otst/osw-studio

Gemini tool calls decoded with thought signatures replayed, Ollama on its native endpoint with a context length setting per local provider, an Ollama setup check in Connections, fix server mode refusing keyless local providers, fix a failed command chain completing the task, fix local providers resolving to IPv6

image generation with a deadline and a stop that cancels it, stopped tasks ask why, terminal API errors end the task instead of pausing, www redirects for custom domains, referrer and setup-path telemetry, fix an expired HuggingFace sign-in looping on Continue, fix Continue rendering in server mode, fix DeepSeek and mesh-llm defaults

quick edit as a second workspace surface, a simple view per account or device, workspace owners managing their own members, the Inspector on a phone, browser-painted screenshots with a preview capture in the composer, fix a local edit reading as another device's, fix project settings stored as a string losing a runtime

contextual chat suggestions scoped to preview pages, project links in server mode with back controls that name their destination, extra headers for custom providers, unresolved review counts on deployments, listings that fold columns and row actions as they narrow, mail settings moved from their own page into Settings, fix a workspace URL rendering another workspace's shell, fix provider fields being uneditable inside a project, fix SQLite timestamps read as local time

email digests for review comments over instance or per-workspace smtp, mail settings page and unsubscribe links, deep links from the review inbox to a comment, fix origin checks matching on string prefix and a handoff token passing as an account session

table views for every listing with remembered layout, shared page shell and section primitives, deployment settings as a full page, muted-orange active states, server-mode web search, gated commands pause for approval

unified settings pane, sidebar flyout sub-menus, models.dev pricing and metadata for all cloud providers, admin edit-user dialog with workspace memberships, dialog height standardization

server-side build delegation to browser, hide python/lua from agent, admin password reset, esbuild cache invalidation, compile-error file attribution, error-log key leak fix, hollow test audit

workspace quota enforcement on all server mode instances, unlimited defaults, ungated UI

preview direct editing with toolbar, Inspector styles and content, Elements tab with source annotations, gate to static and handlebars runtimes, fix the preview accepting messages from a mount with no frame, stop the hidden mount from compiling

skip the auth middleware for static-asset requests so a preview referencing files a project does not contain stops costing a session check per 404, sync binary files whose ArrayBuffer came from another realm as bytes rather than an empty file

compile a project of several hundred pages without timing out, store binary content once and hardlink it into deployments instead of copying, fix a project file path escaping its deployment directory on publish and workspace deletion leaving published sites served

fix publishing a deployment and syncing a custom template failing when they outgrow one request body, check the storage quota once per push rather than per batch

send every project push in batches so a project too large for one request body still syncs, cover backend features and project settings in checkpoints

store and edit php, java, sql and other text formats, next 16 and a clean npm audit, fix desktop publishing writing into the install directory, deploy losing its project, and the agent answering its own question, changelog, what's new and version bump

twenty built-in templates with preview before select and intent grouping, bundle deps instead of CDN fetches, fix quickjs aborting on edge function fetch and teardown, changelog, what's new and version bump

fix a project's database schema being left behind by downloads, backups and server sync

download a project as an editable archive and import one back with a preview of what would change

split the agent shell into one module per command behind a shared command registry, add rmdir, and fix the write escapes it exposed: runtime writing past a restricted agent's directory, generate-image and runtime writing in read-only chat, and coding instructions prepended to transcription and skill checks

recognise redirects written without spaces while leaving markup, quoted text and fd redirects alone, document head -c/tail -c in the shell reference

browse templates in one searchable list instead of picking a runtime first, add DeepSeek provider, sort providers alphabetically, fix agent shell head/tail -c and unspaced operators, changelog, what's new and version bump

let an explicit push from Server Sync overwrite a conflicting server copy, while background syncs still report the conflict instead of picking a side

fix .osws restores keeping the sync timestamps of whichever instance the backup came from, leaving restored projects stuck as a conflict or local-only until pushed by hand

sync reliability, any file type in a project, fix analytics/edge/scheduled functions reading the wrong database, .osws backup and template binaries, analytics authorization, changelog, what's new and version bump

free DuckDuckGo web search and live Codex model discovery, changelog, what's new and version bump

chatgpt subscription browser sign-in changelog, what's new and version bump

fix agent bash-tool parsing: html-entity file corruption, chained multi-file writes, inline python/lua, clearer offline runtime errors

server-mode sync & generation reliability fixes, HuggingFace sign-in nag, analytics dialog focus outline (#18)

preview reliability (client/hash router navigation respected, recover from escaped frames, stable readiness so reload/save work, compile watchdog, only writes reload the preview), fix mobile voice duplication and piped local-curl web-access prompt

fix stale HuggingFace connections (display name stored as username) failing Space publish with a raw 403 — detect and prompt reconnect, and reopen Deploy after the OAuth round-trip

fix HuggingFace Space publish failing under the display name instead of the username (403), stop misreading it as a missing permission

deploy from the workspace (HuggingFace Space / OSW instance / ZIP target picker, HF token paste), consolidate Project + settings into one Settings dialog, surface broken images to the agent, guard against reverting manual edits between tasks

neutral interface pass (orange fill > neutral on buttons, nav, cards; orange kept for accents), fix HuggingFace sign-in blocking chat when already connected (#17), light-mode dashboard cards, docs page/TOC fixes

first-run onboarding improvements (workspace-first landing with a starter project, one-click HuggingFace sign-in, project kept in the URL, starter pills), global model selection replacing per-project config (immediate picks with save/reset, read-only built-in templates, connecting a provider readies every project)

desktop fixes (generation auth, auto-update 404, update failure reporting, logs folder), project-wide release notes

web access and permission gating (search + external curl through an SSRF-guarded proxy, Auto/Ask/Custom approval modes with per-command matrix), fix desktop server-mode generation 401 (routes now use the desktop-aware session), connections split into AI + search sections, collapsible model-picker groups

expand anonymous usage analytics (init in all modes so server/desktop report, ~35 new privacy-scrubbed events, event registry drives the consent disclosure), rework the disclosure dialog

user-managed interview templates (editor, built-ins duplicable, Interviews main-menu entry, server-mode sync), custom items sorted before built-ins in listings, narrower skill/interview editor dialogs

fix zip/json export dropping assets and settings, server-mode import/duplicate deployability, deployment assets 404 on first publish

fix server-mode generation using global model instead of project model, preview 404 on runtime-fetched files

opencode go provider (models from models.dev, per-model openai/anthropic routing)

default custom-model context length to 128k

fix chat readiness keying off global provider instead of project agent

fix provider connection inference, mobile model picker (loading, 404), empty-picker → connections, /admin redirect query params

per-project models & templates (agent/voice/image slots, connections, recommended templates), template sync, voice input (on-device or transcription model), image generation, streaming tool-name early-abort, hide internal reminders from chat

interview mode (templates, picker, completion gate, build handoff), duplicate chat message fix, chained status-complete fix, shell find . path fix, desktop boot fix (bcrypt addon, failure diagnostics, fedora self-test

desktop workspace recovery (stale state re-init, legacy data migration), recovery smoke test, mac upgrade warning in release notes

desktop fixes (linux/windows boot, consent-based updates, failure diagnostics), atomic release pipeline, single-source desktop shell

agent reliability pass (truthful results, reasoning models, sub-agent isolation, compaction), file tree subdir fix, debug messages tab, blank-canvas starters

transparent sync (non-blocking, removed setup modals), immediate conflict alerts, sync perf (parallel pulls, request dedup), input modality indicators