Skip to content

Raven 0.2.8 (2026-11-17)

Latest

Choose a tag to compare

@github-actions github-actions released this 17 Nov 14:06
· 17 commits to main since this release
814e67f

Raven 0.2.8 is a patch release. It brings OpenShell sandboxing and network permissions, DAG replanning that carries unfinished work forward, task delegation as the host's default, a context-window breakdown with better prompt-cache reuse, and more reliable Task Board updates across multiple matters. Other additions include live switching between local EverOS and EverOS Cloud, DeepSeek Harness and MiniMax Code ACP presets, Agent Plugin settings, session search on Cmd+K, shared web/TUI prompt history, SOUL.md merges that preserve local edits, and faster loading and streaming in the web UI.

Review the upgrade notes: the boxlite sandbox backend and the raven sandbox command group are removed, and an install that ran boxlite sandboxes needs the OpenShell SDK, a gateway and Docker before a sandbox starts again; tools.sandbox.allowNet moves to permissions.network; push skill discovery and its config keys are removed; the Skill Hub source is on by default on every install; an explicit message send to another chat now needs that chat in message_targets; the MCP SDK moves from 1.x to 2.x; Agent Plugin installs recorded before this release must be re-imported before Raven will overwrite or remove them; and a DeepSeek model's reasoning-effort row offers Low, High and Max only.

Highlights

OpenShell sandboxing and network permissions

  • Sandboxed commands run in NVIDIA OpenShell on its Docker driver, and the boxlite micro-VM backend is retired. The model of one sandbox per executor is kept: the main loop owns one, and every sub-agent run and DAG node gets its own. The workspace, extra volumes, /agent-home and /plugin-data are mounted at the guest paths boxlite used. A sandbox needs the sandbox extra (openshell>=0.1.8,<0.2, not on Windows), an OpenShell gateway, and Linux 6.2 or newer with Landlock ABI 3 and Docker 28 or newer, or macOS with Docker Desktop or colima; raven onboard checks that the SDK imports and that a gateway answers, and raven upgrade keeps the extra when the SDK is installed (#1269).

  • The network is part of the permission set. permissions.network replaces tools.sandbox.allowNet: a default of open or closed, plus allow and deny rules by host, port and program. A connection the sandbox blocks while a command runs is settled the way a tool call is: a deny rule refuses it, ask asks you (a sandbox.network card on the web, the usual approval prompt in the TUI and IM, a permission request over ACP), smart lets a reviewer open what the command plainly needs, and full opens it for the session. With nobody to ask it stays blocked, the result says so, and a one-shot raven agent -m exits with status 3. IP addresses and internal names always require a person. An approval tells the agent to retry; Raven does not rerun the command automatically. Every sandbox reads the live set and applies a change before its next command, process start or MCP tool call (#1269).

  • Raven's own agents inherit the host's permission set, including ask, allow, deny and network rules, and use OpenShell when the host does. A mode pinned by the agent's hosting still applies: Raven-Code asks over ACP and runs one-shot work at full, while Raven-Oncall applies the owner's mode in its own gate. The design engine's renderer also moves to OpenShell and always runs offline (#1269).

DAG replanning that preserves ongoing work

  • A replan starts a successor run immediately and carries the unfinished nodes its new plan lists or depends on into it: a running attempt keeps running, a suspended one keeps waiting, and each carried node keeps its id instead of being dispatched twice. The exception node being answered is given up; redo its work under a new id. Finished nodes remain readable as context, while unfinished nodes the new plan leaves out stop with the old run. The sealed record marks carried nodes as carried and names their successor in carried_into; both front ends follow the successor, which has its own run record and card. A plan that adds or redefines nothing is refused (#1206).

Task delegation by default

  • When delegation is available, the host's role now sends task work that needs a tool to an agent, however small, because the host's own turn holds the conversation: a message sent meanwhile is merged into that turn and every finished agent's announcement waits behind it. A quick lookup therefore comes back as an agent's announcement on the next turn instead of inline, and the host explains a returned report in full. The host keeps orchestration and its own skill maintenance; a first task-work result from its own tools reminds it to delegate the rest. These host-role changes do not apply to launched product-agent processes. The prompt grows by about 210 tokens per call (#1259).

