New features
- Daytona sandbox provider. Run agents on Daytona alongside the existing sandbox options, including from the opencode inline harness. See Architecture.
- S3 artifact storage. Files generated by an agent are now uploaded to artifact storage and surfaced in the session view, with support for custom S3 endpoints like Cloudflare R2.
upload_artifactMCP tool. Agents can publish files from the sandbox to artifact storage and share a hostable link in the session.read_filesandbox tool. Pull files out of the sandbox without dropping into a shell.setup.shsandbox convention. Drop asetup.shat the root of a connected repo and the sandbox runs it on provision. Edit it in the UI or have an agent update it via the newlap_update_sandbox_setuptool.- opencode inline harness. A new shared, inline opencode harness ships as a first-class option in the harness picker — no per-agent pod required. Skills, MCP tools, system prompts, and memory all carry over.
- Inline LiteLLM MCP servers. Linear, Slack, and GitHub MCP servers are wired into the inline harness out of the box.
- E2B sandbox tools over MCP. Inline-harness agents can provision and execute against E2B sandboxes through stdio MCP.
- Session-scoped memory for the inline harness.
search_memoryandsave_memoryare available in inline sessions, with secret scrubbing on save. See Authentication. - Loadable skills in opencode. Skills attached to an agent now load in both pod and inline opencode sessions.
report_issueMCP tool. Agents can file an issue directly from a session when something goes wrong, with the session linked for context.- Inspector Context tab. A new Context tab in the Inspector shows the MCP servers, tools, and skills loaded for the current session.
- Reviewer agent. A reviewer now runs as a real LAP agent session and produces proactive assessments and log analysis on the sessions it watches.
ghCLI in the sandbox. The GitHub CLI is preinstalled in the sandbox template so agents can open PRs, comment on issues, and check CI without extra setup.- Per-agent egress allowlist. Each agent can declare the hosts it needs to reach, and the vault credential swap is scoped to that host list.
Updates
- Reviewer surface is opt-in. The reviewer panel and inline banner are hidden by default — turn them on per workspace when you want them.
setup.shreplaces the oldsandbox_filesflow. Existing sandbox-file configs are migrated to the new convention.- Longer command timeout. Per-command sandbox timeout raised from 120s to 180s so UI screenshot and build steps don’t get killed mid-run.
- Autonomous opencode. The inline opencode agent no longer prompts for confirmation on
bash,edit, or follow-up questions inside the sandbox.
Bug fixes
- Fixed the sessions list polling at ~65 MB per request — rows are capped and history is no longer pulled into the list view.
- Fixed agents losing progress across rolling deploys — the harness now drains on SIGTERM, retries fetches, recovers sessions, and persists state to SQLite.
- Fixed inline-harness turns not persisting until completion — turns are now checkpointed to the database every 15 seconds.
- Fixed Report issue failing when the agent ID wasn’t known — it now uses the session-scoped endpoint, resolves the agent name, and retries transient errors.
- Fixed E2B sandboxes timing out around the 15-minute mark, which reset agent progress mid-task.
- Fixed opencode chats freezing after ~10 minutes — the event stream now reconnects automatically.
- Fixed inline opencode chats stuck on thinking… and tightened skill slug handling.
- Fixed agents hallucinating session URLs — the real session link is now injected into the prompt.
- Fixed
provisionSandboxerroring when a sandbox of the same name already existed. - Fixed system prompts referencing the wrong opencode tool name for issue reporting.
New features
- REPL mode for Claude Agent SDK agents. Talk to a running agent from your terminal with
lap <agent>— a single streaming surface over the same session you’d see in the UI. See the Claude Managed Agents quickstart. - Interrupt a running turn. Stop a long-running agent turn from the session view without killing the sandbox, then send a follow-up message in the same thread.
- Image attachments in chat. Paste screenshots, drag-and-drop images, or attach files to the chat composer. Agents receive them as part of the turn.
- Screen recording for every agent. Claude Agent SDK agents now ship with built-in screen recording tools, so agents can capture what they did in the sandbox.
- Live preview button. Coding-agent sessions expose a View Preview button that opens the running app from inside the sandbox.
- Lazy sandboxes. The agent brain now runs in the platform and only spins up a sandbox pod when one is actually needed — faster first turn, cheaper idle agents.
- Slack thread streaming. Slack-triggered sessions stream progress live back into the Slack thread, with a single View session button linking to the full session in the UI.
- Connected repo on agent detail. Each agent’s detail page now shows the GitHub repo it’s wired up to.
- Admin toggle for the inline harness. Workspace admins can enable or disable the inline harness from settings.
- Extended thinking for opencode. Toggle extended thinking per model when running agents on the opencode harness.
Updates
- Session UI overhaul. Thinking blocks render inline, tool calls are grouped, the Tasks sidebar updates as the agent works, and session rows in the sidebar are labeled with the first user message instead of a short ID.
- Slack-triggered sessions now paint in any open tab without a refresh, and prior thread messages are pulled in when an agent is
@mentionedmid-thread. - Helm-style template versioning. Agent templates are now versioned, and existing agents can resync to the latest template version.
- Unified agent form. Creating and editing an agent now use the same form fields, so options stay consistent across both flows.
- Environment variables on agent detail. Configured env vars are shown as key/value pairs on the agent detail page.
Bug fixes
- Fixed memory tool calls returning 401 —
search_memoryandsave_memorynow route through the credential layer correctly. See Authentication. - Fixed session threads losing state across navigation — open sessions now persist as you move around the app.
- Fixed user messages occasionally rendering blank or out of order in the session view.
- Fixed double scrollbars in long session threads — one scroll for the whole thread, not one per message.
- Fixed Open in LAP links from integrations pointing at the agent page instead of the session.
- Fixed a unique-constraint crash when spawning a new session from an integration.
- Fixed a 500 on message send when the database was briefly unavailable — now returns a retryable 503.
- Fixed the sandbox MCP provision route returning 404 in some workspaces.
- Fixed agent session heartbeats during the first turn so long initial tasks don’t get marked stale.