Repository navigation
What's Changed
Features
- New Lightpanda browser skill. A local headless browser is driven from nine lazy tools (
lightpanda_fetch,_read,_act,_extract,_screenshot,_run,_mcp,_status,_stop), so an agent can read, click, script and screenshot pages without a browser baked into the image. The skill installs its own binary, ships an SSRF guard, a telemetry opt-out, an MCP server and per-agent session handling, andevonic doctorgained a generic skill check/fix hook (backend/skill_doctor.py) reported next to the plugin checks. Requirement checks acceptsearch_paths, so a skill-local binary resolves without being onPATH. - Lightpanda cookies from a
cookies.txtdump. The skill settings page holds one textarea for a Netscape cookies dump — the format every browser cookie-export extension writes — and materializes it as a0600jar. A probe showed Lightpanda already filters the jar per host, so one jar from the whole dump is safe and simpler than one jar per site: unrelated sites cannot mix, while same-site subdomains and redirects keep the session.lightpanda_cookiesmanages status, set (with merge), delete, clear and verify, and never echoes a cookie value back;lightpanda_fetchreports only the cookie source, count and names. An earlier per-site profile editor was removed for this. - 3D knowledge graph. The agent knowledge graph is rendered with WebGL (three.js) instead of d3 and SVG, so it stays responsive with tens of thousands of nodes: the force layout runs in a Web Worker, nodes and edges are single draw calls, and distance-based level of detail hides lightly referenced nodes while heavily referenced ones are drawn larger. The graph orbits slowly by default; clicking a node or a search result stops the orbit, flies the camera to it and shows a balloon with its title and description, and clicking the balloon opens the editor. Thumbnail nodes are ordered by depth, free orbiting is unrestricted, and the native search-cancel control no longer shows through the graph search box.
recall(mode="paths")returns the directories an agent has actually worked in, so it looks a path up instead of searching the filesystem again. File tools,tree/statsandbashcdcalls are logged per agent infrequent_dirswith no LLM call.- Patch and diff files are first-class.
.patchand.diffattachments are accepted in web chat, render as a highlighted diff viewer (per-file cards, old/new line gutters, word-level change highlighting) in the artifact modal and the Agent Artifacts tab, and str-replace and diff-type tool parameters render with the same viewer inside chat bubbles, falling back to plain highlighting when the text has no hunk headers. - Plugins and skills can be updated from a zip with a diff review. A newer zip is previewed as a per-file diff, confirmed in a second dialog and applied with automatic rollback if it fails; a skill update keeps its saved
config.json. Zip validation ignores junk (.pytest_cache,__pycache__,.DS_Store,*.bak-*) instead of rejecting the upload, plugin template directories are resolved dynamically so a plugin installed after startup no longer raisesTemplateNotFound, and plugin zips may ship compiled handler, routes and tool module files. - MCP over Streamable HTTP. A server entry in
MCP_SERVERStakes{url, headers}instead of{command, ...}: messages are POSTed with the configured headers, theMcp-Session-Idreturned byinitializeis reused and re-established once when it expires (404), replies may be JSON or an SSE stream with pings answered, and the stdio transport is unchanged. A browser User-Agent is sent on HTTP transports so a Cloudflare-fronted server no longer answers 1010, large tool listings are kept whole, and a newmcp_list_serverstool lets an agent see which servers are configured. - Reasoning effort per model and per agent. Models store a
reasoning_effortvalue from the model modal and forward it on every call, and an agent can override it — for example one DeepSeek agent atmaxwhile the rest keephigh— without cloning the model row. Precedence is agent override, then model value, then key absent; values are normalised case-insensitively, an empty string or null clears the override, and agent portability (clone, export, import) carries it. - Evonet join-token auto-registration. An Evonet client can self-enroll with a global join token instead of a one-time pairing code: the server exposes
/api/connector/join, gated by a settings toggle and the token, and creates or re-binds a workplace keyed by a persistent device UUID. The Go client gained anevonet joinsubcommand, a device-id config helper and an install script; Evonet is now 1.3.0. - Connection test for decision models. Decision models cannot be pinged with a chat prompt and had no quick check, so their card and the model form gained a test button. The provider test endpoint accepts an optional
base_url, so a model's connection override can be tested before saving, and a native SystemOne decision model no longer fails the test. - Podman sandbox backend.
SANDBOX_BACKEND=podmanruns agentbash/runpyin a rootless Podman container: the image defaults tolocalhost/evonic-sandbox:latest(SANDBOX_PODMAN_IMAGEoverrides it),--userns=keep-idwith--security-opt label=disablekeeps the bind-mounted workspace writable without SELinux labels, and--restart=noreplaces the restart policy rootless Podman rejects, with the startup sweep restarting persistent containers. Selection, cleanup, prompts and pool guards are runtime-neutral,evonic doctorchecks the runtime the backend actually selects, and the backend is documented and covered by claim-independent tests. - Agent Panel drag-and-drop reordering (#922). Every button in edit mode gains a grip handle; dragging it to a new slot, or
Alt+↑/Alt+↓on a focused button, saves the new order immediately, so it survives a reload and reaches every open tab. - A
web_fetchtool returns a public web page as readable plain text, defined bytools/web_fetch.jsonand the regular tool module. - Setup seeds a capable super agent (#917). Both creation paths (the setup wizard/CLI and the legacy
/api/setupbranch) apply one shared default so the platform's default operator accepts images, voice notes and attachments from the first login.
