Changelog — shot2code

Published Windows builds of shot2code, newest first. Tags in this hub are product-prefixed (shot2code-vX.Y.Z); the upstream project tags the same build as vX.Y.Z.

The authoritative list of published builds is the releases page, which is generated from the release feed; this file summarises each one.

The full engineering changelog lives with the code: CHANGELOG.md. Versions before 0.3.0 were not tracked in a changelog, so the earlier builds mirrored on the releases page have downloads and notes but no entry here.

[0.6.1] — 2026-10-05

Download · hub release · upstream notes

The current stable release adds bounded research/design inputs, clearer tool consent and a complete local-history browser while keeping imported code and third-party content behind the existing never-execute boundaries.

Added

• Local Ollama quick setup. The existing Copilot SDK BYOK card can configure http://localhost:11434/v1 as a credential-free OpenAI-compatible localhost endpoint. Local inference needs no paid API, but shot2code does not install Ollama or a model: users supply hardware and choose a model whose own licence and capabilities include image input and tool calling.
• Separately consented bounded page reading. The provider-neutral read_web_page tool accepts public HTML, text, Markdown and JSON through public-only DNS pinning, redirect, MIME, byte, text, timeout and call budgets. Copilot's runtime-owned built-in web_fetch remains blocked because its full result reaches the model before shot2code can inspect or bound it.
• Full history across local projects. A searchable browser can load every saved project and expand its complete version, option, prompt, model, message, attachment, activity, status, timing, error and retry/branch ancestry.
• Built Storybook context. Import a built folder, selected JSON files, ZIP or public HTTPS root. Only fixed index.json, manifests/components.json and manifests/docs.json metadata is parsed; stories, bundles, addons, loaders, play functions and iframe.html are never loaded or executed.
• Opt-in Iconify search. The canonical search_icons tool uses only https://api.iconify.design, accepts the service's currently keyless public access without promising a permanent quota, filters automatic results to a permissive SPDX allowlist, sanitizes SVG, stores it locally and embeds source, licence and retrieval provenance plus a trademark warning.
• Chat Tools inventory. The composer reports project editing, preview verification, web search, page reading, public-domain photos, generated images, icons, MCP write scope and enabled Skills without silently enabling a paid, external, trusted or write-capable feature.
• Full-page design evidence. URL inspection scrolls bounded lazy content and reports full-page desktop/tablet/mobile dimensions plus explicit blank and truncation state. Figma can preview rendered frames before a model call and reuse the same evidence for generation.

Changed

• The GitHub tab now distinguishes a fully local open from an optional model-assisted first refinement, exposes the edit-model lineup and design system for that refinement, and preserves the detected repository stack.
• Stitch conversion now exposes a labelled model picker in the tab, keeps the control visible while provider setup is incomplete, and offers only image-capable models for the generated Stitch screenshot.
• Review now combines deterministic source findings with bounded per-frame runtime evidence, categories, filtered select-all, per-frame failure isolation, explicit healthy/stale/partial coverage and schema-v2 JSON reports. It remains an automated aid, not WCAG certification.
• A non-retryable image-provider failure trips a per-generation circuit breaker for credential, billing, quota, permission, model or configuration problems, preventing repeated provider calls and directing the run to licensed photos, localized icons, extracted assets, CSS or SVG instead.

Fixed

• Screenshot preview now returns bounded console/page diagnostics and nearly blank-render metadata so a model can repair a mount failure rather than treating an empty screenshot as success.
• The Electron shell inspects a top-level renderer after load, reloads one blank window once, then shows a static recovery screen while preserving local projects if the renderer remains empty.
• Optional discovery waits five seconds after core startup, and Chromium warming plus Copilot authentication/model discovery run on isolated worker event loops, so first-run process scanning cannot freeze Health, History or Settings or create a false backend timeout.
• Full history now includes the separately persisted active draft, so a first local HTML/project import remains one version with its saved options before and after a renderer restart.

Documentation

• Public guides, product copy and generated native documentation now describe the same v0.6.1 feature, consent, licensing and safety boundaries.
• The existing screenshot presentation work keeps current captures full-width, centered and uncropped; planned new captures remain versionless and are not referenced publicly until their PNG files exist.

[0.6.0] — 2026-10-04

Download · hub release · upstream notes

This release expanded shot2code from screenshot generation into a local-first design-source workspace.

Added

• Figma REST import now preserves original image fills and export-marked nodes beside rendered frames as bounded, reusable local assets.
• Google Stitch defaults to Stitch only: localized HTML, screenshot, images, stylesheets, nested CSS assets/fonts, srcset and available DESIGN.md open directly without a second model request. Conversion is explicit.
• The URL tab can inspect a public website through a bounded public-only Chromium proxy, produce desktop/tablet/mobile screenshots and export a browser-computed DESIGN.md.
• A dedicated GitHub tab opens public repositories without a token or private repositories with a separate repository-limited Contents: read token, preserving bounded image assets without executing repository code.
• PNG, JPEG and WebP screenshots can be pasted directly into refinement Chat, with duplicate, count and size controls.
• A new capture-to-code application identity ships across the EXE, NSIS, MSI, portable build, shortcuts, sidebar and browser favicons.
• Public documentation includes a current packaged-app screenshot and detailed usage guidance for every input tab: Upload, URL, Text, Import, Figma, GitHub and Stitch.

Changed

• Imported Figma, Stitch and GitHub assets now use restart-safe local references across History, retries and later chat, and binary project files export as decoded bytes.
• Preview and CodePen unwrap real brace-wrapped public URLs and replace unresolved IMG.* pseudo references with a safe placeholder; generation prompts explicitly forbid those malformed forms.
• Public-site, design-tool and repository text is labelled untrusted evidence. Stitch and website networking uses pinned public-address validation, bounded redirects/downloads and fail-closed behavior.
• MCP and Agent Skills retain enabled-plus-trusted, read-only-by-default and disabled-by-default safety rules and were reverified across backend, renderer and packaged-app surfaces.

