# Changelog

All notable changes to Open CoDesign are documented here. Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Versioning: [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

## [0.1.4] — 2026-04-23

AI image generation, ChatGPT subscription login, and a large reliability wave for third-party API relays. No breaking changes.

### Highlights

- **AI image generation** — the agent can generate bitmap assets (heroes, product shots, illustrations, logos) inline during design. gpt-image-2 via OpenAI, or any OpenRouter image model. Off by default; opt in from Settings.
- **ChatGPT Plus / Codex subscription login** — one-click OAuth, no API key required.
- **CLIProxyAPI one-click import** — auto-detect a running CLIProxyAPI instance on startup; preset + smart model auto-discovery for manual setup.
- **Clearer errors for third-party relays** — timeouts, SSE truncation, and Messages-API-incompatible gateways now produce actionable diagnostics instead of `undefined` / silent stops.
- **First-time contributors** — @yangjunx21 (image generation), @DavidgFernandes (pt-BR locale), @L4b0R (Codex Ollama visibility), @1WorldCapture (tooltip restoration).

### Fixed

Provider & relay compatibility:

- Gemini via third-party relay: keys with `models/` prefix now work (fixes #175, reported by @CaioGS06).
- OpenAI-responses gateways no longer reject with "instructions are required" (fixes #134).
- Gateways without `/v1/messages` now return a clear protocol-mismatch error (fixes #158).
- Third-party relay SSE `response`-event-name truncation now surfaces a diagnostic hint (fixes #167, #180).
- Custom OpenAI-compatible endpoints without `/models` can now still be tested (fixes #179).
- `reasoning=true` is only sent to known OpenAI reasoning models (fixes #183).
- First-turn agent generation retries on transient upstream errors (fixes #125).

Desktop:

- Generation timeout preserves its reason instead of collapsing to "Request was aborted" (fixes #169).
- Settings active-model card no longer misrepresents the model when it's not in the fetched list (fixes #136).
- ChatGPT OAuth login can be cancelled mid-flow (fixes #172).
- Add-menu tooltip behaviour and labels restored (fixes #156).

Codex:

- Token-store wraps JSON parse errors and uses atomic writes (fixes #128).
- Screenshot attachments now reach ChatGPT Codex correctly (fixes #157).
- Ollama hidden from Codex import until manually added (fixes #170).

Diagnostics:

- Generate failures bridge into `diagnose()` hypotheses for targeted fix suggestions (fixes #130).

Release pipeline:

- Both macOS arm64 and Intel x64 DMGs ship with correct native modules (fixes #176, reported by @bytegh).
- Coding-plan app allowlists surface a clear warning when provider expects them.

### Added

- Brazilian Portuguese (pt-BR) locale (by @DavidgFernandes).
- CLIProxyAPI preset + auto-detect banner.
- Per-row model selector in custom provider modal with auto-discovery.
- Diagnostics panel: actionable hints for third-party relay SSE truncation and missing-Messages-API gateways.

### Install

```bash
winget install OpenCoworkAI.OpenCoDesign           # Windows
brew install --cask opencoworkai/tap/open-codesign # macOS
```

Or direct download from [GitHub Releases](https://github.com/OpenCoworkAI/open-codesign/releases/tag/v0.1.4).

---

## [0.1.3] — 2026-04-21

Fast-follow patch. Provider reliability and release-pipeline polish.

### Fixed

- Packaging manifests for Homebrew Cask and winget now track release assets deterministically.
- Unsigned-installer reliable manual-install workflow documented in README for Gatekeeper on macOS Sequoia 15+.

---

## [0.1.2] — 2026-04-21

First patch release. Release pipeline wired end-to-end.

### Added

- Packaging manifests scaffolded for Homebrew Cask, winget, and Scoop.
- SHA256SUMS.txt + CycloneDX SBOM attached to each release.

---

## [0.1.1] — 2026-04-21

Post-v0.1.0 maintenance release. No breaking changes. Focused on runtime correctness, release-pipeline polish, and prompt quality.

### Highlights

- **JSX previews unblocked** — replaced the broken React UMD vendor bundle so React/JSX artifacts render reliably.
- **Context pruning** — size-based prune of tool-call input and tool-result payloads during orchestration, with an artifact-wrapper override, so long agent sessions no longer balloon the prompt.
- **Interactive depth mandate** — prompts now require a craft-surplus minimum across every artifact type (real interactions, real states, not a wireframe).
- **Keyless providers in Codex import** — the Claude Code / Codex import flow now correctly pulls in IP-allowlisted proxies that have no API key.
- **Release pipeline** — workspace build + bundler smoke test added to CI; electron-builder auto-update metadata disabled until signing ships; all i18n menu keys supplied.
- **Packaging manifests** scaffolded for Homebrew Cask, winget, and Scoop.

