Skip to content

Releases

v0.26.1 — Claude, Codex, and Chrome Fixes

Choose a tag to compare

@a-mart a-mart released this 30 Sep 19:01

v0.26.1 — Claude, Codex, and Chrome Fixes

Forge v0.26.1 is a patch release on top of v0.26.0.

Fixes

  • Claude subscription models: the Anthropic OAuth client version now matches the bundled Claude runtime (2.1.280), which fixes newer Claude models being rejected on the Pi subscription path.
  • Native Claude missing folder: if a session's working folder no longer exists, native Claude now reports the missing folder before launch instead of a misleading native-binary error.
  • External Chrome control: Forge now removes other extensions' frames that block debugger control (for example, Bitwarden-style autofill menus and focus-trapped fields) and retries attach/detach. A frame that keeps returning fails with a retryable error that names the blocking extension, and leases no longer get stuck.

Models

  • GPT-6.1 Sol is available for Codex native managers. GPT-6 Sol remains the default.

v0.26.0 — Native Managers, History Recall, and Browser Previews

Choose a tag to compare

@a-mart a-mart released this 28 Sep 21:26

v0.26.0 — Native Managers, History Recall, and Browser Previews

Forge v0.26.0 is the first stable release since v0.25.1. It adds native Codex and Claude managers, agent history recall, experimental Context v2, OpenRouter routing controls, automatic browser previews, side chats, and GPT-6 / Claude Opus 5.5, plus new recommended defaults and a broad set of reliability fixes.

Highlights

  • Native Codex managers run through the Codex app-server and keep the Forge sidebar, conversations, workers, history, and browser tools.
  • Native Claude managers run on the Claude Agent SDK with Claude subscription sign-in.
  • History recall lets agents browse, search, and read earlier messages and tool results.
  • Context v2 (experimental) continues long work from task notes and recoverable history instead of a summarizing pass.
  • Automatic browser previews float over Chat while an agent works in a browser tab.
  • /side chats fork a temporary side conversation without disturbing the main session.
  • New models: GPT-6 Sol, GPT-6 Luna, and Claude Opus 5.5.
  • Recommended defaults: GPT-6 Sol · High on Codex native, Hands-on work mode, and the shipped Default roster.

Native managers

  • Codex native: choose it in the manager model selector. Requires Codex CLI 0.155+ (or a desktop app that includes it) and an OpenAI/Codex account. Uses Forge's credential pool and its own Codex home.
  • Claude native: dedicated sign-in under Settings → Authentication; the Claude runtime ships bundled.
  • Both keep Forge tools, secure execution, progress, task notes, and session lifecycle. All Forge tools (including browser, history, notes, and secret requests) reach native Claude.
  • Native threads resume after stops and tool-set changes; late input carries into the next turn; asynchronous questions appear as Forge choice cards.
  • Background commands stay active and visible, and command waits respond to steering.
  • Native runtimes can serve as roster workers, and native usage now counts in Stats and the usage gauges.
  • Native managers run with full host access and no per-command approval prompts. Secret delivery uses forge.secure_bash with the nono executor (FORGE_SECURE_EXECUTION_BACKEND=nono).
  • Existing sessions never switch runtime automatically.

History and context

  • Agent-only history tool for local Builder managers and workers: browse context windows, list items, run ranked or literal search, and read canonical evidence.
  • Settings → History shows indexing progress and can pause or resume it. Secure Sessions secrets and restricted content stay excluded.
  • Project Settings → Context management chooses Summary (default) or experimental Context v2. Sessions can override the choice beside Send.
  • Session task notes (checkpoint.md and related files) survive compaction and restart and appear under Artifacts → Context v2.
  • Routine compaction and Context v2 transitions show as informational Context update rows.

Chat and workspace

  • /side [question] opens a temporary side chat in a resizable panel. Native sessions fork their full thread; side chats get read-only Forge tools and ask for confirmation before being discarded.
  • Drag a sidebar session into chat to reference it as a peer. Both managers can then message each other, including across projects.
  • Needs you now appears for an unanswered choice while work continues, and returns correctly after worker cleanup and in-flight input.
  • Inbox: Recent shows five rows with Show more / Show less, local rows get the full context menu, and project headers can be reordered by dragging.
  • Work-graph step inspector: select a Graph or List step to see its assignment, acceptance criteria, status, and attempts.
  • Stopped sessions recover when new user input arrives.