[0.5.2] — 2026-09-28

Download · hub release · upstream notes

The previous stable release. It includes the v0.5.0 feature set plus the fixes and additions below.

Added

• Free image search — shot2code sends no key of its own. A new canonical search_free_images tool finds real, ready-to-use photographs through Openverse's public image API, and is offered to every runtime. It needs no credential, so it works on a machine with no Replicate, Cloudflare or OpenAI-compatible configuration at all, and configuring one of those never switches it off. It is a separate tool from generate_images, never a silent fallback inside it: the model picks whichever the design needs, and the activity feed says "Found 3 free images" rather than "Generated". Off by default, because the query text leaves the machine.

• Free image results are restricted to CC0 and Public Domain Mark, so an exported project cannot quietly inherit an attribution, share-alike or non-commercial obligation shot2code could not enforce on the user's behalf. The licence is re-checked locally rather than trusting the search server's filter, and a result missing its source page or licence URL is dropped because it could not be verified. Title, creator, source page, provider, licence name and licence URL travel with every image, along with an explicit warning that Openverse aggregates other platforms' metadata and should be verified before commercial use.

• Found images are downloaded and served locally, so generated and exported projects never hotlink somebody else's server. Every URL is checked first: http(s) only, DNS-resolved with private, loopback, link-local and cloud metadata addresses refused (a mixed answer is rejected outright as a rebinding attempt), redirects re-validated rather than followed, declared content type required to match the sniffed bytes within an image allowlist, and byte and decoded-pixel ceilings enforced. Web image search is deliberately not used: a picture on a web page grants no reuse right.

• Optional image-generation providers. Placeholder images can now be generated by Cloudflare Workers AI (account ID + API token) or by any OpenAI-compatible image endpoint (base URL, plus an API key unless it runs on localhost), in addition to Replicate. The alternatives are additive: Replicate stays the default and the existing Replicate, OpenAI, Anthropic and Gemini keys are never read, replaced or re-routed by the choice. Cloudflare's daily allocation is quoted from Cloudflare's own pricing page — 10,000 Neurons a day on both the Workers Free and Workers Paid plans, resetting at 00:00 UTC, $0.011 per 1,000 Neurons above it — attributed to the user's Cloudflare account, dated, and marked able to change, never as a guarantee. CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN supply the same credentials to headless runs.

• Both validated Replicate image models are now selectable — prunaai/z-image-turbo (default) and black-forest-labs/flux-2-klein-4b — each with cost wording that names who bills, quotes the provider's own published figure with the date it was read, links that provider's pricing page, and says the terms can change. No provider is described as permanently free anywhere in the product.

• Checked custom Replicate models. Another Replicate model can be used once it passes a schema check: POST /api/image-models/validate reads that model's own published OpenAPI schema and accepts it only if it really declares a string prompt input and an image output and requires nothing else. Replicate models do not share one input schema, so no model outside the curated list is claimed to be compatible without that evidence.

• GET /api/image-models exposes the curated catalog — providers, defaults, capabilities and cost notes — as a pure read that carries no credential.

• Provider-neutral web search. A shot2code-owned search_web tool is now offered to every runtime — native OpenAI, Anthropic and Gemini, GitHub Copilot and Copilot SDK BYOK — instead of only to Copilot SDK sessions. Tavily is the default provider (its published plan is 1,000 API credits a month with no credit card, plus a documented keyless trial that needs no account); Exa is available as an alternative and always needs a key. Those allowances are the providers' own and can change — shot2code links their pricing pages rather than promising anything. Off by default; Settings states exactly what leaves the device before it can be switched on, and offers a test-connection button.

• Web search requests go only to the provider's fixed HTTPS endpoint, with an explicit timeout, no redirects and no raw page content. Each search returns at most 5 results with bounded titles, snippets and total size, prefixed with an untrusted-content warning, and is capped at 3 searches per turn and 10 per generation. Search-provider keys are backend-only and are excluded from history, commit snapshots and tool arguments.

• Copilot's built-in web_search remains available for people who prefer it and have the entitlement. When the canonical search is configured and usable, it is enabled instead of the built-in, so a model can never take the unbounded route past the budgets and the domain allowlist.

• Google Stitch generation now reports real SDK phases — connecting, project creation, screen generation, output download and import — with elapsed time, an accessible live status and bounded error messages. The SDK client now uses the same ten-minute generation timeout the UI promises.

• Dedicated Figma and Stitch input destinations sit beside Upload, URL, Text and Import. Figma uses its documented REST/PAT frame-rendering path; Stitch uses the bundled experimental SDK and preserves the selected shot2code stack/model routing.

• The Preview toolbar now includes accessible 25%–200% desktop canvas zoom, Fit and 100% controls. Magnified canvases pan without blocking iframe interaction.

• Help now includes an in-app Bug, Feature request and Feedback form with authenticated gh submission plus copy, Markdown-download and prefilled browser fallbacks.

Changed

• Copilot's built-in web_fetch is assessed and deliberately not offered. It is a real runtime built-in and ToolSet.add_builtin("web_fetch") would reach it, but a built-in's result is produced inside the Copilot runtime and handed to the model by the runtime — the SDK only notifies the application afterwards. There is therefore no point at which shot2code could cap that text, mark it as untrusted, or count it against a budget, and web_fetch returns a whole page. The SDK also accepts no URL allowlist at session creation. Settings now explains this next to the Copilot web-search switch, as its own unconditional disclosure rather than something coupled to that switch, and points at the canonical search_web tool, whose results are capped, labelled and budgeted.
• The deny-by-default Copilot permission handler is now installed on every session, subscription and BYOK, instead of only on sessions that configured MCP servers. Without a handler the runtime fell back to its own default policy, so a run using a network-capable built-in had no shot2code-owned answer to "may this URL be fetched?". Any URL the runtime asks to open is denied with a reason naming the host and path — never the query string, which can carry a token. MCP approvals are unchanged.
• A ToolSet that names a blocked built-in — directly or through a builtin:* wildcard — now fails when the session is built, rather than silently granting a capability whose output cannot be bounded.
• Image results are normalized in one place. Remote URLs, raw bytes, base64 and data: payloads all pass through one adapter that either keeps a public URL or writes the image to the served local asset directory, and refuses anything unsafe (file:, private or loopback addresses that are not our own assets, oversized payloads). A locally served image is handed to the model as bytes, because such a URL is not fetchable by Anthropic, OpenAI or Gemini.
• Background removal is stated as Replicate-only and behaves that way: the remove_backgrounds tool is simply not offered without a Replicate key, and no other provider's model is substituted. Image editing runs on Replicate, or on an OpenAI-compatible endpoint that actually implements /images/edits — a missing route there is reported as a capability gap, not a credential problem.
• Figma MCP is no longer advertised or activated. Figma officially limits both its desktop and hosted MCP servers to clients in the Figma MCP Catalog, so shot2code filters registry results, disables migrated entries and uses the supported scoped-PAT REST workflow instead. Figma 429 responses now preserve retry, plan, limit-type and upgrade guidance.

