v0.1.0

Release notes for the terminal client and shared agent runtime

Upgrading from v0.0.12 or earlier

Flywheel v0.0.12 and earlier cannot update to this release by themselves, because this release's archives are larger than those versions accept. They show no error and stay on their current version. Run the installer once. It keeps your sessions, logins, and settings, and later releases then update automatically again.

macOS and Linux:

curl -fsSL https://github.com/wsauret/flywheel/releases/latest/download/install.sh | bash

Windows (PowerShell):

irm https://github.com/wsauret/flywheel/releases/latest/download/install.ps1 | iex

This release turns workflows into the main way Flywheel runs larger jobs. Agents can write, launch, and steer workflows, run steps in parallel, and message and wait on each other, with live progress in the terminal. It also bundles a Claude engine for the Anthropic Subscription provider and fixes crash recovery, startup, and leftover tool-process cleanup.

Added

• Run independent workflow steps at the same time, with explicit phases and step order, and fan a chain of steps out over a list and join the results
• Loop finder rounds until a round finds nothing new, with votes and quorum joins that merge duplicate findings and report what each join dropped
• Give each workflow step only the earlier results it names with reads, keep a run going when one step fails, and resume a failed run without redoing finished steps
• Launch a workflow from its content, a file path, or structured input, and let agents write, launch, and resume their own workflow files, with refused files listing every problem and its fix
• Start workflows without a confirmation question by default
• Run workflow steps on the launching chat's model and effort unless a step or a Workflow Agent setting chooses another, and fix the model and effort of named steps at launch
• Flywheel changed agent prompting so a chat delegates substantive work to a subagent, to parallel subagents, or to a workflow only when a later agent reads an earlier agent's output
• Show the workflow card in the terminal as a live tree with the plan, step strip, progress counts, lanes, loop round and stop reason, and each running agent's current activity as an openable row, with moves between sibling agents and lanes
• Show a workflow as one background row with its progress, and stop or interrupt the whole workflow from that row
• Interrupt and resume individual roster agents, and send guidance to some lanes of a running workflow without reaching others, with changes sent to a running workflow reaching every later step
• Let agents message their parent, siblings, children, and other open chats, ask their parent instead of interrupting you, wait for replies, and resume waiting or stopped children in their own sessions
• Let agents check the progress of and interrupt the subagents and workflows they started, and show when an agent interrupted one
• Show each subagent's card with the same live state as the subagent's own session
• Give subagents larger turn budgets, and keep paused workflows resuming under them
• Let agents wait on shell commands, subagents, and workflows for a stated estimate, four minutes by default, with the wait ending on completion or expiry and the work left running
• Serve the Anthropic Agent SDK provider through a Claude engine bundled in every release archive, so Claude Code does not need to be installed and the bundled engine never reads or changes your own Claude Code
• Log in to each Claude account once from Authentication, keep saved model choices and paid-use decisions for that account, and ask you to log in again instead of ending refused Claude runs
• Show each Claude account's plan, its five-hour and seven-day usage, and its status, with paid use off unless it is allowed for that account in Provider Settings, and use metered billing for usage-only seats
• Report the engine version and each account's status in /doctor
• Accept self-update archives up to 512 MiB, with installations at v0.0.12 or earlier needing one reinstall with the bootstrap installer
• Run Opus 5.5, Sonnet 5.5, and Fable 5.1 chats on instructions written for those models
• Show shell commands as plain sentences such as "Searched rows in models; Read parsed.sql" instead of raw commands, with mutating commands shown as running edits that name the files their captured changes touched
• Group shell, memory, skill, papercut, experiment, session note, and child-session calls under the Ran heading
• Run flywheel --headless as a normal chat on standard input and output with durable history, attach an existing chat with --session, and return after the handoff with --one-shot
• Print durable request evidence as JSON with flywheel events evidence --session, including the workflow step that sent each request
• Set the model and effort of named roles for one process with FLYWHEEL_ROLES without changing the saved configuration
• Rank available models by quality, speed, and cost, let workflow agents inspect the same model list, and search the model picker without changing another model's remembered effort
• Ask temporary contextual questions with /btw
• Show each startup phase as a counted step on one progress bar
• Keep the reading position across recovery and resize in the terminal
• Enter in the slash popup runs the highlighted command
• Edit a session's tab title and description separately
• List only chats in session lists and tabs, with workflows and subagents opened from their parent chat and deleting a chat removing the work it owns
• Show cross-session messages as compact rows that expand, with a direct parent's message fully visible
• Remove and enable MCP servers by global or project scope
• Connect MCP servers that need a pre-registered OAuth client with flywheel mcp add --client-id
• Read Python files by symbol, file map, or summary
• Set prompt-cache retention for an API provider, and checkpoint large idle chats before their prompt cache expires
• Delete session scratch folders that no agent has written to for 3 days

Fixed

• Stop tool commands that outlive a crash at the next start on macOS and Linux, and prove each process's identity before signaling it
• Keep a chat that a crash cut idle when its tab opens, with no new turn started and a Resume option when no work is left
• Stop recovery hangs when a chat and its subagent both need recovery after a crash
• Start and accept input at once after a crash or restart, without waiting to recover interrupted sessions
• Keep a subagent running while it waits on a background command
• Restore every shared setting on Reset to Defaults, including the theme and Desktop notifications
• Report a refused auth profile change with its reason instead of reporting success
• Start a chat when text or attachments are submitted with no session open, instead of reporting an unknown command
• Restore Git AI attribution in projects that are Git repositories
• Report a provider failure with its own error, and name what failed when every provider for a model fails
• Stop Claude subscription and OpenAI sessions from failing on an "input.content is not supported" error
• Report an unknown headless model as model_unavailable
• Stop the terminal from being left in a bad state when its window closes
• Stop a chat delete from waking the chat into a model request
• Stop large MCP results from stalling other sessions
• Stop a Linux bus error when a second process opens a project's database
• Stop Copilot requests from waiting on a usage check each time
• Remove read-only scratch folders correctly

Changed

• Remove named workflow launch, the /workflow: commands, and the /qa and /fly.* slash commands, so workflows launch by content or by path, and a workflow file in a workflow folder is ignored with a warning
• Remove the separate Ad hoc workflows setting, since agents now use workflows when they fit
• Remove the headless --description and --workflow launch flags, and use flywheel --headless for chats
• Rename the workflow model settings to Workflow Agent and Workflow briefs, with existing settings converted automatically

Performance

• Cut about 14 MB from each tool guardian and command supervisor, so the account service with no project open holds about 20 MB less and each open project about 13 MB less
• Start the account service on Windows while the interface loads, so the prompt is ready sooner
• Open the tab you are looking at first after a restart, with the other tabs loading afterward
• Let new chats and subagents reuse the cached tools, system prompt, and workflow step prefix, which cuts the cost and first-request time of each run
• Bound store and history reads so long sessions and large workflow histories stay fast
• Make Code Mode batches about twice as fast, and keep the interface responsive while tools return very large results