a shared fx, Codex, or Grok login once.
  • A cancel no longer cuts off a sign-in refresh, so other fx processes stay signed in.
  • A new Vercel sign-in starts on the team you used before.
  • A request waiting for a sign-in resends with the team you picked.
  • Claude Fable 5.1 and later no longer reject a continued session with "Invalid signature in thinking block".
  • Starting many HTTP MCP servers no longer crashes fx.
  • MCP logins save while other fx processes use the Keychain.
  • MCP logout says when a login couldn't be removed or revoked.
  • A named pipe in place of ~/.fx/mcp.json no longer hangs fx.
  • fx says when it can't write its output and stops quietly when piped into head.
  • Startup-file wrappers such as Socket Firewall's npm work in fx.
  • Large shell snapshots restore in full with macOS's bash 3.2.
  • A startup file that defines builtin no longer breaks fx's shell.
  • A backslash before a newline continues the line, as in the shell.
  • A daemon started by a command keeps running after the command ends.
  • Ctrl+Backspace deletes a word, and repeated Cmd+Backspace deletes across lines.
  • /rename, /image, and /shell picked from the slash menu wait for their argument.
  • Wide edit rows keep their +N / -M counts.
  • Resizing or resuming a session no longer drops turns from the screen, and long tool-heavy turns leave no blank rows in scrollback.
  • Theme switches and bracketed paste keep working after a terminal reset or Ctrl+Z.
  • Ctrl+G updates fx without moving or repeating the conversation.
  • A damaged session says it is unreadable and points to fx doctor.
  • New session IDs never start with a dash.
  • fx says why a terminal recording could not start.
  • libfx no longer sends the same Codex request twice.
  • /credits shows the AI Gateway amount used.
  • Skills load from a linked FX_HOME, with multi-line descriptions and nested keys.
  • fx keeps running when herdr doesn't answer.
  • A relative HOME no longer crashes fx mcp.
  • MCP calls wait up to 5 minutes, up from 60 seconds.
  • Common tool-call mistakes from models no longer fail the call.
  • A blank custom answer no longer submits a question.
  • Quitting while the transcript loads no longer crashes fx.
  • Deeply nested provider events no longer crash fx.
  • Fixed a data race in parallel grep_files searches.

Security

  • Control codes in model replies, prompts, and MCP tool errors show as text instead of acting on the terminal.
  • fx logout codex revokes the login at OpenAI.
  • Shell commands no longer inherit web_fetch connections.
  • fx doctor fix lines can't run commands from a project's .mcp.json.
  • When other users can write to ~/.fx, fx removes their access and still loads your settings.
  • Control characters in error messages show as visible escapes.

Breaking Changes

  • Auto mode now holds an action only for prompt injection or harm to your machine.
  • The subagent tool's one-off run action is gone.
  • fx doctor exits 1 when a check fails.
  • Usage errors exit with status 2 instead of 1.
  • statusLine is now status_line, and notifications is now sound.
  • fx setup, /setup, fx mcp --profile, and /allowlist user and local warn now and stop working in fx 0.0.18.
  • Old resume spellings, such as fx resume and --resume-id, are removed.
  • fx ask --verbose and FLASH_SYNC_UPDATES are removed; use FX_SYNC_UPDATES.
  • Esc during a retry ends the turn, and the separate Esc-to-pause is gone.
  • fx ask --json and ACP report recovery as state, attempt, need, and message.