  • What the host carries on every call is shorter elsewhere: the spawn and run_subagent_dag tool texts lose about 175 tokens (#1235), and the DAG orchestration guide a model reads before its first run_subagent_dag call goes from 11,000 to 5,940 tokens (#1228).

Context window breakdown and prompt cache reuse

  • Clicking the context ring on the web composer opens a card that says what is filling the window: the fill as one number, a grid of cells coloured by part, and each part's own count -- system prompt, tool schemas, conversation, files, memory and skills. The parts are the loop's local estimate of the prompt it sent and ride the turn's usage as context_parts (#1205). Attachment notes and pictures are counted as Files rather than conversation (#1229, #1286), and the always-on skills digest, the per-turn skill menu and the results of find_skill, read_skill and use_skill are counted as skills (#1212).

  • The runtime context stays stable across turns. Each user message opens with a runtime-context block (time, channel, chat id, notices) that the session write used to strip, so the next turn replayed the message without it and the request stopped matching the provider's cached prefix at the previous turn. A later turn now replays each earlier user message with the block it was sent with, and an MCP plugin's state is told once in a note that stands until one replaces it, rather than on every turn. Replay also works through the curator fast path and overflow recovery. Existing saved turns without this metadata are not reconstructed, and a model, tool or system-prompt change can still invalidate the cached prefix (#1276, #1294, #1331).

Task Board tracking across multiple matters

  • A turn that moves several matters at once now writes a step on each matter's own group, instead of folding them into the group the turn opened on. Its post-turn update now reads the current turn's window rather than the entire conversation and checks evidence for each candidate matter. This reduces the input sent to that call, though a reasoning model can take longer to settle the board; the reply does not wait for it (#1329). A group that still needs you keeps its open questions when a later turn writes a step with none: one more call carries them forward through a new carry action on the task_board tool (#1178). More of the questions a reply asks land on the board, including one asked about a message sent mid-turn (#1236).

  • The task_board tool declares the type of the lists an entry carries, so a model that builds tool calls from the schema (MiniMax, for one) no longer has every post-turn update refused; the final schema also declares the per-matter evidence checked by the multi-matter update (#1320, #1329).

  • On the web, a message about a Task Group sent while a turn runs takes its group into the queue with it, so the Move on bar goes at once and later messages join the running turn (#1210); a message sent with a Move on opens with On "<title>":, so the bubble and the stored conversation say which matter it is about (#1268).

Upgrade notes

  • Breaking - sandbox backend: boxlite is removed. The first load rewrites tools.sandbox.backend "auto" or "boxlite" to "openshell" and removes debug and diskSizeGb, with a notice per change, and the sandbox refuses to start until the OpenShell SDK and a gateway are installed; there is no fallback to running on the host. Windows is not supported (a Raven inside WSL 2 is a Linux host that can use it), and on a root host the sandbox runs as the image's user (UID 1000 when the image declares none) and stops with an error when that user cannot write the workspace. Inside the sandbox only /tmp and the mounted volumes are writable, so pip install, apt-get and npm i -g do not work there, and file-heavy work runs several times slower than on the host. An install that never enabled a sandbox is unaffected. Follow the OpenShell installation guide: its gateway is separate from raven gateway and must allow Docker bind mounts. Rolling back needs the config edited by hand: an older build fails to load a config that says "openshell" and does not read permissions.network (#1269).

  • Breaking - raven sandbox removed: the command group, the sandbox debug server and scripts/boxlite_cli.py are gone; openshell sandbox list, exec and connect replace them (#1269).

  • Breaking - sandbox network: tools.sandbox.allowNet moves to permissions.network on load, with a notice: true becomes the open default, false becomes closed, and a host list becomes allow rules with the default closed. An entry without a port allows ports 80 and 443, an explicit host:port keeps its port, and "*" keeps the open default; unsupported entries are omitted and named in the migration notice. The open default allows hosts with three or more labels on any port plus common service domains on common ports, so it is narrower than boxlite's fully open network: an unlisted two-label domain such as example.com needs a rule or approval. Wildcards must occupy whole labels (api*.example.com fails validation). What an approval opens in the main sandbox reaches every conversation until Raven restarts (#1269).

