shot2code v0.5.2

[0.5.2] - 2026-09-28

Released as shot2code v0.5.2.
Checksums for the published Windows artifacts:
SHA256SUMS.txt.

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.