Changelog
New In Hoplite
Everything we ship, as we ship it — releases, improvements, and fixes.
- New
Send your local changes to a thread's sandbox
hoplite push, and /sync in an interactive session, send the work on your machine that isn't on GitHub yet (staged, unstaged, and untracked files) to a thread's sandbox. The changes apply on the same commit as your checkout and merge three ways with the sandbox's own changes, naming any conflicts. Credential files such as .env are never sent, and --dry-run lists what would be.
- New
Hoplite CLI 3.0: the coding agent in your terminal
The Hoplite CLI is now a full interactive client for Hoplite threads. Run hoplite in a linked repository and every prompt runs on a thread in your project's cloud sandbox, streamed back with tool calls, approvals, and pull request status. Script it with hoplite ask, reach the sandbox with hoplite exec and hoplite shell, and plug Hoplite into Zed or Paseo with hoplite acp. Install or upgrade with npm i -g @usehoplite/cli@latest or brew upgrade hoplite.
New
- An interactive terminal session on a Hoplite thread: queue or steer messages while a run works, answer approvals and questions, switch models, and see teammates who are in the same thread
- hoplite ask runs one request from a script, with plain text or --json output, and --resume to continue a thread
- hoplite exec runs a command in a thread's sandbox and exits with its status; hoplite shell opens an interactive shell there, like ssh
- hoplite acp runs Hoplite as a local ACP agent for Zed, Paseo, and other ACP clients
- hoplite login and hoplite logout, with a workspace choice in the browser; on macOS, keys are kept in the Keychain
- hoplite -c and hoplite -r resume your latest or a recent thread in the repository
- Slash commands for the thread's pull request (/pr, /merge, /autofix), previews, usage, rewind, and export
Upgrading from 2.x
- hoplite with no command now opens an interactive session; run hoplite onboard to import local agent history, and pass --headless or --org-id to onboard rather than to hoplite
- hoplite version is now hoplite --version
- You'll be asked to sign in once after upgrading: keys stored by 2.x lack the thread permissions the agent needs, so the first hoplite, hoplite ask, or hoplite acp signs you in again in the browser and stores a new key. In scripts, hoplite ask and hoplite exec stop with a message to run hoplite login once
- onboard, update, undo, handoff, skills, mcp, and auth work as before; hoplite completion is gone on macOS and Linux
- The interactive session, ask, exec, shell, status, and acp run on macOS and Linux; on Windows the CLI keeps the 2.x commands and adds login and logout
ACP endpoint
- API keys used over ACP need project read and thread create, read, update, and stop permissions, so a key can continue and cancel its runs; compacting or retrying also needs thread retry, and merging a pull request needs repository update, as in the web app
- Open ACP connections re-check their credential about once a minute, so an OAuth connection ends when its access token expires after an hour. Refresh the token and send it on the next request, or with _hoplite/auth/refresh over a WebSocket, to keep the connection; otherwise reconnect
- Improved
Cloud-connected local agents and smoother merges
Local-agent threads now publish to your organization by default, so the work on your Mac can stay visible and controllable from Hoplite everywhere else. GitHub merge queues are reflected in the PR rail, and GPT-5.6 Sol is now the default platform model.
New
- New local projects share their local-agent threads with the organization by default, so teammates can follow the work, send messages, approve tools, and stop runs while the desktop host executes them
- Pull requests accepted into a GitHub merge queue now show a Queued to merge state, so agents do not try to merge work GitHub already owns
Improvements
- GPT-5.6 Sol is the platform default, and Free members with a ChatGPT subscription can run Codex-supported GPT models
- Long tool runs collapse into readable activity summaries, making thread timelines easier to scan
- Improved
A smarter model picker and steadier integrations
The model picker now has search, hover tooltips, and clearer truncation for lookalike names. GPT-5.6 and GPT-5.5 gain paid 1M-context options, sandbox sizing moved to simple presets, and a stack of MCP and Linear reliability fixes landed alongside.
New
- 1M-context variants for GPT-5.6 Sol, Terra, and Luna, and for GPT-5.5, selectable right from the model picker
- Sandbox CPU and memory settings replaced with a single Box size preset dropdown
Improvements
- The model picker gained search, hover tooltips with pricing and context window, and smarter truncation for similar model names
- Adding an MCP server now tells you upfront when its OAuth setup can't complete, instead of saving a connection that fails silently at sign-in; connectable servers save first, then connect from the server list
- Linear lookups recover automatically from brief database or timeout hiccups instead of surfacing an error
- Improved
Prebuilds and warm previews, now on by default
Projects with a supported lockfile now bake dependencies and a warm preview ahead of time by default, so opening a workspace no longer starts from a cold install. Existing projects switch over automatically the next time their setup succeeds, and the setting can still be turned off per project.
New
- Prebuilds now default to on for projects with a supported lockfile; existing opt-outs are respected
Improvements
- Prebuilds warm a preview by default alongside dependencies, so the first preview doesn't start from a cold server
- Project settings show when prebuilds turned on automatically versus were explicitly enabled
- Improved
Mobile composer, activity, and login fixes
A round of mobile fixes: the threads composer collapses into a compact button so it doesn't crowd a docked list, run activity now renders correctly instead of dropping into ordinary chat, and a legal notice no longer overlaps the login form on small screens.
Improvements
- On phones, the threads composer collapses into a New thread button and expands on tap, keeping your draft and attachments intact
- Run activity now shows in collapsible sections on mobile instead of being dropped or shown as plain chat
Fixes
- Fixed the legal notice overlapping the email field on short mobile screens at login and sign-up
- Fixed duplicate "Worked for…" activity headers appearing after a run compacted mid-way
- Improved
Desktop beta: smoother updates and sign-in
The desktop app replaces its native update dialog with an in-app notification that installs now or waits until you're idle, and sign-in and file attachments are more reliable end to end.