  • OpenShell network limits: under default: open, a deny rule cannot carve a host with three or more labels out of OpenShell's broad allow rule. An allow rule that only partly overlaps a deny rule also remains open, with a warning; use default: closed and non-overlapping rules when a target must be unreachable. An address already approved directly in a sandbox remains open there until that sandbox stops, even after a deny rule is added. Allow once lasts for that sandbox's lifetime, and Allow for this session reaches the conversation's sandboxes until Raven restarts. External ACP agents still have their own permission boundary; the host's rules do not provide review of every operation inside a third-party agent (#1269).

  • Breaking - design renderer network: the renderer is always offline. render.allowNetwork and render.networkAllowlist are retired and ignored with a notice; render inputs that fetched remote resources must make them available locally (#1269).

  • Breaking - plugin contract: CONTRACTS_VERSION is 44. A plugin that implements ApprovalResponder must accept the needs_person keyword on await_approval, and the plugin ServiceLocator adds sandbox_gateway for render backends (#1269); the response double on the provider socket gains truncation_reason (#1265).

  • Breaking - push skill discovery removed: skills reach the model through the pull path only (a short menu plus find_skill and read_skill), and no turn spends a model call choosing skills. A config that set skillForge.discovery: "push" gets the menu; the push-only keys (discovery, rewriteMaxTokens, injectionMode, injectMax, router.topK and the gate's settings) are removed from the config on load, the skill gate role leaves the settings page, and settings.set refuses its keys. Out-of-tree code that imported SkillsSegmentBuilder, QueryRewriter, LLMGateFilter, resolve_refs or skill_gate_pin breaks (#1295); so does a caller of AgentLoop.set_skills_sink (#1292).

  • Breaking - Skill Hub on by default: the conversational Skill Hub source now defaults on for every install, upgrades included, with https://skillhub.evermind.ai as its endpoint, so conversation text is sent to the shared marketplace on a qualifying search without an onboarding write. Only an explicit off is preserved: skillForge.router.hub.enabled: false, or a legacy null or empty endpoint. Flip the switch on the Skills page or set enabled: false to stop it on the next search. RAVEN_SKILLHUB_URL now also steers conversational retrieval, and a value the client refuses turns that retrieval off for the process (#1238).

  • Breaking - explicit sends to another chat: a message(channel=..., chat_id=...) call is now delivered to the destination it names, and that destination must be listed in the channel's new message_targets setting, which is empty by default. allow_from governs only who may message the bot and no longer authorizes a send, a wildcard included, so a deployment that relied on it for cross-chat sends must add the chats to message_targets (#1195).

  • Breaking - MCP SDK 2.x: LiteLLM moves from 1.85.0 to 1.108.0 and the MCP SDK from 1.30.0 to 2.3.0 (mcp>=2.3.0,<3.0.0), which brings httpx2, boto3 and huggingface-hub into the install. Out-of-tree code that imports the 1.x MCP SDK from Raven's environment needs the 2.x names. OpenRouter discovery and missing price or window rows now depend on its live catalogue, and MiniMax-M3 prompts above 512k tokens are priced at 0.60 per million instead of 0.30 (#1174).

  • Breaking - Agent Plugin ledger: the ledger is now version 2, recorded per workspace, and Raven verifies its own markers before it overwrites or deletes an installed directory. An install recorded by an earlier build is kept as unverified: move its directories aside and re-import before import or remove will touch them. There is no reverse migration, so keep matching backups of the ledger and the installed-skill directories if you may roll back (#1264).

  • DeepSeek reasoning effort: a reasoning model served directly by DeepSeek offers Default, Low, High and Max; a gateway serving that model uses its own declared ladder. Under the final fallback rule, a stored Medium pick runs at Low, Xhigh at High, and Minimal falls through to the conversation mode or global default because DeepSeek declares no lower rung. The saved pick is retained for a later model that supports it. In general, an unsupported pick uses the highest declared rung below it, falling back to the mode or global default only when no such rung exists (#1176, #1123).