Fixed

• Image generation no longer reports success for a batch that produced nothing. Every per-prompt failure used to collapse into None, the tool still returned success, and the activity feed counted the empty results as generated images — three blank grey tiles under "Generated 3 images". Each prompt now keeps its own classified outcome (429 rate limit, 402 billing, authentication, timeout, or a generic provider error), a batch that produced nothing returns a failure with the reason and what to do about it, a partial batch truthfully reads "Generated 2 of 5 images", and no success-shaped result ever carries an empty URL. The same counting fix applies to remove_backgrounds and edit_images.
• The activity feed replaces the unexplained grey "Failed" tile with an accessible alert carrying the classified reason and the action to take, so a screen reader announces the failure instead of reading an empty box.
• Image-provider credentials are excluded from commit snapshots, project history and streamed tool arguments, mirroring the existing BYOK, MCP and web-search secret handling.
• A manually configured Copilot SDK BYOK model is tested directly when an OpenAI-compatible gateway does not expose a working optional /models endpoint. Credential failures still fail immediately, and a connection with no configured model still receives an actionable discovery error.
• Non-maximized preview resizing is coalesced through one ResizeObserver and one animation-frame write. Unchanged geometry and hidden zero-sized panes no longer rewrite iframe dimensions or repeatedly report scale, eliminating the resize flicker without rebuilding the preview document.
• The selected preview tab, HTML/Stack source, view mode and zoom now survive restart through validated preferences. The desktop also restores validated on-screen window bounds/maximized state, while the active project, version and file continue to restore from SQLite rather than being duplicated in local storage.
• Restoring a saved project no longer reopens on a cancelled option when the same generation has a completed option. shot2code selects and persists the completed sibling; when every option was cancelled or failed, the sidebar now offers Retry instead of incorrectly telling the user to select a completed option that does not exist.
• The packaged file:// renderer now obtains the runtime backend HTTP and WebSocket addresses synchronously from the Electron main process instead of relying on a mutable environment variable visible to preload. History and API calls can no longer fall back to invalid file:/api/... URLs.
• The Electron preload no longer imports Node's crypto module. Electron 44 runs the preload in a sandbox where that module is unavailable; the import prevented every desktop bridge from loading even though the backend itself had started successfully.
• The in-app feedback IPC handler is registered during application startup rather than inside the Stitch import callback, so Bug, Feature request and Feedback submissions work before any Stitch action has run.

[0.5.1] — 2026-09-27

The v0.5.1 source tag was created during release validation, but no GitHub Release was published. Packaged smoke testing found a cancelled-option restore defect and two Electron preload/IPC defects. The draft release was deleted, the tag was not moved or reused, and the corrected build was rolled forward to 0.5.2.

[0.5.0] — 2026-09-27

hub release · upstream notes

Marked a pre-release and superseded by 0.5.2. The final stable build includes everything described here plus the corrections documented above.

It signs you in to GitHub Copilot from the app itself with no command-line tool installed, opens the official MCP Registry and an Agent Skills library, makes Figma and Google Stitch first-class design sources, restores the whole conversation in Chat, shows the exact files an export will contain before you download them, and keeps your projects when the installation directory is replaced.

Everything new here is additive and off until you switch it on. If you configure none of it, the app behaves as it did in 0.4.0 — the native OpenAI, Anthropic, Gemini and Copilot routes are untouched.

