{"version":1,"updated":"2026-09-24","entries":[{"title":"Your tools move with the work","date":"2026-09-24","category":"Workspace","summary":"Code, Changes, Review, and Worktrees now open only when you need them. Resize them, expand them, or keep two open together while the conversation stays at the center.","slug":"workspace-tools","url":"https://www.winskel.com/changelog/workspace-tools","details":"## What changed\n\nThe fixed panel beside the conversation is gone. Four tools open from the title bar when you need them, and the rest of the time the conversation has the width.\n\n- Code holds Files and Terminal.\n- Changes shows the real Git diff for the objective.\n- Review holds Tasks, Checks, and Result.\n- Worktrees separates active working copies from results kept on their branches.\n\n## What you can do now\n\n- Keep up to two tools open together, side by side when the window has room or stacked when it does not.\n- Resize the conversation edge and the split between two tools, expand a tool, and restore it.\n- Open a tool with Command 1 to 4.\n- Open a real login shell in Code. Terminal uses the same shell sessions as before, so hiding the pane does not end a running shell.\n- In a narrow window, work with one tool at a time instead of a squeezed layout.\n\n## Limits\n\n- Winskel is not a code editor. Files and Changes are for reading and reviewing the work.\n- Opening a third tool replaces the oldest one.\n- Terminal opens in the trusted project root, not inside a run's worktree."},{"title":"See what actually changed","date":"2026-09-24","category":"Review","summary":"Winskel now brings the real files, Git changes, observed checks, executed models, and final result back into one review flow, so you can inspect the work instead of trusting an agent summary.","slug":"see-what-changed","url":"https://www.winskel.com/changelog/see-what-changed","details":"## What changed\n\n- Files lists the objective's real files as Added, Modified, Deleted, or Renamed, grouped by folder, with links into the diff.\n- Changes is a line-numbered Git diff. Each hunk has Ask Winskel, so you can ask about exactly that change.\n- Checks read Passed, Failed, Not verified with a reason, Not required, or Not run yet, with the evidence Winskel observed.\n- Result brings the work, files, checks, and the models that actually executed into one coordinated result, and shows only what still needs your attention.\n- Model replies render as clean, safe Markdown.\n\n## Why it matters\n\nWhen several models work on one objective, their summaries are not evidence. Review is built from the files, the Git history, and the commands Winskel watched run.\n\n## Limits\n\n- An agent saying tests passed does not count as a check. When Winskel did not observe it, the check says Not verified.\n- A GitHub pull request workflow is not part of this release."},{"title":"More control without more babysitting","date":"2026-09-23","category":"Control","summary":"You can choose how much autonomy Winskel gets, attach context, stop work, rerun it, and step in only when the work genuinely needs you.","slug":"safe-autonomy","url":"https://www.winskel.com/changelog/safe-autonomy","details":"## What changed\n\n- Choose Ask Me, Smart Auto, or Autonomous for each conversation, right where you write the objective. When you do not choose, your project and account defaults apply.\n- Attach screenshots and images to an objective. The agents working on it receive them.\n- Stop active work, including work waiting in Needs You, without erasing the objective or its history.\n- Edit & Rerun a prompt as a new objective linked to the original.\n- Revert a finished objective's applied changes with a new inverse commit, so history is preserved.\n- Needs You is for real decisions. A failure is shown as a failure, not turned into a question.\n\n## Boundaries that hold in every mode\n\nForbidden actions such as a hard reset, a forced clean, a force push, privilege escalation, and targets outside the worktree are blocked. Unknown commands and complex shell compositions still come to you. No mode widens a provider's own sandbox.\n\n## Limits\n\n- Revert refuses when later work changed the same code, when the result branch is checked out, or when Winskel cannot prove the result can be reversed safely.\n- Code that a project's own tests or build scripts run can still do whatever that project does, within the provider's sandbox."},{"title":"Winskel chooses the model team","date":"2026-09-22","category":"Orchestration","summary":"One objective can now become connected work across Claude Code and Codex, with Winskel choosing the model for each task instead of making you route the work manually.","slug":"multi-model-routing","url":"https://www.winskel.com/changelog/multi-model-routing","details":"## What changed\n\nWinskel now classifies the work inside an objective and routes each meaningful task to a model that fits its type and difficulty. Connected parts of one objective can run on different model families, for example architecture on one model, implementation on another, and an independent review on a third.\n\n- Routine work stays on Tier 1 models: Claude Sonnet 5 or GPT-5.6 Terra.\n- Hard, high-risk, or broad multi-file work uses Tier 2 models: Claude Opus 5 or GPT-5.6 Sol.\n- The hardest architecture work has a dedicated route to Claude Fable 5.1, and the hardest terminal and DevOps work has a dedicated route to GPT-6 Astra.\n- Tiny tasks stay as one Tier 1 step instead of being sent to a frontier model.\n\n## Why it matters\n\nYou no longer have to decide which model should own each piece of the job, open each agent yourself, and carry context between them. You give Winskel the objective and it builds the model team.\n\n## What you can do now\n\n- Give Winskel one objective and let it keep simple work simple or split larger work into connected tasks.\n- See the model assigned to each task.\n- Pin an exact model when you want explicit control. A pin holds through delayed starts and retries. If the pinned model cannot run, Winskel stops and says so rather than quietly substituting another.\n\n## Limits\n\n- Provider availability and your own Claude Code and Codex plan limits still apply. When a provider is known to be unavailable, Winskel uses only fallbacks its routing policy allows.\n- If a Tier 1 attempt fails verification, Winskel may escalate once to the matching Tier 2 model inside the existing retry budget. It does not keep escalating.\n- Routing follows a fixed, versioned policy. It does not yet learn from past outcomes."},{"title":"Winskel becomes a local Mac workspace","date":"2026-09-21","category":"Platform","summary":"Winskel became a real Mac application for running Claude Code and Codex against repositories you trust, with the work and the tools around it in one window.","slug":"local-workspace","url":"https://www.winskel.com/changelog/local-workspace","details":"## What changed\n\nWinskel now runs as a packaged macOS application. It uses the Claude Code and Codex sign ins already on your Mac and starts agents only in repositories you have trusted on that machine.\n\n- Chats keep related objectives together inside a project, and working surfaces open as tabs.\n- A real login shell opens in the trusted repository root.\n- Current Codex usage is visible without starting a run.\n\n## Why it matters\n\nCoordinating agents needs a home on the machine where your code, your repositories, and your provider sign ins already live. This release made that home a real application instead of a browser page and a separate command line tool.\n\n## Limits\n\n- Winskel is a macOS Alpha. Your Claude Code and Codex plans, and their limits, stay with each provider."}]}