Browser

  • Automatic previews: up to four thumbnails float over Desktop Chat while agents use embedded or Chrome-backed tabs. Drag to move or resize from the corner; double-click an in-app card to open it in the Browser workspace. Previews are view-only and never saved.
  • New tabs and empty-session workspaces open embedded Forge tabs. Agents can request one with browser_open target='managed'.
  • Pi managers load browser tools on demand instead of attaching every schema to each turn.
  • External Chrome: profiles resume after a clean Desktop quit, stuck debugger leases release, control recovers when another extension's frame (such as a password-manager autofill menu) blocks the debugger, and Windows permission/identity setup is repaired.

Models and routing

  • GPT-6 Sol and GPT-6 Luna for Codex-backed managers and specialists. Saved GPT-5.6 selections migrate to their GPT-6 equivalents.
  • Claude Opus 5.5 is the Opus family default.
  • OpenRouter routing and privacy under Settings → Models: shared defaults plus per-model overrides for provider order or allowlist, exclusions, price ceilings, quantization, Zero data retention, and Block data collection. Hard filters fail closed.
  • Per-model OpenRouter context-window caps.
  • Credential pools: pause or resume individual accounts.
  • OpenAI weekly usage windows display correctly in the sidebar gauge.

Secure Sessions and secrets

  • Per-project switch: new projects ask whether to enable Secure Sessions (default off). Project Settings → Secure Sessions toggles it later without deleting saved secrets or trust.
  • Inherited project secrets: granting an alias to a project covers eligible agents in it. The chat shield can pause the task, block a worker, or revoke an inherited grant; blocks survive restart.
  • Agents continue after you approve or deny a secret request.
  • Bitwarden: Sync items now, locked-vault warnings, and lighter startup probes.

Work modes and defaults

  • Hands-on is the product default. Delegate first and Adaptive remain available.
  • The shipped Default roster pairs a Plan consultant, an Independent reviewer, and a Researcher for bounded consultations.
  • Existing Desktop installs see a one-time dialog: Use recommended defaults or Keep current defaults. Explicit session overrides are always preserved.
  • Adaptive workers can send requested interim findings and blockers to their manager.
  • Changing a project default model no longer rewrites existing sessions.

Fixes and improvements

  • Work-graph dispatch binds workers to attempts before delivery, and stale failures no longer overwrite newer state.
  • Runtime startup stays consistent when preparation fails late.
  • Long-running Pi bash commands yield, and bash_process can wait for, list, or stop them.
  • Chat copy works over non-secure LAN HTTP; scrollbars follow the app theme; secret dropdowns stay anchored.
  • Compaction no longer depends on the Secure Sessions runtime; exit code 125 is preserved.
  • Pooled Codex auth recovers after a successful usage validation.
  • Brave Search runs in packaged Desktop installs.
  • Stats history caching no longer blocks conversation loading.
  • Expo push notifications can authenticate from a private token file.

Changes to know

  • Default Desktop builds no longer bundle the Stream Deck plugin installer. Set FORGE_STREAM_DECK_SETUP_ENABLED=true when building to include it.
  • Forge's Pi compaction settings and Context v2 do not apply to native Codex or Claude managers.

Known limitations

  • Automatic Browser on Windows remains beta, and native Codex on Windows still needs live validation.
  • The nono secure executor is validated on macOS only; Windows uses the Docker path.
  • Context v2 is experimental and limited to eligible Pi Builder managers.
  • No Linux Desktop artifact is shipped.

Thanks

Thanks to Jerins71 for eight contributions, covering Chrome resume and recovery, native Codex stop handling, runtime reservation, Brave Search packaging, and test hardening.

v0.26.0-beta.5 — Native Claude Managers

Choose a tag to compare

@a-mart a-mart released this 24 Sep 18:06

v0.26.0-beta.5 — Native Claude Managers

Forge v0.26.0-beta.5 adds native Claude managers that use your Claude subscription, and fixes several Native Codex and browser-preview rough edges.