Added

• Sign in to GitHub Copilot without a CLI. The desktop app now runs its own GitHub OAuth device flow against an application registered for shot2code. It shows a one-time code, opens your browser, and completes without the GitHub Copilot CLI or the GitHub CLI being installed. The app uses a public client id and no client secret, because a desktop application cannot keep one.
• The token is encrypted at rest. The access token and its refresh token are stored under the app's own user-data folder, encrypted with Electron safeStorage — the operating system's own key store. They are refreshed automatically when they expire.
• Disconnecting is app-only. Disconnect GitHub from shot2code clears the credential this app holds and nothing else: a gh auth login or copilot session on the same machine stays signed in, because shot2code did not create it and has no business ending it. The app also says so, rather than leaving you to guess what it touched.
• The official MCP Registry, in the app. Settings → MCP servers → MCP Registry searches registry.modelcontextprotocol.io and installs an entry as a disabled, untrusted draft with its URL and headers laid out for review. Nothing a registry returns can start a server: the two existing switches still have to be turned on by hand. Only remote https:// entries are offered, and duplicates collapse to the latest active version.
• Featured integrations you would otherwise have to hand-configure. Figma Desktop MCP, Figma Remote MCP and Google Stitch MCP are offered as one-click drafts with the correct transport and endpoint pre-filled — and, like every other server, arrive disabled and untrusted.
• An Agent Skills library. Settings → Agent Skills imports a skill from a local folder or from a public GitHub folder URL (https://github.com/owner/repo/tree/main/path). Front matter, paths and sizes are validated, the origin is recorded as provenance, and each skill can be enabled, disabled or removed. Every imported skill is disabled by default.
• Figma, four ways. Exported screenshots, exported SVG files (which are rasterised locally before they are sent), the Figma Desktop MCP server on http://127.0.0.1:3845/mcp, and a REST import that takes a Figma URL plus a personal access token with file_content:read, resolves the file and node ids in the URL, asks Figma to render those frames, and brings the images in as local data URLs.
• Google Stitch, two ways. The official Stitch MCP endpoint at https://stitch.googleapis.com/mcp, and a bundled experimental @google/stitch-sdk in the desktop app that validates a Stitch API key, generates a screen from a prompt, and imports an existing Stitch project or screen. HTML and image downloads made on its behalf are HTTPS-only and size-bounded.
• Chat shows the whole conversation again. The active branch is reconstructed in order — your prompts, the screenshots or recording you attached, any selected-element context, the model identity, the generation state and the assistant's own responses, in expandable, scrollable blocks. Older versions no longer collapse to "Option ready."
• A first instruction on Upload and Import. Both tabs now accept an optional instruction — and a model choice — before the first generation, so an imported multi-file project can go straight into the refinement flow.
• Review can fix what it found. Fix selected findings sends a targeted update against the exact commit and option that was reviewed, rather than whatever happens to be on screen. Findings can be narrowed with severity filters and a text search, and selected in bulk with Select visible and Select errors + warnings.
• An optional, bounded AI review. Beside the deterministic local audit, a second opinion can be requested from the model that option actually ran on. It runs with no tools, no MCP servers, no skills, no web search, no shell and no file writes, and its findings are kept separate from the local ones, which remain authoritative. It is a real provider request and may use quota.
• A Design Inspector. Review can extract repeated colours, CSS variables, typography, spacing, radii, shadows, motion and semantic component counts from the generated source, and export them as DESIGN.md, SKILL.md and a palette PNG.
• The Code tab shows what you will actually download. Beside Current code there is now a read-only Export project view listing the exact text files and assets the ZIP will contain for the selected stack, so the export layout is no longer a surprise at download time.
• A sandboxed Stack preview. Preview still defaults to the composed HTML document. Stack preview additionally renders the controlled Vite HTML, React and Preact files in the same browser sandbox. No package script and no project configuration is executed to produce it.
• Copilot web research, opt-in. Settings → Copilot web research → Allow web search lets GitHub Copilot and Copilot SDK BYOK options search the public web when a prompt needs current documentation. Search queries leave the device, which the setting says in as many words. It is off by default.
• ScreenshotOne says what went wrong. A failed capture is now reported as a rejected key, a billing or credit problem, a rate limit, a timeout, an invalid URL or an unavailable provider — and the URL tab can test a ScreenshotOne key with one minimal request before you rely on it.
• One endpoint, several selectable models. A Copilot SDK BYOK connection is no longer limited to a single endpoint model: every model you discover or type becomes its own selectable identity, so several models from the same endpoint can run in one generation and be compared.

Changed

• Projects survive an upgrade. History lives in %LOCALAPPDATA%\shot2code\history.sqlite3, outside the installation directory, and the installer is explicitly configured not to remove application data. Replacing the installed program therefore leaves every project, version and prompt in place.
• History records the exact run identity. Each option stores and displays the identity it really ran as — a native model id, or the sdk-byok/<provider>/<base model> twin — so a retry replays that runtime rather than the provider whose model it borrowed.
• Image tools are gated on a usable key. generate_images and edit_images are no longer advertised to a model unless an effective Replicate key exists. The preference still works when that key comes from backend/.env.
• Skills and web search reach Copilot runtimes only. Like MCP tools, they are exposed to GitHub Copilot subscription options and Copilot SDK BYOK options. An option running on your own OpenAI, Anthropic or Gemini key never sees them.
• The shipped desktop stack was updated: Electron 44.4.3, electron-builder 26.15.3 and electron-updater 6.8.9, with npm audit --omit=dev reporting 0 vulnerabilities for what is actually distributed.
• First-run browser startup has a longer budget. The bundled headless Chromium is allowed more time on its first launch, because that is when antivirus software scans a newly written tree — the previous budget could report the browser unavailable on a machine where it was merely slow.

Fixed

• A generation started during a cold start no longer fails. The deferred route loader accepts the WebSocket handshake before the frozen generation graph has finished importing and replays the connect event, instead of dropping the first run after launch.
• Long runs no longer look dead. The backend sends a heartbeat every 15 seconds so a long Copilot run is not mistaken for a stalled connection.
• A completed run is no longer reported as a failure. An abnormal socket closure after every option has reached a terminal state is treated as completion, and when the selected option was cancelled or failed while another one finished, the usable option is selected for you.
• "Check the console" is a last resort again. A failure that the backend actually diagnosed keeps its own message and points at the diagnostic log.
• A root-level skill folder on GitHub imports correctly. The Contents API does not return file contents in a directory listing, so each file is now fetched individually rather than imported empty.
• A 204 route no longer breaks the backend. A FastAPI route annotated -> None for a no-content response made the whole application unimportable; it returns an explicit Response now.

Security

• No client secret ships in the app. The GitHub device flow uses a public client id only. A desktop application cannot keep a secret, so it does not pretend to have one, and it does not borrow another product's client id.
• Credentials are encrypted by the operating system. GitHub tokens obtained in the app are written through Electron safeStorage. The backend receives the token in its process environment on a restart that is serialised, so two sign-in or disconnect actions cannot race a half-started backend.
• Disconnecting never reaches outside the app. Clearing the credential shot2code holds does not touch a gh or Copilot CLI session on the same machine.
• Capture-only credentials never reach a model. The Figma personal access token and the Google Stitch API key are used only by the code that calls those services — the Figma token is sent to api.figma.com and nowhere else, and the Stitch key is passed through desktop IPC to the bundled SDK. Both are stripped from generation requests, from project history and from exports.
• A registry install is a draft, not a running server. An entry installed from the MCP Registry arrives disabled and untrusted, and only remote https:// entries are listed at all. The two switches, the read-only default and the separate Allow write tools gate are unchanged.
• Imported skills cannot execute. A skill may contain scripts, and those files are stored as resources — but shot2code exposes no shell tool and no unrestricted host-filesystem tool, so nothing in a skill can be run. Skills are bounded in file count and size, path traversal is rejected, and every import is disabled until you enable it.
• Copilot's own file and shell tools stay excluded. shot2code continues to expose only its own create_file and edit_file tools, which act on the project in memory. Enabling web search adds search and nothing else.
• The AI review runs without tools. It is given no tool surface at all — no MCP, no skills, no web search, no shell, no file writes — so a second opinion cannot become a second agent.
• Stitch downloads are constrained. HTML and image fetches made on the SDK's behalf must be https:// and are size-bounded.

