Forge 0.10.0

2026-08-12

• Forge Remote — your agents on your iPhone. A companion app mirrors every workspace's live status, notifies you the moment a run finishes or stalls waiting on you (approve/deny tool permissions right from the notification), shows each project's board, and sends the next prompt straight into a live Claude pane. On your home Wi-Fi it talks to your Mac directly over an encrypted local socket ("Live" badge); away from home it rides your own iCloud — no accounts, no third-party servers. Every loading state says exactly what it's doing.
• Streaming a chat no longer lags the whole app. Each Claude transcript now renders in its own layout island, so a busy agent can't freeze typing, hovering, or the other panes — even with several workspaces streaming at once. The idle CPU cost of a multi-workspace window dropped from "fans on" to ~0%.
• Long sessions reopen instantly. A restored chat shows the newest ten conversation turns; "Show earlier messages" reveals ten more per click instead of rendering the entire history at once.
• Select as much of an answer as you want. A new I-beam button on each answer switches it to one continuous text view, so a single drag can select any range — alongside the existing drag-select within paragraphs and code blocks.
• Select & copy from agent answers in UI mode. Assistant answer text is now drag-selectable — highlight any part and ⌘C (or right-click → Copy), the same way code blocks and thinking already were. The per-message Copy / Edit / Retry actions are now always shown in the message header (previously hover-only, which the new text selection hid).
• Graceful first-run & failure states. Opening an AI pane (Chat, Compare, review) with no API key now shows an upfront "Add your API key" card that deep-links straight to Settings → AI Providers, instead of failing only on send. Provider errors are mapped to plain language (offline, unreachable host, bad key → check Settings, rate-limited, server error) rather than raw URLError/HTTP dumps. The Review pane on a non-git folder offers a one-click "Initialize repository"; Settings → Agents shows an install hint + copy button for CLIs that aren't on PATH.
• Public-launch polish. The Claude pane header now shows a clickable account chip (your plan, email in the tooltip, one click to the Account panel — or a "Sign in" chip when the session drops auth). Pane-header and sidebar ⋯/✕/toggle controls got accessibility labels so VoiceOver announces them. Board cards no longer overflow their column at narrow pane widths (columns size to the pane, label/due pills truncate). Split-divider drags now persist across store updates instead of snapping back, and no longer accelerate mid-drag. Settings → About links out to the Privacy Policy and Terms.
• Security hardening for public release. The iCloud auto-clone path now validates git remotes against a transport allowlist (rejecting command-executing ext:: helpers and --option injection) and contains clone destinations strictly under the clone root (no ../ traversal); git clone also runs with protocol.ext.allow=never. Synced workspace environments are stripped of code-execution-hijacking variables (BASH_ENV, DYLD_INSERT_LIBRARIES, NODE_OPTIONS, …) on ingest. Release builds drop the disable-library-validation entitlement (kept Debug-only for hot-reload), so the shipped app enforces full Hardened Runtime library validation.
• Re-login without leaving Claude Code UI mode. An Account panel (options menu, or type /login) shows who you're signed in as and lets you sign in, switch accounts, or log out in place — claude auth login runs right in the pane, surfacing the sign-in link and taking the pasted code. When a turn fails because the session isn't authenticated, the pane now says so and opens the sign-in panel instead of a cryptic error, so there's no more dropping to Terminal mode just to re-authenticate.

Forge 0.9.1

2026-07-19

• Goals — Claude Code's native /goal, first-class. Type /goal in a Claude pane for a native panel of saved goal templates (one-tap Run) or a free-text verifiable end-state with an optional time budget. A goal bar tracks the run — elapsed, Stop, then met / not met — and .forge/goals.json + an inbox make goals agent-scriptable. Old Loops specs migrate to templates automatically; the Loops pane is retired.
• Services pane — a process manager for dev servers. Long-lived processes (Metro, APIs, docker compose) with health checks, restart policies, and dependency-ordered Start all; agent-scriptable via .forge/services.json + inbox.
• New workspaces open with your default agent — Claude Code out of the box, or pick any detected CLI (Settings → Agents).
• Workspace status you can trust. Working dots survive workspace switches and long tool runs; a pane genuinely waiting on you shows an orange "waiting for input"; finished checkmarks clear themselves once seen (and expire if never visited); interrupts no longer fire phantom "finished" notifications.
• Claude Code UI fixes. Transcript rows no longer vanish until you scroll; the idle 100% CPU spin is gone at the root; errors render as unmissable red alert cards; the worktree "Isolate" banner no longer hides the header row and its dismissal finally sticks.
• Model picker, finished. "Default" now names the actual model ("Default — Opus 4.8 (1M)"), and your pick is remembered per workspace across restarts.
• Sync fixes. BYO API keys actually roam via iCloud (both directions, stale copies cleaned); .forge/board.json refreshes after agents edit board.md directly.

Forge 0.8.1

2026-07-14

• Claude Code UI mode — a native chat for Claude Code — new Claude panes open as a real UI over the same claude session (the terminal stays one toggle away): markdown with syntax-highlighted code, collapsible thinking, tool rows with live diffs and checklists, permission cards with one-tap "Always allow", answerable multiple-choice questions, / command autocomplete, @ file mentions, and image attachments. Resumed panes rehydrate their full transcript from disk.
• Slash commands as real UI — /model, /effort, /config, /permissions, /mcp and /agents open native pickers and editable panels in headless mode, reading and writing Claude's on-disk config atomically.
• Ultracode effort tier — the effort picker now includes the CLI's session-only ultracode mode (xhigh reasoning + multi-agent workflow orchestration); applied as --effort ultracode on the session, never saved to settings.json.
• Skimmable transcripts — collapse any answer (or all at once) to a smart one-line summary with word and code-block counts, and jump to any prompt you've sent with the message navigator.
• Active Agents in the sidebar (optional) — every pane currently working or waiting across all workspaces, with live status dots and click-to-jump.
• Send Feedback + opt-in crash reporting — a feedback sheet with optional diagnostics, composed locally and sent through your mail client; unclean exits offer to send the crash report on next launch.