Native Claude managers

  • Native Claude managers. Create managers that run on the Claude Agent SDK runtime with Claude subscription sign-in, while keeping Forge tools, secure execution, progress, and session lifecycle.
  • Simpler sign-in. Settings → Authentication has a dedicated native Claude sign-in flow, and the bundled Claude runtime is updated.
  • All Forge tools reach Claude. Browser tools (click, type, navigate, resize), history, notes, and secret-access requests are now visible to native Claude managers. Pi Anthropic models also receive the full browser-tool parameters.

Native Codex

  • Codex threads resume correctly after Forge's tool set changes.
  • Codex's asynchronous questions now appear as Forge choice cards.

Browser previews

  • Resize the floating browser preview stack by dragging its corner.

Fixes & Improvements

  • Forge's Expo push notifications can authenticate with a private token file.
  • Removed low-value internal tests and dead exports; no user-facing behavior change.

Known limitations

  • Automatic Browser features on Windows remain beta.
  • No Linux Desktop release artifact is shipped.

v0.26.0-beta.4 — New Models and Runtime Reliability

Choose a tag to compare

@a-mart a-mart released this 22 Sep 20:43

v0.26.0-beta.4 — New Models and Runtime Reliability

This beta adds GPT-6 Sol, GPT-6 Luna, and Claude Opus 5.5, updates Forge's model defaults, and fixes several cases where native Codex and work graph sessions could lose or mishandle work.

Models

  • GPT-6 Sol and Luna: Both are available for Codex-backed managers and specialists. Native Codex now defaults to GPT-6 Sol.
  • Claude Opus 5.5: Available in the model catalog and selected by default for the Opus family.
  • Saved model migration: GPT-5.6 selections are mapped to the corresponding GPT-6 models, including the low-reasoning replacement for Terra.

Session and Work Graph Reliability

  • Native Codex continuation: Input arriving near the end of a turn is carried into the next turn, and a stopped runtime releases its thread writer so the session can resume.
  • Work graph dispatch: Workers are bound to their graph attempts before assignments are delivered. Late failures and cancelled attempts no longer overwrite newer state.
  • Runtime startup: Runtime reservations remain consistent if preparation or attachment fails late.
  • Secure Sessions: Project policy now reaches native runtimes, and disabling Secure Sessions updates their guidance as well as access checks.

Fixes & Improvements

  • Updated model selection and specialist defaults for the new catalog.
  • Improved documentation for empty Browser workspaces and managed browser opening.

v0.26.0-beta.3 — Native Codex, Hands-on Defaults, and Browser Previews

Choose a tag to compare

@a-mart a-mart released this 20 Sep 21:57

v0.26.0-beta.3 — Native Codex, Hands-on Defaults, and Browser Previews

Forge v0.26.0-beta.3 follows v0.26.0-beta.2. It adds a Codex native manager runtime, makes Hands-on the product default with a shipped Default roster, adds automatic floating browser previews on Desktop Chat, a project-level Secure Sessions switch, and pause/resume for pooled accounts.

OpenRouter routing, inbox Needs you for unanswered choices, history recall, experimental Context v2, inherited project secrets, Adaptive worker feedback, and the work-graph inspector from earlier 0.26 betas remain in this train.

Treat this as a validation build, not a stable promotion.

Native Codex managers

Choose Codex native in the manager model selector to run a local Builder manager through the Codex app-server. The family is labeled Preferred. Existing Codex choices continue to use Pi. Native managers keep the Forge sidebar, conversations, task notes, history, workers, and browser integration tools. No existing session changes runtime automatically.