Notes

• The GitHub device flow is a feature of the desktop app. Where it is not available, the existing delegated route is used instead and shot2code never receives or saves the token in that mode. The gh auth login / copilot ladder and the optional token field are unchanged either way.
• Figma's remote MCP server currently admits only clients in Figma's own MCP catalogue. shot2code does not impersonate another editor or reuse its OAuth identity, so the REST import with your own personal access token is the route that works today for most accounts.
• @google/stitch-sdk is experimental. Google Labs states it is not an officially supported Google product. It is bundled at version 0.3.5 (Apache-2.0) and is used only when you supply a Stitch API key.
• The AI review, like a connection check, is a real provider request on a metered account. The local source audit remains free, deterministic and offline, and it is still an automated check of generated source — not a WCAG conformance assessment.
• Validation for this build: 976 backend tests (13 skipped), 851 frontend tests (9 skipped) and 47 desktop tests, plus 15 exported projects checked across the 12-stack browser matrix. The packaged file:// app, history retention across an install-directory replacement, a root-level GitHub skill import and the bundled Chromium were all verified on the packaged build.

[0.4.0] — 2026-09-20

Download · hub release · upstream notes

It adds a second, optional way to reach a model — your own endpoint through the GitHub Copilot SDK — lets Copilot runs call MCP servers you configure, turns the preview into a side-by-side responsive Review with a local source audit, and puts a real Windows menu bar on the app.

Superseded by 0.5.2, which includes CLI-free GitHub sign-in, the MCP Registry, Agent Skills, Figma and Google Stitch.

Everything here is additive. If you do not configure any of it, the app behaves exactly as it did in 0.3.3.

Added

• Copilot SDK BYOK: bring your own endpoint. A single, separate connection in Settings with its own credential — OpenAI-compatible, Azure OpenAI or Anthropic, an optional base URL, an API key or a bearer token, an optional endpoint model, and an Azure api-version. No Copilot subscription is required. The OpenAI-compatible mode works with any endpoint that speaks the OpenAI wire format — a vendor API, a gateway, a self-hosted server or one your organisation runs — over any https:// URL, or http:// when the host is localhost. An OpenAI-compatible endpoint on localhost may run without a credential; everything else must carry its own. The direct OpenAI and Anthropic keys are never used as a fallback for it.
• The wire API is chosen for you. Automatic is the default: an endpoint with its own base URL gets Chat Completions, the interface almost every OpenAI-compatible server implements, and a provider's own endpoint gets Responses. You can pin either protocol if your endpoint needs one, and a pinned choice always wins.
• Endpoint models, discovered or typed. If the endpoint lists models at /models, Test model access fills a picker with the ids it reported — de-duplicated and capped — and you choose one. Azure and Anthropic do not expose that route, so the name is typed by hand there; typing is available everywhere. Model ids may be up to 128 characters and may contain the . _ : / @ + - characters real deployments use.
• A custom endpoint model is one honest option. Name a model the catalog does not know and Settings → Models shows exactly one entry for it — <model> via <provider>, with the run identity sdk-byok/<provider>/custom/<model> — instead of a list of catalog names that would all reach the same model. No reasoning effort is sent for it, because an arbitrary endpoint model has no thinking level.
• BYOK models are their own selections. Each model the connection can serve appears in Settings → Models under Copilot SDK (BYOK) with the run identity sdk-byok/<provider>/<base model>. Because that id can never collide with a direct model id, you can select a model and its BYOK twin in the same generation and compare the two results side by side. The reasoning effort you picked is part of the base model, so it carries across unchanged.
• The identity survives History and retries. Each option records the identity it actually ran as, so retrying a BYOK option re-runs it on your endpoint rather than on the provider whose model it borrowed. A custom endpoint model keeps its own identity through both.
• Sign in to GitHub Copilot from the app. Settings → GitHub Copilot offers Sign in with GitHub when nobody is signed in. It runs the official GitHub Copilot CLI web flow — falling back to the GitHub CLI — which opens your browser and stores the credential in that tool's own keychain. shot2code never receives, stores or sees the token; it only learns whether a session now exists. The flow can be cancelled, and if neither CLI is installed the app says so and links the official install instructions. The optional token field and the terminal route are unchanged.
• Test a provider before you generate. A compact Connection checks area under the API keys tests OpenAI, Anthropic, Gemini and Replicate individually, and the BYOK card gains Test model access. Each test makes one deliberately tiny request — a single-word prompt capped at 16 tokens — so a key that is malformed, revoked, out of credit or pointed at a model your account cannot see fails here rather than halfway through a generation. Leaving a field empty tests the key the backend holds in its own configuration instead. Replicate is checked against its account endpoint, so it starts no prediction.
• Results say what to do next. An answer is reported as Ready, Out of credit (with a link straight to that provider's billing page), Key rejected, Rate limited, Access denied, Model unavailable, Configuration problem or Could not reach the provider — never as an unexplained failure. An account with no credit left is the case this exists for: it looks like a rate limit and waiting never fixes it.
• Re-check the screenshot preview without restarting. The Screenshot Preview warning gains Check again, which re-probes the backend. Running from source it gives the exact install command and asks you to restart the backend; in the packaged app — where the browser is bundled — it says the bundled browser could not start, suggests a restart or reinstall, and opens the diagnostic log.
• MCP servers for Copilot runs. Up to 8 servers in Settings, over stdio, HTTP or SSE, with an optional tool allowlist and timeout. Their tools appear in the activity list as MCP · <server> · <tool>.
• A responsive Review workspace. A Review destination renders the generated page at two to four real widths at once — 1440, 768 and 390 by default, any whole width from 320 to 1920 if you add your own. Frames are rendered at their actual width rather than scaled screenshots, and horizontal overflow is measured in the running frame.
• A local source audit beside the frames. A deterministic pass over the generated source reports semantic and accessibility problems — missing lang, no <title>, missing viewport meta, heading order, missing main, unlabelled images, form controls and interactive elements, duplicate ids, positive tabindex, nested interactive elements, missing table captions and headers, and fixed widths that will overflow your narrowest frame — each with the evidence, the affected file and what to do about it.
• Send findings to Chat. Tick the findings you want and they are written into the composer as a grouped instruction. It is not sent for you — read it, edit it, then send.
• A source-safe JSON report. Export the review as JSON: severities, rule ids, messages, evidence and guidance, with file paths reduced to a leaf name and no credentials of any kind.
• A native Windows menu bar. File, Edit, View, Window and Help, driving the same commands as the keyboard: new project, upload, import, export, the Preview/Code/Chat/History destinations, Show Chat panel (Ctrl+Alt+C), Settings, zoom, the Help centre and the shortcut reference. Items that cannot work without a project are disabled rather than offered.