### All changes

- feat(prompts): mandate interactive depth + craft-surplus minimums (1f3b913)
- fix(core): size-based context pruning + artifact-wrapper override (b692ec7)
- fix(core): aggressive context-prune — compact `toolCall.input` + `toolResult` (b614196)
- fix(desktop): support imported Codex keyless providers (1cdf006)
- fix(runtime): inject comment overlay into HTML previews (bed4458)
- fix(runtime): replace broken React UMD vendor + unblock JSX previews (4536297)
- fix(release): disable electron-builder auto-update metadata (16be6cc)
- fix(i18n): add provider import menu keys (29a3f97)
- fix(ci): release workspace build + add bundler smoke to CI (4cc21df)
- docs(packaging): scaffold Homebrew / winget / Scoop manifests (d342eae)

---

## [0.1.0] — 2026-04-18

First public release. Electron desktop app, GitHub releases for macOS (DMG), Windows (EXE), and Linux (AppImage). Installers are unsigned in v0.1 — code-signing + opt-in auto-update land in v0.5.

### Added

- **Unified provider model** — Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, SiliconFlow, local Ollama, and any OpenAI-compatible relay. Switch in Settings with no hardcoded shortlists.
- **Dynamic model picker** — every provider exposes its real model catalogue fetched at connect time.
- **Keyless providers** — IP-allowlisted proxies (enterprise gateways, local Ollama) supported with empty API key.
- **Claude Code + Codex config import** — one-click import reads `~/.codex/config.toml` and `~/.claude/settings.json` and pulls in every provider, model, and key.
- **Prompt → HTML or JSX/React component** — rendered in a sandboxed iframe with vendored React 18 + Babel, fully on-device.
- **Twelve built-in design skill modules** — slide decks, dashboards, landing pages, SVG charts, glassmorphism, editorial typography, heroes, pricing, footers, chat UIs, data tables, calendars. The agent picks the skill that matches the brief.
- **Fifteen ready-to-edit demo prompts** — cosmic animation, organic loaders, landing page, case study, dashboard, pitch slide, email, mobile app, pricing page, blog article, event calendar, chat interface, portfolio gallery, receipt/invoice, settings panel. All localized EN + 简体中文.
- **Anti-AI-slop design Skill** injected into every generation.
- **Inline comment → AI patch loop** — click any element in the preview, leave a note, the model rewrites only that region (str_replace against stable `data-codesign-id`).
- **AI-generated tunable sliders** — the model emits the parameters worth tweaking (color, spacing, font); drag to refine without re-prompting.
- **Live agent panel** — streaming tool calls rendered in real time while the model edits files.
- **Instant design switching** — preview-iframe pool keeps the last five designs hot in memory, Hub ↔ Workspace switching is zero-delay.
- **Phone / tablet / desktop preview frames** — true responsive canvas switching.
- **Files panel** — inspect multi-file artifacts (HTML, CSS, JS) before export.
- **Connection diagnostic panel** — one-click test for any provider with actionable errors.
- **Light + dark themes**; **English + 简体中文 UI** with live language toggle.
- **Five export formats** — HTML (inlined CSS), PDF (via user's local Chrome through puppeteer-core), PPTX (pptxgenjs), ZIP asset bundle, Markdown. All exporters lazy-loaded.
- **Generation cancellation** — stop mid-stream without losing prior turns.
- **Four-tab Settings** — Models, Appearance, Storage, Advanced.
- **Local-first storage** — SQLite (better-sqlite3) for design history and snapshots; TOML (encrypted via Electron `safeStorage`) for config. No electron-store blob. XDG-compliant paths.
- **Zero mandatory telemetry, zero cloud account, zero API proxy.**
- **Install size ≤ 80 MB** across macOS and Windows installers (CI-enforced).
- **MIT license** with third-party NOTICE file and machine-readable `CITATION.cff`.

### Technical foundation

- pnpm workspace + Turborepo + Biome + TypeScript (`strict: true`, `verbatimModuleSyntax: true`)
- Electron + React 19 + Vite 6 + Tailwind v4
- `@mariozechner/pi-ai` as the multi-provider abstraction
- Vitest for unit tests (~700 tests across 11 packages), Playwright for E2E
- Sandboxed iframe renderer using esbuild-wasm + import maps

[0.1.1]: https://github.com/OpenCoworkAI/open-codesign/compare/v0.1.0...v0.1.1
[0.1.0]: https://github.com/OpenCoworkAI/open-codesign/releases/tag/v0.1.0