  • Install Codex CLI 0.155 or newer, or a desktop app that includes that version. Configure an OpenAI/Codex account in Authentication. Native managers honor Forge's credential pool and Forge Auth broker mode.
  • On macOS, Forge looks in ChatGPT.app or Codex.app, then PATH. On Windows, it launches codex.exe from PATH or a standard npm install — not .cmd / PowerShell wrappers. Set CODEX_BIN to a native codex.exe for a custom install. Windows execution still needs live validation.
  • If the CLI is missing or too old, the error includes the install or update command for that OS. Restart Forge after installing so PATH changes are picked up.
  • Native data lives under Forge's own Codex home, separate from the desktop Codex app. Workers, Collaboration, and system sessions stay on their existing runtimes.
  • Native managers run with full host access and without per-command approval prompts. Genuine clarification questions still use Forge's choice UI.
  • Hands-on is now the product default, so the manager keeps the critical path. Use Delegate first when workers should own substantive project work, or Adaptive when the manager should start directly and hand off only when the total path is better.

Native Codex can use saved secrets through forge.secure_bash when the project has Secure Sessions enabled. Ordinary Codex shell and file tools are not intercepted — do not write credentials into the workspace and then read them with those tools. Secret delivery for native managers uses the nono executor (FORGE_SECURE_EXECUTION_BACKEND=nono); Docker remains the default when that setting is omitted. Nono was exercised on macOS; Linux still needs platform acceptance, and native Windows stays on the existing Docker path. Start a new session or fork to pick up the secure tools on a thread created before this beta. Browser login delivery is outside this integration.

Forge's Pi compaction settings and Context v2 do not control native Codex compaction.

Recommended defaults

New installations default to:

  • Manager model GPT-5.6 Sol · High on Codex native
  • Work mode Hands-on
  • Roster Default — Plan consultant (openai-codex/gpt-6-astra, Extra High), Independent reviewer (anthropic/claude-fable-5-1, Low), and Researcher (xai/grok-4.6, High)

The manager owns implementation and integration. Specialists answer bounded planning questions, review a concrete risk, or do source-backed research. Worker results still return automatically. Accept by integrating and continuing; a thanks or acceptance message to a completed worker starts another assignment.

Existing Desktop installs keep their current project defaults until the one-time post-update dialog. Use recommended defaults updates user projects and preserves explicit session model, work-mode, and roster overrides. Keep current defaults leaves those project settings unchanged. New sessions still copy the current project default; existing sessions are not rewritten.

The earlier Roster actions → Add hands-on support roster helper is removed. Balanced remains available as a compatibility roster derived from stored tier bindings.

Automatic browser previews

When a local Builder agent starts using an exact embedded or Chrome-backed tab, Forge Desktop places a small thumbnail over Chat. There is no Preview button and no separate Preview window. Up to four thumbnails form one compact, frameless stack.

  • Drag the cluster to move it. Click a card to bring it forward. Double-click a managed (in-app) card to open it in the Browser workspace.
  • Opening the Browser workspace hides the stack without dropping it. Chrome thumbnails appear only after a successful snapshot supplies real pixels; they are not live and do not attach the debugger.
  • The stack does not accept page clicks, typing, scrolling, refresh, recording, or Take Control. Preview images stay in memory and are not saved to history, artifacts, Remote Projects, or Collaboration.

New browser tab and empty-session workspace opens create embedded Forge tabs. The workspace can stay empty until you or the agent open a tab. Agents can request an in-app tab with browser_open target='managed'. Chrome-backed tabs stay in Chrome and are not shown as Browser workspace cards.

Eligible Pi managers load Automatic Browser tools on demand (discover_tools with the browser bundle) instead of attaching the full browser_* schemas on every turn.

Secure Sessions per project

New local projects ask whether to Enable Secure Sessions and default to off. Project Settings → Secure Sessions can turn it on or off later. Enabling permits the existing grant flow; it does not grant secrets by itself. Existing projects without the setting keep their previous enabled behavior.

Turning it off removes secure tools from that project's agents, stops its secure environments, and keeps saved secrets, SSH host trust, and grant mappings for later.

After you approve or deny a secret-access request, the waiting agent can continue instead of stalling. Bitwarden Password Manager can Sync items now so catalog entries match the current collection items, not only the saved collection list.

Inbox and accounts

Drag local and connected remote project headers in Inbox the same way as in Projects. The shared order is a local display preference.