Changed

• The model picker has a fifth group. Copilot SDK (BYOK) sits beside GitHub Copilot, OpenAI, Anthropic and Google Gemini. The four native groups are untouched: each still needs its own key or sign-in, and none of them is ever relabelled or re-routed because BYOK is configured.
• Every API key field is masked. The OpenAI, Anthropic, Gemini and Replicate keys are password inputs with autocomplete off, matching the Copilot token and BYOK credential fields. The OpenAI base URL stays readable, because it is not a secret.
• A failed generation keeps its explanation. When the backend diagnoses a problem and then the connection closes, the actionable message is what you see. The generic "check the console" text now appears only when there really was no diagnosis, and it no longer replaces a specific one that arrived first.
• MCP tools reach the SDK runtimes only. A GitHub Copilot subscription option and a Copilot SDK BYOK option can call them. An option running on your own OpenAI, Anthropic or Gemini key runs on that provider's own client and never sees an MCP tool. The picker says so.
• Review results are bound to what produced them. A result records the version, the option, a hash of the source and the widths it ran at. Change any of those and the result is marked stale instead of being presented as current.

Fixed

• Every Settings control remains reachable. Settings now owns a viewport-bounded scroll area whether the app is empty or a project is open, so the final Screenshot by URL and capability controls no longer extend below the fixed desktop shell.

Security

• Sign-in is delegated, never implemented. shot2code runs the official CLI's own login command with a fixed argument vector, spawned directly and never through a shell. Nothing a request sends can influence that command line. The CLI's output is drained but never returned, logged or stored, because a login flow prints one-time codes and can echo tokens. The run is bounded by a timeout, can be cancelled, and is killed when the backend stops.
• Starting or cancelling a sign-in is origin-guarded. Both spawn or kill a process, so they are refused unless the request comes from the app on this machine — localhost, 127.0.0.1, ::1, or the null origin the packaged file:// app sends. A missing origin is refused too, because CORS does not stop a cross-site POST from being sent. Reading the status is unguarded only because it changes nothing and returns no token.
• The server's key is never lent to a caller's URL. A connection check that names its own OpenAI base URL must carry its own API key in the same request. The key configured on the server is only ever used with the server's own endpoint, so no caller can point shot2code at a host it controls and have the server's credential sent there.
• An MCP server needs two switches. It must be enabled and explicitly marked trusted before shot2code will start it or approve its tools.
• Read-only by default. A trusted server still cannot use write tools until Allow write tools is turned on for it, which the settings page labels as the risk it is.
• A local server is spawned as an argument vector, never through a shell, so nothing in a command or its arguments is re-interpreted by cmd or PowerShell.
• Remote servers must use https:// unless the host is localhost.
• Server secrets stay secret. Environment values and request headers that look like credentials are masked in the list and stay masked, read-only, in the editor until you ask to reveal them. Only key and header names appear in diagnostics; values never reach a log line, a validation response, a project snapshot or the exported review report.
• Connection checks return no credential. A result carries a category, a message and the model that was tried. Provider error text is scrubbed of any value that was sent before it is shown.
• Validation contacts nothing. Validate connection and Validate servers check the configuration only — no endpoint is called and no MCP server is started. Test model access and the per-provider checks are the opposite and say so: they contact the endpoint and may use a little quota.

Notes

• Configuring neither BYOK nor MCP changes nothing: an incomplete or switched-off BYOK connection, and a disabled or untrusted MCP server, are reported as notices and never block a direct generation.
• There is no Gemini BYOK: the Copilot SDK has no Gemini provider, so Gemini models always use the Gemini API key directly.
• A model reached through your own endpoint must accept the wire API in use and support image input and tool calling. Screenshots are sent as images and the agent works by calling tools, so a text-only model will fail. Neither the connection check nor the endpoint's model list can tell you which models qualify — only the endpoint's own documentation can.
• A connection check is a real request to a real provider and may consume a small amount of quota. It is deliberately tiny, but it is not free on a metered account.
• The source audit is a local, deterministic check of generated source. It is not a WCAG conformance assessment and does not replace testing with real assistive technology.

[0.3.3] — 2026-09-14

Download · hub release · upstream notes

It makes the packaged app start in seconds instead of minutes, lets you decide how the workspace splits its width, and replaces the shortcuts-only dialog with a Help centre that can open the log you need to attach to a bug report.

Superseded by 0.4.0, which adds Copilot SDK BYOK, MCP servers and the Review workspace.

