Caide Releases

Commands stream their output while they run, long tool groups window themselves, a right-gutter minimap makes long chats navigable, and every tool call renders through one typed surface.

• run_command streams its output into the live tool row (ANSI parsed, 64k budget) instead of showing a motionless spinner, and the live view disappears when the real result lands
• Every tool call now renders through one typed activity surface: failed rows are marked, each row shows its duration and a hover timestamp, and the raw caide-tool tags are gone
• Long chats get a right-gutter transcript minimap with bucketed marks, tone priority, and pointer or keyboard scrubbing that jumps through the transcript
• Expanded tool groups above 24 rows window their rows, so long runs no longer mount every row
• read_url rows show the site's favicon
• The agent stays quiet by default; its repetition guard now sees short status lines, and the narration setting round-trips
• Ticking consent no longer causes hackathon submissions to be rejected

A step's narration collapses to its latest status line, the model picker keeps its full layout while a turn runs, and the unused handoff menu is gone.

• Within a step, the live text collapses to its latest status line when the model starts calling tools, so narration stops piling into one paragraph
• The model picker keeps its full layout during a turn: the provider rail is disabled with a "Locked to … until this turn finishes" hint, while effort and mode stay adjustable
• The thread menu no longer lists Handoff to …; the engine's model-switch handoff is unchanged

Only build and lint can block a turn now; design, security, performance, and the rest run after the turn and land in the Problems tab.

• Only build_project and lint_project block a turn; taste, edge, security, performance, benchmark, coherence, and color checks are advisory and run after the turn instead of holding it
• The fixer runs a single pass on blocking issues instead of looping up to three times on advisory ones
• Advisory checks warm in the background after a completed build, so the Problems tab fills without adding time to the turn
• Advisory issues appear in Problems with their location, capped at 30, and never fail the turn

Copy actions across the desktop app work again, and clipboard permissions are granted only to Caide's own window.

• Copy and paste work again in the desktop app: clipboard read and write permissions are granted to Caide's own renderer instead of being denied app-wide
• Browser webviews stay denied, so third-party pages can neither read nor overwrite your clipboard in the background
• The share-project dialog now uses the shared copy helper, so its Copy link gets the same fallback as the rest of the app

Send a project to anyone with a Caide link that imports straight into their app, browse and enter the Caide hackathon from the site, and earn 30% when someone you refer buys a Caide Agent pass.

• Share project from the project menu: the shared package carries checkpoints, agent memory, and decisions as provenance, never your .env, and opens on the web before importing into Caide
• Add a project gains a Caide link source, and caide://receive-project links import a shared project directly
• The new /hackathons page has a live countdown, prize ladder, submission form with automatic image downscaling, rules, and an FAQ; the gallery stays locked until judging
• Referrals pay 30% (₦900) of every Caide Agent pass bought by someone who signs up through your link, with a dashboard share link, masked referrals, balances, and a bank-withdrawal request
• The terms now cover the referral program, including which payments count and how payouts work

Commit messages, PR copy, and thread recaps generate again, the preview stage drops dead controls for element annotation, and @mentions, local images, automation setup, and short-transcript jumps all work.

• Blank commit messages, Create PR titles/bodies, and the Environment recap are generated again; Create PR still falls back to the repository template or newest commit when no text model is usable
• The preview header loses rotate, quality/controls, screenshot-save, and QR branches; the component picker becomes a browser-style element inspector that annotates the picked element with its tag, selector, source file, and styles, and captures are instant
• Mentioning a chat with @ now hands the agent a bounded transcript of that chat, and $skill references survive delegation
• Automation setup stops looping: real intent extraction, wall-clock one-offs like "by 6 15 pm today", a 4s generation budget with local inference, and a three-question cap
• Local chat images render again in the desktop app, and the blank space below a settled turn collapses
• Pinned-message and message-trail jumps now scroll short transcripts instead of silently doing nothing
• The chat header's add-script editor is removed (saved scripts still run, and the sidebar dialog still saves them), niche keybindings are pruned, and agent routing no longer offers integration credentials as models

React Native, Flutter, Website, and Blank projects stop leaking preview chrome, prompts, and tooling into each other, transient provider failures retry instead of parking, and steer bubbles stay between the replies they split.