New
- Desktop updates now surface as an in-app notification with "Update now" or "Update when idle", instead of a native dialog
Fixes
- Desktop sign-in surfaces a clear retry option if the app-to-browser handoff fails, instead of leaving the app silently signed out
- File attachments added by picker, paste, or drag-and-drop now upload reliably from the desktop app
- Improved
A steadier automatic PR review loop
Threads waiting on checks and review now resume reliably: a passing pull request with no outstanding feedback gets a final report instead of sitting parked, the loop wakes up when the target branch moves, and it recovers automatically from brief GitHub outages instead of stalling.
Improvements
- Threads parked on a passing pull request now get a final status report instead of silently going stale
- The review loop wakes up automatically when the target branch advances, so a rebase is no longer needed to get its attention
- Trusted review-bot feedback is accepted automatically while every command still requires a human
Fixes
- A brief GitHub API outage no longer strands a review loop; it retries and resumes on its own
- New
Muse Spark 1.3 joins the model lineup
Muse Spark 1.3 is now available across every plan, alongside Claude Fable 5.1. Free workspaces default to its Contributor route automatically, and your model selection carries over cleanly if you upgrade or downgrade.
New
- Muse Spark 1.3 added as a paid model option, alongside Claude Fable 5.1
- Free workspaces now default to the Muse Spark 1.3 Contributor route
Improvements
- A saved Muse Spark selection now maps automatically between the regular and Contributor routes when your plan changes
- New
Parallel agents, now in isolated sandboxes
We're beginning the move from parallel agents sharing one sandbox to each working in an isolated fork. For organizations included in the rollout, every subagent that needs to work on the code gets its own sandbox, so agents can build, test, and edit at the same time before Hoplite reconciles their completed work into the parent workspace. Organizations not yet included continue on the shared-workspace path while the rollout expands.
New
- For organizations included in the rollout, code-writing subagents run in isolated sandbox forks instead of sharing the parent agent's checkout
- The parent agent and its subagents can build, test, and edit concurrently without competing for files or processes
- Each subagent starts from the parent workspace's exact state, including its uncommitted work
- Completed changes are reconciled back into the parent workspace, with conflicts surfaced instead of silently overwriting work
- Remote MCP servers can now be connected through OAuth directly from Hoplite, with authentication discovered and configured in-product
- Authorized commands on GitHub pull requests can now invoke Hoplite, creating a durable agent thread and isolated workspace for the requested work
- Improved
Sturdier live previews
A stack of preview fixes: far more dev servers detected out of the box, slow starts no longer flap the preview between loading and error, and a crashed server shows you its output instead of a blank frame.
New
- Review checklist items can be un-checked; a stray tick is no longer permanent
Improvements
- Framework detection covers far more dev-server setups without configuration
- Previews wait out slow server starts instead of flapping offline
- When the dev server crashes, the preview surfaces its crash output
- Improved
Opening threads got a lot faster
Opening a thread no longer waits on a chain of REST round trips, because the sync layer already has the data. The JavaScript every visitor downloads also shrank by megabytes.
Improvements
- The thread-open path reads straight from the sync layer; its REST round trips are gone
- Thread queries warm as soon as the sync client connects
- 4.5 MB of syntax-highlighting grammars trimmed from the client build
- Another 149 KB (compressed) cut from the eager payload every visitor downloads
- hoplite.sh and the docs now serve from a global edge CDN
- New
Bring your own model gateway
Connect a custom model gateway and route threads through your own endpoints — your keys, your quotas, your bill. Once connected, a gateway model behaves like any other: pick it per thread, set its reasoning effort, and switch away whenever you like.
New
- Add OpenAI-compatible gateways from workspace settings
- Pick the gateway per thread, straight from the model picker
- Reasoning-effort control carries over to gateway models
Fixes
- The model picker no longer flashes the default model while a thread is loading
- Improved
Faster workspace wake-ups
Workspace restores now stream in far larger chunks, so big repositories wake in a fraction of the time. Warm sandboxes still resume in under a second — this release is about the cold path.
Improvements
- Restores stream in 4 MiB chunks — dramatically fewer round-trips on large repositories
- Setup logs stream live while the workspace wakes
- Wake time now shows in sandbox status
Fixes
- Provisioning no longer stalls on repositories with very large dependency trees
- Reconnecting mid-restore resumes the restore instead of restarting it
- Improved
Threads on your phone
Thread rows stack into two lines on small screens, so triaging a backlog from your phone no longer means horizontal scrolling.
Improvements
- Thread rows reflow into a two-line layout under 640px
- Tap targets across the thread list sized for thumbs
- Verified end to end at 390px
- New
Sandboxes that grow with the job
Sandboxes watch their own disk and scale it up before a big install or build runs out of room — no more mid-run failures on monorepos.
New
- Automatic disk growth when a run approaches capacity
- Disk usage surfaces in sandbox status, alongside CPU and memory
Fixes
- Large monorepo installs no longer fail mid-run on the default disk size
- New
Delegate straight from Linear
Assign an issue to Hoplite in Linear and an agent picks it up, works it in an isolated sandbox, and links the pull request back to the issue when it's done. Fan out as many as you like — every delegation gets its own machine.
New
- Assign an issue to the Hoplite app in Linear to start an agent
- The pull request links back to the issue automatically
- Progress updates post to the issue as the agent works
- New
Automations
Standing agents on a schedule: dependency bumps every morning, a test-flake sweep every hour — down to every minute on Scale.
New
- Create automations from a prompt and a schedule
- Every run opens a reviewable thread with its own sandbox
- Minute-level schedules on the Scale plan