  • Session search: the rail's list is no longer filtered in place, and Cmd+F no longer expands the rail on its way to search (#1310).

  • Deck picture search: ppt_image_search follows the effective deck search vendor inherited from tools.web.search.provider, supporting Serper, SerpApi, Brave, Firecrawl and Serply. An unsupported choice or a selected vendor without a key falls back to Serper; without a usable key on either route, the picture tool and Design-to-PPT route stay unavailable. Tavily remains supported by the host's image_search, but its hits lack the dimensions and source page the deck needs. The Search service settings explain the effective vendor or fallback, and the Raven-Design card reports route readiness. A deck-specific imageSearch.apiKey overrides Serper only. Serply also joins the host's image_search vendors (#1258).

  • Generated participants: a generated function now receives each tool call as a plain id / name / arguments mapping instead of a dataclass string; one that parsed the old string must read the fields (#1321).

  • Desk and settings ids: the desk's stored size and position are reset once, so the new full-height default reaches everyone (#1243); the settings section id plugins now names the Agent Plugin section, and the MCP servers' section is mcp (#1314).

Other improvements and fixes

Live memory backend switching

  • Settings -> Data & memory switches long-term memory between the local EverOS server and EverOS Cloud without reloading the gateway. The page offers a checked API-key panel for Cloud and a Start button for a local server the plugin manages; an externally managed server gets instructions and a Check again action. Only installed backend plugins can be selected. A switch requested during a turn waits until it finishes, then updates recall, storage, skill retrieval and after-turn feedback together; pending records finish against their original backend. A settings write through raven_config uses the same live path where the settings service is available, and otherwise reports that a reload is needed. New RPC methods include memory.switchStatus, memory.applyRuntime and memory.startLocal. The section's Know-how tab is renamed Skills (#1186).

SOUL.md updates that preserve your edits

  • SOUL.md is now a maintained prompt. The workspace sync records the product text it last wrote in .soul.md.template and, when a release ships a new voice, merges the release's changes into your copy with a line-level three-way merge. A product change that conflicts with your edit leaves every section it touches as you have it and names them in a notice; changes no conflict touches follow the release. A custom copy that cannot be attributed to a shipped template is left alone with a notice. soul.local.md remains the place for local additions, and editing or deleting it updates the appended block on the next sync while the recorded block remains at the end of the file (#1227).

Agents and sub-agents

  • A reply cut by the model's context window is told apart from one cut by the output limit: Anthropic's model_context_window_exceeded is read as a truncation (#1215), and such a turn shrinks the prompt, or fails naming the context window when nothing can be shrunk, instead of doubling max_tokens (#1265).

  • A DAG node's exception report fences the judge's evidence as untrusted content and cuts it at 500 characters (#1138). A broken turn's stored failure reason keeps the exception class, as the live event already did (#1216).

  • An acp row that redeclares a built-in agent under the same name without a command now reaches the roster, running this install's own raven acp (#1257). A sub-agent's unprompted report (an armed wake, a watch round) names its agent, so its row wears that agent's mark and says "From " (#1311). A failed direct-chat turn is handed off as {prompt.md,error.md} instead of naming an out.md that was never written (#1315). A worker seat with an instruction addendum and no brief no longer receives the addendum twice.

  • The on-call agent no longer stops waking during a long run: its trunk gate runs at full and the owner's approval mode is applied in a gate of its own, so a busy reviewer's timeout cannot refuse the call that schedules the next wake (#1171). A Playbook started over RPC charges its model usage to the requested session and its owner (#1244).

  • A malformed harness-generation capability catalogue now reports one loader error naming the file and invalid module or section, instead of leaking an AttributeError or failing later as an unknown capability (#1298).

External agents

  • DeepSeek Harness (DSH) joins the ACP presets (dsh), and Raven's ACP client now continues a session through session/resume for an agent that declares it without loadSession; such an agent reads as stateful, gains instance handles, and is closed by the pool when idle (#1211). MiniMax Code joins as minimax_code, reached through a local mcode acp, with an install hint and mcode login as the sign-in remedy. If a desktop login is stored under ~/.minimax, set MINIMAX_DATA_DIR to that directory's absolute path for the ACP process; a literal ~ is not expanded (#1199).