• Framework isolation: preview chrome, slash commands, build/plan prompts, doctor probes, and lint/test/package tooling follow the project's framework; React Native releases explain EAS instead of offering APK/AAB/IPA targets that always fail, and Blank loses the mobile chrome
• A clean pre-stream provider failure now retries once after five seconds with a retrying-provider state instead of ending the turn; failures that produced output still park for Resume
• Steer bubbles render between the pre-steer and post-steer assistant replies, with a monotonic clock keeping same-millisecond order
• The Turbo edits toggle is removed; its batch-edit rules now live in the base build prompt, so turns behave the same with one less control
• The model picker popup stays pinned above its trigger instead of flipping below it
• React Native docs describe the actual expo-router/token scaffold instead of NativeWind

Turns no longer collide after a restart, the cost guard runs to 24M tokens with a real budget setting, and the composer footer folds into the model picker.

• Every send mints a unique turn id, so a restart can no longer merge a new message into an old bubble; steered follow-ups appear as their own user rows and queue-mode sends queue instead of diverting
• The cost guard rises from 6M to 24M billable input tokens, with an Agent budgets ladder (12M/24M/48M), and every pause now names its cause
• The agent speaks one short sentence per phase, and prose alongside routine tool calls is capped structurally instead of by wording alone
• Agent-started previews open immediately instead of waiting for a navigation remount
• Build/Plan mode and access mode move inside the model picker, leaving the composer footer as a single trigger
• Fast mode and the composer provider-setup strip are removed because neither changed what the engine did
• The sidebar backdrop runs full width and the activity rail sits below the project image
• Docs search on caide.us works again, including partial words

Goals are gone from the product, long runs no longer pause on segment or time caps, and the agent speaks once per phase instead of narrating every tool.

• Goals are removed across the engine, contracts, client, prompt, and catalog
• Long runs auto-continue past segment and time caps; cost, idle, entitlement, and no-progress stops still pause, with a Settings toggle
• The agent stays silent by default, speaks one short sentence per phase change, and guards against repeated prose
• The composer keeps queue/steer send reachable during a live turn, with Send now, edit, and delete on queued rows
• UX pass: unified model picker, slash-command cleanup, fresher account state, clearer context markers, a sign-in backdrop, and a settings audit
• Chat gains a labelled scroll-to-end pill, and the command palette restores Import Project
• The public changelog now starts at 0.5.0 on both the desktop What's New list and the marketing site

Concurrent threads no longer share live-turn state, Turbo edits are per project, and the token ceiling pauses for Resume instead of auto-continuing.

• Transcript mirror keys are scoped per session and turn, so concurrent threads can no longer settle each other's messages
• Turbo edits are stored per project in caide.turbo-edits.v2 and the legacy global key is removed
• The token ceiling now joins the pause set with Resume offered and never auto-continues into a new segment
• Cached tokens count at one tenth of input, so heavily cached turns no longer trip the ceiling after ~400k of real work
• Recoverable control codes stay out of the transcript and duplicate inline errors are deduped
• Build segments run 15 minutes, with three run segments by default so builds finish

A single build could reach 44 minutes and 26M input tokens; turns now batch file tools, compact at 96k, cap context and screenshots, and stop persisting image bytes.

• Batched read_files/write_files replace one-file-at-a-time calls, with 8-minute build segments and a 6M cumulative input-token ceiling
• Compaction triggers at 96k by default, the in-turn transcript and tool results are bounded, and image bytes no longer count against the text budget
• Screenshots capture at 1x with reduced motion, cap at four per turn, prune to the newest sixty files, and are never persisted as transcript bytes
• Per-step usage rows record input, output, and cached tokens so a turn's real cost is measurable
• Text-only models receive the screenshot path and dimensions instead of pixels; DeepSeek is declared text-only
• The role block and website brief match the actual scaffold, install_package takes a package list and honours the lockfile, and search skips build outputs
• Harness turns now project into the thread read model so history matches the live session

The collaboration preview is out of production, agent access is restored across integrations and tools, and the UI, dependencies, and marketing claims were hardened.