Fixed

• Startup no longer waits for features you have not used yet. The frozen backend used to import the generation, evaluation and project-tool dependencies before health could answer, then launch Playwright synchronously. On a slow machine that delayed readiness by 299–347 seconds and blew past the shell's timeout — the splash screen that never cleared. Core routes now come up first, the heavy routers load on first use, and Chromium and Copilot discovery run as bounded background work.
• Readiness is checked strictly, not hopefully. The desktop shell now requires HTTP 200 and a real {"ok": true} body, gives up immediately if the backend exits or cannot be spawned, records how long readiness took, and uses a 90-second cold-start deadline instead of waiting on optional capabilities. Final packaged validation was 10 of 10 relaunches with no timeouts — 5.2 s fastest, 8.3 s median, 11.2 s slowest to a ready shell; 10.6 s on a fresh profile and 40.9 s for a fresh portable or post-update first run.
• A deferred feature that fails stays failed. If one of the lazily imported routers cannot load, the request returns an error and health starts failing from then on, rather than reporting a healthy backend that cannot generate. Chromium is advertised only after it has actually launched, and it is closed when the backend shuts down.

Added

• A resizable Chat and History panel. On wide windows (≥ 1280px) drag the divider between the conversation/History panel and Preview or Code. It is a real separator, not a decoration: arrow keys move it 16px, Shift+arrow 64px, Home/End jump to the narrowest and widest allowed widths, Enter or a double-click restores the default, and Escape cancels a drag in progress.
• A resizable project file explorer with its own divider on multi-file Code workspaces, and minimum widths for both the tree and the editor. Single-file projects still show no tree, so no space is wasted on an empty one.
• Pane widths that persist — and stay a view preference. Chat and explorer widths survive a restart and a collapse/reopen cycle, and are clamped to the current window so neither side can be squeezed into uselessness. They are kept entirely apart from project data: resizing a pane cannot change or create a version, an option, a retry or anything in History.
• An in-app Help centre. The rail's Help action and Ctrl+/ open four sections — Get started, Guides, Support and Keyboard shortcuts. Its links point at this hub: the product page, the complete release history, and the install, user, FAQ, troubleshooting, architecture, data-handling, security, changelog, contributing, third-party and releasing guides, plus the issue tracker and the source repository.
• Open diagnostic logs. In the packaged desktop app, Help opens the folder holding the backend log and says plainly whether that worked. The browser development build has no such log and explains why the action is unavailable instead of offering a dead button.
• Real desktop zoom. Ctrl+= and Ctrl++ zoom in, Ctrl+- zooms out, the numpad add and subtract keys work with Ctrl, and Ctrl+0 resets. Zoom moves in deterministic 10-point steps and is bounded between 50% and 300%.

Changed

• Below the desktop split breakpoint nothing moves: the Preview, Chat and History destinations behave exactly as before and no drag handle is shown.
• Chat spacing and the composer reflow at the narrowest supported panel width without hiding Send, the model picker or the design-system controls.
• The shortcuts-only dialog became the Help centre; the full shortcut reference is still inside it.

Accessibility

• Every pane handle is a role="separator" with vertical orientation, current, minimum and maximum values, readable value text, an explicit link to the pane it controls, keyboard instructions, visible focus and a 44-pixel interaction gutter.
• Help keeps focus trapped while open, labels each external link descriptively, uses 44-pixel tabs and rows, scrolls its tabs horizontally on narrow windows, and shows honest disabled states.
• The recognised zoom keys suppress Chromium's duplicate handling without swallowing ordinary editor input, AltGr or composition events.

[0.3.2] — 2026-09-14

Download · hub release · upstream notes

It protects an update that starts from an older client, extends model selection to every supported code provider, and finishes the responsive Preview and History work.

Superseded by 0.3.3, which makes the packaged app start in seconds.

Fixed

• An old client can no longer be replaced while its backend is running. The 0.3.1 guard lives in the app, so a machine still on an older build could not use it. The NSIS installer now runs its own pre-install safeguard before it uninstalls or replaces anything: it identifies processes by executable path under the installed resources\backend directory, terminates each captured process tree synchronously, and verifies that no matching process is left. It never kills by process name, so an unrelated program that happens to share an executable name is left alone.
• That safeguard fails closed and says why. If the installed backend tree cannot be confirmed stopped, the replacement is aborted — with an actionable message for an interactive install, or a distinct exit code for a silent one. Diagnostics are written to %TEMP%\shot2code-installer-preinstall.log.
• The 100% preview no longer hugs the left edge. A fixed-width desktop canvas is centred inside a neutral framed viewport. Narrower windows keep deliberate horizontal scrolling instead of clipping the start of the canvas.
• History is discoverable at every width. The user-facing name is now History everywhere it used to read Versions, with its own labelled destination beside Preview and Chat when the window is too narrow for the split view. History rows are real buttons with better focus, larger targets and stronger contrast.
• A strict Windows console can no longer crash prompt logging. Prompt-preview diagnostics are encoded for whatever the active output stream can actually represent, including a strict cp1252 console, instead of letting box-drawing characters or prompt text raise UnicodeEncodeError mid-generation. This affects running from source; the packaged app logs to a file.

Added

• Model selection for every code-generation provider. Settings and the compact picker beside the composer group the available models under GitHub Copilot, OpenAI, Anthropic and Google Gemini. Tick as many as you like: one option is generated per selected model, up to the per-run variant limit. Selecting nothing keeps the automatic behaviour.
• A credential-aware model catalogue. /api/models reports which providers are usable and what each model can do — without returning any secret. Copilot models are discovered from the signed-in account; the API-key providers use maintained, validated catalogues.
• Safe migration and honest handling of stale picks. An existing copilotModels list, or a codeGenerationModel that was actually changed from its old default, is migrated into the provider-neutral selection. Removed credentials, retired models and choices the current input mode cannot use are reported explicitly instead of silently dropped.
• Installer-level regression coverage. Desktop tests cover the NSIS wiring, the installed process-tree shutdown, the survival of an unrelated process with the same executable name, the fail-closed path, and the updater lifecycle.