Settings credential pools can Pause or Resume an account. Paused accounts are skipped when routing requests. If every pooled credential is paused, Forge fails closed instead of using a paused singleton.

Fixes & Improvements

  • Rooms worker rows indent under their session in the sidebar.
  • Long-running Pi bash commands yield so the manager can keep working; bash_process waits, lists, or stops that command.
  • Native Codex progress appears between tool calls instead of waiting for the final message.
  • Managed browser tabs initialize before automation, and resize keeps the chosen viewport.
  • External Chrome recovers more reliably across restart and lifecycle gaps.
  • Context v2 continuation prefers the current checkpoint.md and recorded decisions before broadening history search.
  • Stream Deck plugin preparation is opt-in (FORGE_STREAM_DECK_SETUP_ENABLED=true) for Electron development and packaging; default Desktop builds no longer bundle the installer.

v0.26.0-beta.2 — OpenRouter Routing and Inbox Attention

Choose a tag to compare

@a-mart a-mart released this 11 Sep 20:00

v0.26.0-beta.2 — OpenRouter Routing and Inbox Attention

Forge v0.26.0-beta.2 follows v0.26.0-beta.1. It adds OpenRouter provider routing and privacy controls, keeps unanswered choices visible in Needs you while work continues, and fixes Chrome resume, usage gauges, and packaged Brave Search.

History recall, experimental Context v2, inherited project secrets, Adaptive worker feedback, and the work-graph inspector from beta.1 remain in this train.

OpenRouter routing and privacy

Settings → Models stores shared OpenRouter provider routing and privacy filters on the selected backend. Shared OpenRouter defaults apply to every OpenRouter manager and worker call, including models not added yet. Configure routing on an added model stores an exact-model override. On a Collaboration server this is admin-only.

The dialog leads with Privacy and Providers. Zero data retention and Block data collection are separate filters and a privacy floor that a per-model override cannot weaken. Provider mode can be automatic, preferred order, or a hard allowlist, with optional exclusions, price ceilings, quantizations, and endpoint fallback.

A saved change applies to the next OpenRouter call. Hard filters fail closed and suppress Forge automatic model fallback; preferred order or sort alone do not. These filters constrain OpenRouter inference routing only — not local transcripts, tools, or non-OpenRouter compaction. Endpoint discovery is advisory.

Added OpenRouter models also support a per-model Context window cap. The effective limit never exceeds the model default; changes apply when an agent next starts, while running agents keep their current limit. Reset restores the default.

Inbox attention

An unanswered Input requested pending choice raises Needs you even while work still runs or follow-ups are queued. Answering or cancelling retracts that item without marking the session complete, so remaining work can stay Active. Unread badges still do not create Needs you. Input consumed during an already-running manager turn now clears its continuation fence correctly, preventing sessions from getting stuck after in-flight input.

Fixes & Improvements

  • Routine compaction and Context v2 transitions appear as informational Context update rows instead of amber system warnings.
  • Chrome profiles resume after a clean Desktop quit.
  • Hidden Chrome tab URLs are proven before status is serialized; disconnect errors are consumed before native-port ownership checks.
  • Sidebar controls stay beneath the sticky Needs you panel.
  • OpenAI weekly primary usage windows classify correctly in the sidebar gauge.
  • Brave Search skill dependencies are staged beside the packaged skill so Desktop installs can run it.

v0.26.0-beta.1 — Context v2 and History Recall

Choose a tag to compare

@a-mart a-mart released this 08 Sep 20:39

v0.26.0-beta.1 — Context v2 and History Recall

Forge v0.26.0-beta.1 is a prerelease after stable v0.25.1. Managers can recover earlier conversation evidence, continue long work from task notes instead of a summarizing pass, inherit project secret access with durable controls, and inspect work-graph steps in place.

History recall

Local Builder managers and ordinary workers can recover earlier messages and tool results with the agent-only history tool: browse context windows, list items without guessing a search term, run ranked or literal search, and read canonical evidence.

Indexing starts after local Builder hydration. Settings → History shows progress and can pause or resume indexing. Pause does not stop conversation recording or direct same-session recovery. There is no human history drawer and no embedding index.