• Collaboration is removed from production and preserved on feat/collab-project-pair-archive; the harness reconnect fix stays
• Agent access restored across GitHub, Neon, Supabase, preview, and the tool surface, including Plan-mode tool allowances
• New agent capabilities: APK artifact paths, preview tunnel status, .caidepkg inspection, project-scoped history search, and local-server visibility
• Desktop renderer boundaries and engine lifecycle hardened; Electron, updater, Next, router, and Vitest security advisories patched
• UI pass: unified Geist/Hugeicons surfaces, responsive shell, docs footer, keybinding groups, project controls, and readable stage banners
• Marketing now states only published release evidence, install verification steps are documented, and a path-traversal hole was closed
• Claude Code is out of the provider picker while it is unavailable

Pairing links and desktop auth work again under the caide:// origin, and chat reconnects after the local engine restarts on a new port.

• Auth routes answer CORS preflight for the caide:// renderer, so invite and pairing link creation no longer fails with "Failed to fetch"
• Harness sockets re-resolve the bridge URL on every reconnect instead of pinning a dead loopback port after a backend restart
• Settings probes the selected pairing address and shows an actionable firewall warning when the host is unreachable
• The Team pane opens at a 30rem conversation width instead of a 50/50 split

A second computer can drive your Caide natively, shared projects get team chat with presence and invites, and first-run users get a guided tour.

• Native desktop remote mode drives a host machine through a loopback bearer proxy, with no local database, project, or provider keys on the guest
• Team collaboration: shared project chat with presence and creator attribution, one-time invite links with QR pairing, and per-device revoke
• Project handoff bundles attachments, imports into a path with no existing project, and a lease witness guards remote secrets
• First-run onboarding and a guided tour walk new users through the workspace
• Feedback reports can attach screenshots, and streaming turns coalesce transcript writes instead of rewriting the thread every ~100ms
• Fixes: integration panes return real data, local images and PDFs render again, Ctrl+P reveals the preview, and design-reference uploads save

App inspection moves to the Preview console while the browser tool refuses the preview origin, run status ticks live, and the chat shell keeps the transcript as the only scroll surface.

• Browser tools refuse the preview origin; app inspection routes through the Preview console and the prompt explains the split
• The long-run strip stays hidden on ordinary first segments and active time ticks between server updates
• A composer-footer toggle shows or hides the run and to-do panels, with todos reconciled at turn end
• The chat shell is viewport-clipped so only the transcript scrolls
• Cold start restores the most recently used thread instead of a skeleton draft

Streaming replies now parse markdown incrementally, and decorative animations stop painting when they are off screen or motion is reduced.

• Streaming replies reuse the completed markdown prefix instead of reparsing the whole message
• Decorative animations run only while visible, window-visible, and allowed by reduced-motion
• Skeleton, spinner, consent strip, and timeline indicators follow the same visibility gate
• Smooth streamed text keeps its pacing without keeping the GPU awake off screen

Project panes now follow the active worktree and stay safe under agent edits, turns carry the canonical project root, and the desktop release history mirrors the marketing changelog.

• Project panes resolve against the active worktree and stay agent-safe
• Turns carry the canonical project root so tools, previews, and panes agree
• Desktop What's New projects the shared release history instead of its own list
• Account details moved into a sidebar profile popover
• Checkout buttons show a pending state and reject double submits

Long tasks now run as checkpointed segments, Plan mode has a single accept path, and every model keeps one resolved context window for the whole turn.

• Long-running turns execute as checkpointed segments and resume from the newest todo rows
• Plan mode has one accept path; the accepted plan stays in the dock and the gate clears
• One context-window resolver per model, pinned for the turn; Caide Agent runs a 1M window
• Admin access grants support explicit 30-day and permanent passes with expired status
• macOS builds are ad-hoc signed, so they open with the standard right-click > Open path

Caide's first public desktop build ships for macOS, Linux, and Windows with the full agent workspace: seven model providers, Build and Plan modes, DeviceLab previews, and free BYOK.

• macOS (Apple Silicon and Intel), Linux, and Windows installers published on the public releases repo
• Seven model providers: OpenRouter, OpenCode Console, Caide Agent, ChatGPT/Codex, Claude Code, DeepSeek, and Google
• Build and Plan modes, with subagents and consent-gated tools
• DeviceLab mobile previews and local-first SQLite history
• Free BYOK forever; optional Caide Agent at ₦3,000 for 30 days after a 7-day / 500-credit trial