Changed

• Preview controls are grouped by purpose: a Fit / 100% segmented control and a labelled History n/m control instead of loose buttons.
• A retry reuses the provider and model choices its source generation actually used, and history records the concrete model behind each variant.
• The published screenshots and these guides were retaken and rewritten for the centred preview, the History naming, the responsive destinations and the multi-provider model interface.

[0.3.1] — 2026-09-13

Download · hub release · upstream notes

A corrective release: it closes a race in the silent updater shipped in 0.3.0, and reworks the workspace around the chat panel, the code editor and the provider status the app can honestly report.

Superseded by 0.3.2, which moves the shutdown guard into the installer so it also protects machines updating from an older client.

Fixed

• The updater could start a second installer. Both install paths — the automatic one after a download, and Restart & install in Settings — called the install directly, so a queued click or an install-on-quit could launch a second installer while the first was already replacing files. There is now a single guarded entry point: the in-progress flag is set before any blocking work, and install-on-quit is disabled for that explicit path.
• Backend shutdown is verified instead of assumed. The process-tree kill is bounded by a 30-second timeout, its exit status is checked, and the process ID is re-probed to confirm the tree is gone. An already-exited backend counts as stopped.
• A failed shutdown aborts the install and stays retryable. If the tree cannot be confirmed dead, the update is not started at all — no more overwriting a live PyInstaller tree. Settings reports that shot2code could not shut down safely and asks you to quit and try again, instead of looking stuck.
• The new install path is packaged. It is listed in the electron-builder files array, so it ships inside the app rather than being dropped from the build.

Added

• Collapsible chat panel on wide layouts (≥ 1280px), so the preview or the code editor can take the full width. The choice is remembered; Ctrl+3 toggles it and Ctrl+1 / Ctrl+4 bring it back.
• Conversation empty state with starters. A new project offers three starting points — Make the layout responsive, Fix accessibility issues and Polish spacing and typography for imported code; Match the reference more closely, Make the layout responsive and Add the missing interactions for a generated first version. They insert the full instruction into the composer so you can edit it; nothing is sent on your behalf.
• Provider status callout that states only what the app can see: "No model provider saved in this browser" (with the reminder that credentials it cannot see from there, such as a Copilot sign-in, may still work) or "… saved on this device". It never claims a provider is connected or verified.
• Format action in the Code tab — a whitespace-only formatter for HTML, CSS, JavaScript and JSON. It is never automatic, never evaluates the file, refuses component source (JSX/TSX/Vue), and bails out unchanged on ambiguous or unbalanced syntax.
• Collapsible project file explorer, defaulting to open only when a project has more than one file.
• Editor status bar with the language and line count, the keyboard hint, and the reason CodePen is unavailable as a polite status message rather than a banner over the code.
• Read-only and Generated badges beside the existing Entry and Preview badges.

Changed

• The icon rail's Editor entry is now Chat, with an explicit collapse/expand control, aria-label/aria-pressed/aria-controls/ aria-expanded on rail buttons, decorative icons hidden from assistive technology, and targets at least 44px.
• Code tab toolbar labels shortened to Format, Copy, Download and CodePen, with the full descriptions moved into their accessible names.
• File tabs appear only when a project has more than one file, and the active tab is underlined rather than merely tinted.
• The editor highlights the active line and gutter and uses theme-correct surfaces in both themes.
• The variants strip renders only when a version actually has more than one variant.

[0.3.0] — 2026-09-13

hub release · upstream notes

The release that turned shot2code into a local-first project workspace: projects and their version history are saved on your own machine, generated output is a real multi-file project you can edit, and previews run in a locked-down sandbox.

Superseded by 0.3.1, which fixes the updater race it shipped with.

Added

• Local project history (SQLite). Projects, versions, variants, prompts and messages are stored in history.sqlite3 in a per-user data directory (%LOCALAPPDATA%\shot2code\ on Windows), with a versioned schema, tracked migrations and an /api/history route group. A Recent projects panel restores earlier work; deleting a project removes it and its versions from the device.
• Retry ancestry. Commits record both their parent and, for a retry, the commit they re-roll. Retrying reuses the models that produced the original variants, and ancestry walks reject cycles.
• Multi-file editor and project tree. A keyboard-navigable file tree with expandable folders and read-only files announced as such, file tabs with arrow-key navigation, per-file copy/download, and the resolved entry point badged. The tree is the authoritative source for export, CodePen and preview.
• Safe, sandboxed previews. Rendered from srcDoc without allow-same-origin, with an injected Content-Security-Policy, no-referrer, and camera/microphone/geolocation/display-capture denied. Select-and-edit moved to a nonce-checked per-preview message bridge with capped payloads, and an unrepresentable build shows a deterministic fallback instead of a silently broken page.
• Import, export and CodePen across all 12 stacks. Import a folder, ZIP or selected files — parsed as text only, never executed, bounded by explicit archive, entry, file, size and text limits. Export produces a single self-contained HTML file or a project folder with a documented per-stack strategy. CodePen is offered only when the stack can honestly run in a Pen, and always asks first.
• Multi-screenshot modes: pages (the default), responsive, states and references, saved with the project and sent with each generation.
• Shortcut reference (Ctrl+/) and conflict-free Ctrl+Alt project shortcuts, plus accessibility work across the workspace.
• GitHub Copilot model discovery — the app asks the signed-in account which models it can use and whether each supports vision, instead of shipping a hardcoded list. Credentials are resolved without persisting or logging a token.
• Desktop updater and packaging — auto-update for per-user NSIS installs with the state surfaced in Settings, silent install and relaunch, managed-install detection under Program Files, and NSIS/MSI/portable builds with the backend frozen by PyInstaller and only headless Chromium bundled.

Security

• Previews run in an opaque origin under a restrictive CSP with a nonce-checked edit bridge.
• Imported projects are parsed as text only and never executed; traversal paths are rejected and count/size/text limits enforced.
• Raw imported source is never written into persisted project context.
• CodePen sharing requires explicit confirmation before any code leaves the device.