History recall is local Builder and ordinary-worker only. Remote Projects search the selected remote origin, not the viewing client. Collaboration and Cortex do not get the history tool. Restricted runtime content and Secure Sessions secrets stay excluded.

Context v2

Project Settings → Context management chooses Summary (default) or experimental Context v2. Eligible local Builder managers can inherit or override that choice beside Send. Saving a mode does not clear the current conversation; it applies at the next Compact, Smart compact, or automatic transition.

Context v2 continues from task notes and recoverable history instead of a Smart LLM summarizing pass. Compact and Smart compact follow the supported policy. Settings → General → Compaction still chooses only the Summary compaction model, reasoning, and timeout.

Context v2 requires a supported ordinary Pi Builder manager (OpenAI/Codex or Anthropic) and an eligible model. Collaboration, Cortex/system, Cursor SDK, plugin/external threads, and workers use Summary.

Agents can keep session-local task notes (checkpoint.md and related working files) through context transitions and restart without promoting them into profile memory. Artifacts shows a Context v2 section for the selected manager’s task notes and window-recovery files.

Inherited project secrets

Granting a saved alias to a project installs protected access for eligible agents in that task. The chat shield can pause the task, block a worker, or revoke an inherited grant so it cannot silently regrant until restored. Blocks survive restart. Forks copy task pauses and blocked secrets; temporary grants are not copied.

Adaptive ownership and work graphs

Adaptive still starts directly and delegates only when the full path improves. Workers can send requested interim answers, blockers, or approach-changing findings to the owning manager without duplicating the final result. After ten minutes on a running assignment, Forge may send one internal ownership-review notice so the manager can continue, narrow, redirect, or take over.

Select a Graph or List step to inspect assignment, acceptance criteria, authoritative status, and execution attempts. A worker succeeding is still awaiting review until the manager accepts.

Inbox

Recent can show five rows plus Show N more, then Show less. Local Inbox session rows get the same context menu as Projects; remote Inbox rows stay without those mutating actions.

Fixes & Improvements

  • Dismissing a redacted-output warning resets the secure pill and popover.
  • Compaction no longer depends on the Secure Sessions runtime.
  • Shared Secure Sessions runtimes preserve command exit code 125.
  • Secret storage dropdowns stay anchored to their triggers.
  • Chat copy buttons work in the web browser on non-secure HTTP (LAN) by falling back when navigator.clipboard is blocked.
  • Native scrollbars follow the app theme.
  • Changing a project default model no longer rewrites existing sessions. Those sessions keep their stored model; Use Project Default applies the current default to that session only. Later default changes affect new conversations.
  • Sessions can return to Needs you after worker cleanup.
  • Pi workers use the same explicit system prompt as managers, avoiding an Anthropic-rejected documentation block.
  • Granted Bitwarden vaults warn when locked. Bitwarden metadata reads stay off the secure startup queue, and redundant sync/probes are skipped.
  • Windows Chrome setup repairs permission lifecycle and package identity, and Chrome compatibility is no longer tied to Desktop version numbers.
  • Context v2 budgets respect capped context windows.
  • Stats history caching no longer blocks conversation loading.
  • Pooled OpenAI/Codex auth recovers after a successful usage validation instead of staying marked as an auth error.

v0.25.1 — macOS Startup Fix

Choose a tag to compare

@a-mart a-mart released this 04 Sep 23:52

v0.25.1 — macOS Startup Fix

Fixes a macOS startup crash caused by incomplete UI code in the 0.25.0 package. Rebuilt from isolated release source.

Source: clean v0.25.0 tag (b00f86889e1336c63b314e344552e09aac347d2d) plus frozen-release workflow/test/documentation support and the two version-only commits used for beta-first validation and stable promotion. No post-tag product or feature commits are included.

v0.25.1-beta.1 — macOS Startup Fix

Pre-release

Choose a tag to compare

@a-mart a-mart released this 04 Sep 22:58

v0.25.1-beta.1 — macOS Startup Fix

Fixes a macOS startup crash caused by incomplete UI code in the 0.25.0 package. Rebuilt from isolated release source.

