Release history.

October 8, 2026

• Bug fixes and stability improvements. Fixes and polish across the app, found while running Fermata on itself every day.
• Request Changes commits for you. When a Request Changes turn ends, Fermata commits the changes, and pushes them when the Piece has a pull request. Choose Commit automatically or Ask me first in Settings › Pieces › Gates, for all projects or one, or per Piece in its Flow Configuration and from the gear on the Loop board. A failed commit or push still stops and asks.
• Extra folders agents can read. In Settings › Projects, add folders outside the repository, such as a prototype or a design system, that every agent in that project can read. Agents can never change them.
• Quieter blocked steps. When a step’s read-only rules block a command and the agent finds another way, the transcript shows a quiet note instead of a red error.
• Works on gateways without Haiku. Commit messages and pull request descriptions fall back to the Piece’s own model when your gateway does not offer Haiku.

October 8, 2026

• Fixing bugs and stability issues. This release is mostly fixes found while running Fermata on itself every day, plus a few small changes listed below.
• Models per phase. Settings › Models & Effort has a Models per phase panel: choose the Spec, Strategy, Agents and Review models and effort that new Pieces start with, for all projects or one. The new gear on the Loop board opens the same defaults for the project you are in. A finished Piece’s configuration is now read-only, shown as it ran.
• Request Changes, with the New Session composer. Request Changes opens the same composer as an expanded New Session, with a model and effort picker. The run uses the model you pick; it no longer falls back to a fast model on its own.
• Composers behave the same everywhere. Pasting an image or an image file attaches it instead of pasting its path. ⌘↩ submits in every sheet and input; in sheets, Return adds a new line. Tab and Shift-Tab move focus.
• Readable PR reviews on the Done screen. Review comments from bots like CodeRabbit are cleaned up, with collapsible sections and a Show full review link. Acceptance criteria chips open the criterion and how to check it, and the summary names them.
• PR status you can trust. The PR card shows how long ago it was checked and has a Refresh button. A failed refresh says the status is stale instead of keeping an old state, so a merged PR no longer shows as open.
• Agents find your tools. Agents started from Fermata opened in the Finder now get your login shell’s PATH, so node, pnpm and tools installed through Volta or Homebrew are found.
• Fixes. Haiku 5.5 costs are priced correctly. The Effort menu no longer wraps in narrow composer headers.

October 6, 2026

• Settings, redesigned. Settings is grouped by what each control changes, has a search field, and edits a project’s own values inline with a scope switcher at the top. Every description now matches what the setting actually does.
• One model picker. A single grouped, readable picker everywhere you choose a model: Settings, a Piece’s flow configuration, and the compact chip. You can also switch the model of a running session without restarting it.
• The real state of your pull request. The Done screen shows the PR’s checks, reviews, and anything blocking the merge, instead of a bare Merge button. After a merge, Fermata offers to mark the Piece done.
• Sessions end when their work does. Request Changes rounds, conflict resolutions, and code reviews close once their job is finished, and the active count in the toolbar now says what is actually alive. On a Piece without a pull request, Commit ends a feedback round, and commits get a message that describes the change.
• A Pieces-first sidebar. The sidebar leads with your Pieces, lets you pin the ones you are watching, and shows when each was last active.
• Drive Fermata from Claude Code. The Fermata MCP server is no longer an Edge feature. Turn it on in Settings › MCP, and any Claude Code session can file Pieces from a Linear or GitHub issue, change their lane, request changes, read the work review, and open pull requests. It registers itself in every Claude config folder your projects use, and it is off until you turn it on.
• Review the plan at a glance (Edge). At the strategy gate, a Document | Visual toggle shows the plan as a timeline of units, with the files each one touches and the acceptance criteria it serves. A criterion no unit covers shows in amber before anything runs.
• Report card and Validation plan (Edge). A finished Piece gets a report card with findings, cost by phase, and the time its agents actually worked, plus a validation plan you can mark off by hand. Each has its own switch in Settings › Edge, and new Edge switches start off when you update.
• Streams and Tracks (Edge preview). Point a Stream at a Linear or GitHub query, and matching issues arrive as Tracks in their own column on the Loop board. Review what the agent will read, then approve a Track into a Piece with its lane and model. Choose how often Fermata checks, from every minute to every 30 minutes.
• Fixes. Task agents keep the 1M context window you picked. Typing your default Claude config folder no longer breaks sign-in. Skipped tasks no longer count as progress at the advance gate. Request Changes stays in sync when you commit from outside Fermata, and a diverged branch can be recovered from the commit gate. Released Auto Pieces no longer show Armed while their spec is formatted. Fermata checks a folder is a repository before adding it as a project.

September 28, 2026

• Pieces. What was an initiative is now a Piece, everywhere in the app. A Piece runs spec, strategy, build, review, and PR, and plan approval now lives inside Strategy so you approve the plan where you read it.
• Chat-first spec interview. Describe the work in a conversation and the interviewer asks structured questions as cards, single or multiple choice, with room to type your own answer alongside a pick. It explores your project before it asks, can reach your project’s MCP servers, and hands a pasted document over as an attachment instead of burying it in the draft.
• The Loop. A board that runs Pieces on their own, with a column for each phase from Backlog to Done, ordered so whatever needs you sits on top. Press Play to work through a backlog of drafts, cap how many agents run at once across Pieces, and set per-phase approval gates for anything you want to sign off by hand.
• Home. The default surface for each project, with plan usage and the Pieces and sessions that need your attention.
• One window per project. Each window binds to a project with its own settings. Clone a repository straight from GitHub to start a new one.
• Analytics. Sessions, approvals, interventions, and cost per project, with a per-phase cost breakdown on every Piece. Pro adds spend insights split by phase, by Piece, and by rework.
• Models and effort. Pick the model and effort level per session, per Piece, and at spec approval. The catalog includes Claude Opus 5.5 and Fable 5.1, plus Bedrock and Vertex model awareness, and mirrors the curated model list from your Claude settings.
• Recover from failed agents. An agent that hits a resumable error can Continue where it stopped, stopped build runs can be resumed, and an agent whose work already landed can be marked completed.
• Edge features. Switch each one on or off in its own Settings pane: code review for Pieces, starting a Piece from a Linear or GitHub issue, the Fermata MCP server that lets an outside Claude Code drive your Pieces, and the new diff viewer.
• Attachments everywhere. Paste or drop images and files into every composer, including the New Piece sheet and question answers.
• Settings, rebuilt. A list and detail layout, app-wide Text Size, Keep Mac Awake while agents run, and a Welcome Tour you can replay from Settings.
• Crash and freeze reports. Fermata asks at first run before sending anonymous analytics, and crash and freeze reports can be switched off in Settings.
• Stability. Long transcripts, spec interviews, the Review diff, and the hidden console no longer freeze the app, and a pending approval can no longer hang a session for good.