  • A run on the openai lane (a deep-research endpoint, for one) appears in the trajectory waterfall as an external agent, with its usage and the reasoning steps the endpoint reports (#1262). The four Raven agents -- Research, Code, Design and Oncall -- wear their own marks in the roster and the waterfall (#1202).

  • An ACP agent upgraded in place is measured again: the capability snapshot's fingerprint includes the launcher's path, mtime and size (#1179, and on Windows for a relative launcher, #1285). An agent whose boot probe failed gets its mode and model menus from the first session a dispatch opens. That first dispatch still uses the agent's default; the recovered tier applies from the next dispatch, and reasoning-effort choices still require a successful probe (#1183). A pooled connection is reused only by the event loop that made it, so a CLI command run beside the host no longer waits on a connection that answers nothing (#1203).

  • On Windows, a third-party CLI agent whose row names a relative launcher and sets a cwd now starts (#1287), and an old Node.js ahead on PATH is named for a launcher installed through npm's .cmd shims (#1217).

  • The Mirasim adapter finds a host on whatever port its token file names, the mirasim serve default 4939 included (#1175). A failed npx download is named as a download failure on every path and retried once from the mirror, unless RAVEN_NO_MIRROR or an operator's own registry is set (#1184). The copy shown when an agent cannot connect is shorter, and the status reads "Sign-in needed" instead of "Unauthorized" (#1200).

Models and providers

  • LiteLLM 1.108.0 refreshes model and provider metadata, including Responses routing for current Copilot GPT models. Explicit model effort lists take precedence over inferred defaults; Kimi K3 declares Low/High/Max and Mistral Medium 3 High. The token estimator defaults to tokenizer files bundled in the LiteLLM wheel, so a cold offline start does not need a download; an explicit TIKTOKEN_CACHE_DIR is preserved. OpenRouter's curated model ids remain available offline, while discovery and missing price or window rows use its live catalogue (#1174).

  • The Add model list decides a model's kind from its declared output modalities, so an image model the bundled snapshot does not carry is no longer filed under Text (#1172). A vendor that refuses a forced tool_choice (DeepSeek in thinking mode) is retried once under auto on every wire (#1234).

  • Azure OpenAI usage keeps its cached prompt tokens (#1191). A streamed call is recorded under its model, not its tools list; a name longer than 200 characters or carrying a line break is recorded as ? (#1237). Volces Ark image generation is asked for inline bytes, so a picture already billed is not lost behind a fake-IP resolver (#1182).

  • A voice message that arrives with no Groq key, or with one Groq rejects, is answered with the fix (raven provider set groq --api-key <KEY>, then restart the gateway) instead of an empty transcript (#1300).

  • The host and research Firecrawl tools use the v2 API, ahead of v1's shutdown on 2026-12-10 (#1231). The Raven-Code and Raven-Oncall launchers inherit the host's web vendor keys and vendor choice from tools.web.providers, so they no longer start without web_search on a host set up on a current Raven (#1256).

Skills and Skill Hub

  • With pull discovery now the only path, workspace-confined Raven-loop sub-agents can read bundled skills through read_skill under their effective skill allow-list and live blocklist. Memory-backed evolved skills and cases returned by find_skill can be opened through read_skill or use_skill even when no file exists on disk; after a restart, search again to retrieve their bodies. Skill-read traces record the registry's actual source layer, and token-budget estimates no longer render a skill directory absent from the real request (#1295).

  • The Skill Hub switch applies to the next search without a reload, including on a gateway that started without an explicit Hub endpoint. A legacy null or empty endpoint preserves an opt-out, and an invalid RAVEN_SKILLHUB_URL disables conversational retrieval for that process instead of silently using the shared marketplace (#1238).

MCP and Agent Plugins

  • Agent Plugin packages get a section of their own in web settings: version, source, skills, whether a verified import stands in the workspace, each MCP server with its connection state, and a "Start its MCP servers" switch that adds the package to agentPlugins.mcpPackages or takes it out. This switch takes effect on restart, and the row says "needs restart" until the running process matches the config. Package import and removal remain CLI operations; the page supplies a copyable raven agent-plugins import <path> command. The section that listed the MCP servers in tools.mcpServers, formerly called Plugins, is renamed MCP (#1314).