Source: clean v0.25.0 tag (b00f86889e1336c63b314e344552e09aac347d2d) plus frozen-release workflow/test/documentation support and the 0.25.1-beta.1 version-only commit. No post-tag product or feature commits are included.

v0.25.0 — GPT-6 Astra and Adaptive Work

Choose a tag to compare

@a-mart a-mart released this 04 Sep 20:25

v0.25.0 — GPT-6 Astra and Adaptive Work

Forge v0.25.0 adds GPT-6 Astra as a new Codex family, introduces an Adaptive work posture and specialist Rosters, brings PDF previews into the artifact panel, and expands Source Control and Bitwarden workflows. It also tightens Secure Sessions behavior and removes retired GPT-5.4 and Grok-era selections.

Highlights

  • GPT-6 Astra. The new pi-6 family is available for managers, compaction, specialists, and worker spawns, with low, medium, high, xhigh, and max reasoning. High is the family default; GPT-5.5 remains Forge’s global default. Astra has a 1.05M-token context window and 128k maximum output.
  • Adaptive work and Rosters. Adaptive joins Delegate first and Hands-on, choosing direct work or delegation outcome by outcome. Specialist configuration is now presented consistently as Rosters, including project and session selection and the built-in roster-management skill.
  • PDF artifacts in context. Assistant-presented local PDFs can open directly in the artifact panel with page navigation, zoom, fit, and reload controls. Preview reads use one-use authorization tickets and are capped at 16 MiB. Forge Desktop can also open the PDF in the operating system’s default app.
  • Publish a local branch. Source Control now offers Publish Branch when the current branch has no upstream. Forge fetches origin, refuses to overwrite a remote branch that has commits missing locally, pushes without force, and sets origin/<branch> as the upstream.
  • Create Bitwarden Logins from private entry. A saved private value can be written directly to a selected Bitwarden Password Manager collection as a Login item. The entry flow supports username metadata plus cryptographically generated passwords and passphrases. Configured locked vaults can prompt for unlock at launch or when Team Secure Mode needs them.
  • SuperGrok usage in the sidebar. A stored xAI OAuth credential can populate a weekly-only xAI credits gauge with reset timing. API-key-only and environment-only xAI setups do not provide this gauge, and transient polling failures retain the last good reading.

Security and reliability

  • Structurally valid multiline OpenSSH private keys with CRLF or CR line endings are normalized only in the copy passed to ssh-add; saved bytes and non-matching material remain unchanged.
  • Matching manual approvals and automatic project grants reuse an equivalent active lease instead of creating duplicate leases or false conflicts.
  • State-changing secret-settings requests now require an authorized secure-control path.
  • Claude OAuth compatibility now advertises client version 2.1.259 instead of 2.1.75.
  • Opening a local PDF in the default app now returns a structured failure if the file disappears or becomes unreadable during header inspection, instead of rejecting the IPC call.
  • Publishing a local branch still reports success when the remote branch is created even if local upstream tracking cannot be set. Forge keeps origin/<branch> as the published destination and surfaces the tracking failure as a warning.
  • Work Graph cards now remain opaque so connector lines stay behind them.

Migrations and compatibility

  • GPT-5.4, GPT-5.4 Mini, Grok 4, Grok 4 Fast, and the retired Grok 4.20 variants are no longer offered. Persisted legacy selections load through current replacements: GPT-5.5 for retired GPT models and the current pi-grok default for retired native Grok models.
  • Existing roster IDs and saved roster settings remain compatible; Rosters replaces the former “delegation preset” product language rather than introducing a second configuration system.
  • The installed Pi catalog does not yet include GPT-6 Astra. Forge projects the official model ID from its checked-in catalog, and actual availability still depends on OpenAI enabling Astra for the account.
  • The xAI gauge has no session window or historical pace curve. PDF previews remain bounded to 16 MiB, and opening a local PDF in the default app is a Forge Desktop capability.

Attribution

  • All changes in this release are by a-mart.
  • Multiline OpenSSH private-key support: a-mart, PR #7.