Improvements
- The login page now says Password not configured and shows a short quick guide (
./evonic pass→ choose a password of at least 6 characters →./evonic restart→ reload and sign in) wheneverADMIN_PASSWORD_HASHis empty, instead of rendering a form whose every submit failed. On a fresh install it also links to the setup wizard, and the password field, show-password button and Sign In are disabled in that state. - Skill tools can return images: a tool adds
"images": [{mime, data, label}]to its result, each image is saved as a session attachment and replaced in the result by a reference, and a vision-capable model with vision enabled sees the images on its next call within the same turn. Otherwise the agent receives the file path fordescribe_image; PNG, JPEG and WebP up to 4 MB, at most four per result. - Model pickers list only the model name. Renaming a model left the id — frozen at creation time — in the label, so a row could show a stale registry path next to its own provider entry.
- System Settings uses the full page width instead of a narrow column.
- A channel send now reports what became of it.
BaseChannel.deliverandAgentRuntime.deliver_as_botreturn sent, queued, failed or unknown — unknown meaning a timeout after the message was handed over, which may have arrived and must not be re-sent unasked — andPluginSDK.send_messageno longer reports success when the channel refused the message or was not running.send_as_botkeeps its boolean for the admin takeover route, and a timeout with an empty message is no longer stored as no error. - Plugin export skips
.gitandnode_modules, so an exported zip no longer carries a repository or a dependency tree. test_resultsis indexed byrun_id, which the per-run report queries filter on.- The LLM client defaults the HTTP read timeout to 900 seconds when a model sets none, so a slow long-context answer no longer fails on a library default.
- Claude CLI: each image is shown where it is mentioned and pictures are no longer described as unreadable when the CLI can read them.
- Fixtures, tests and docs no longer carry real brand names, personal names or infrastructure identifiers, and the deprecated
auto_improverplugin is gone.
Bug Fixes
- The connection-test toast no longer prints the probe summary, which leaked the resolved endpoint and contract details into a transient notification. It shows a plain result message and rounded latency, and the full probe output stays in the connection-test modal.
- Typing in the agent chat box no longer slows down as the session grows. Every keystroke flipped the textarea to
height:autoand readscrollHeight, forcing a synchronous layout of the whole page (about 11 ms per key in a short chat, 40-65 ms with ~800K tokens of history). The height is now measured on an off-screen mirror and written only when it changes, the eight:has(#chat-input.chat-bash-input)rules became plain.is-bashselectors with the same specificity, and the "needs input" ring animates onlytransform. - An agent working a kanban task on autopilot no longer receives
ask_user, so a question cannot stall an unattended run for good. Plugins hide a tool per agent through a new tool-suppressor hook (register_tool_suppressor); it is checked when the tool list is built and again when the tool is called. Autopilot OFF is unchanged. - Native image input now respects the agent's Vision enabled switch, not only the model's Vision Supported flag, and the "Use the
describe_imagetool" line is never emitted for an image that is already attached — including messages injected mid-turn. Non-vision models and attachments older than the last three user turns still usedescribe_image. - Intermediate responses appear in chat again. The display had been implemented directly in the generated
chat-ui.jsbundle, so the next rebuild fromstatic/js/chat-ui/*.jsdroppedsplitForIntermediate()while both pages still called it, and theresponse_chunk(is_final=false)handler threw before the live text reached the chat. - Disabled agents no longer appear in the
Ctrl+Gquick search. - Login hardening: an IP's login-attempt entry is dropped once its window expires instead of being kept forever, the post-login redirect target rejects backslashes, and the setup page requires a login for test-connection once setup is complete.
- The legacy plan-file fallback is limited to
plan/, so an unrelated Markdown file in the workspace is no longer loaded as a plan. - Panel scripts substitute their parameters in a single pass, so a value that itself contains a placeholder token is no longer expanded again.
- Escaping: channel user fields in the pending list are escaped, the plan viewer's Markdown is sanitized, and children of a disallowed tag are sanitized before the tag is unwrapped.
- Sandbox: the image is checked before a container starts, a missing image is named instead of reported as an anonymous failure, a dead host no longer costs a 300-second wait, and
pull access deniedis treated as a missing image rather than an authentication problem. - The active fallback model is persisted for agents that use the global default, so a reset no longer leaves the badge and the actual model out of step.
- Disabled slash commands are enforced at execution time, not only when the command list is rendered.
- Malformed
*_JSONplugin configuration is rejected instead of silently ignored, and token-pricing parse errors are surfaced. set_owner_nameno longer overwrites a super agent'sSYSTEM.md.evonic doctordrops the obsolete skillsets check, and the Evomem brain-db connection is closed on error paths.- Large MCP tool listings are no longer truncated mid-JSON.
Full Changelog: https://github.com/anvie/evonic/blob/main/CHANGELOG.md
Upgrade: ./evonic update then restart the service.