  • The Agent Plugin CLI distinguishes verified, missing, replaced and legacy workspace installations. A partial removal remains retryable, damaged or unsupported ledgers are refused, and remove --json preserves its kept name list while adding kept_details with the reasons (#1264).

  • The MCP 2.x transport migration preserves the protocol version negotiated by each endpoint relay connection, including concurrent connections negotiating different versions (#1174).

  • A package's MCP args and env are refused when a value carries a .. path component after placeholder expansion (#1208), including one hidden behind an option or a list delimiter such as --config=../secret (#1253); a refused entry is skipped and reported, and the rest of the package loads. A symlink loop in a package is treated as outside the package on Python 3.13 as on 3.12 (#1251).

  • The npm registry probe for a host's own stdio server runs in that server's cwd, so a project .npmrc there is seen (#1214); a package server or a playbook definition reads no project file at all (#1266). A stdio server starts when the host's sys.stderr has been replaced by a stream with no file descriptor, and endpoint and relay diagnostics retain the upstream reason a stream failed (#1177).

Channels, CLI and TUI

  • An opt-in link ties one owner's Telegram private chat to an existing page conversation: an owner-only telegram-project-link.json in the workspace turns it on (POSIX only), /stop cancels active and queued work, and other slash commands in the linked chat are refused (#1195).

  • A cron schedule without an explicit timezone follows the local DST rules instead of a fixed UTC offset, so a daily reminder no longer shifts by an hour across a change (#1196).

  • /logs in the TUI pages the host's real log tail again, over a new system.logs RPC (20 lines by default, 80 at most) (#1198). The TUI resolves RAVEN_HOME the way the host does, heap dumps included, and reads a CRLF or CR-only history file as the host reads it (#1207).

  • raven no longer fails with "refused the recorded token" after a raven web --stop on Windows: the serve record is bound to the gateway generation that wrote it (#1187). raven doctor --help states its three exit codes, and its table names every branch that returns them (#1280).

  • Tracing dashboard startup tells you when it adopts a viewer it cannot manage through raven tracing stop. Viewer records are written atomically, malformed records are discarded, transient read or lock failures keep the record for a retry, and stop clears a record only if it still names the process it acted on. An unreachable recorded viewer blocks its own port rather than every requested port (#1280).

Web UI

  • Session search is a dialog. The rail's magnifier and Cmd+K (Ctrl+K) raise a centred card listing the sessions that match, each with its title and its last line with the match marked; arrow keys move, Enter opens, Escape closes. Cmd+F stays as an alias. The rail no longer carries an inline search field (#1310).

  • The composer recalls sent prompts, reading and writing the TUI's own history file (.raven_history under the Raven home) through two new RPC methods, input.history.list and input.history.append. Up at the first line recalls an earlier prompt; Down at the last line walks forward and restores the draft. Each front end loads its history once at startup, so entries written by another running front end appear after a reload. In the TUI, an edited recalled entry detaches from the history cursor, a double-Esc clear stays recoverable with Up, and a submitted entry stores the expanded paste body. Recalling a slash command reopens its command palette so Enter executes it (#1189).

  • Long conversations open faster. The stage mounts the last 20 turns of a stored conversation and widens by 20 as you scroll toward an edge, where it used to build every turn at once (#1293). A streamed reply is redrawn from its first changed block instead of whole, so a text selection survives the stream (#1316); a fenced code block stays plain while it streams and is coloured once its closing marker arrives, with the highlighter warmed after boot (#1313); a keystroke in the composer no longer re-measures the dock (#1318); the boot splash keeps moving while the bundle loads (#1317); and raven serve sends the page gzipped to a browser that accepts it (#1319).

  • A warmer page. Selection, focus and the cost chart take a warm yellow accent (#1226); the settings dialog becomes warm paper and the main page frosted glass, with the desk palette opening at full height (#1243); the chat header gathers into one capsule with a Conversation | Trajectory switch and a live desk chip (#1232); the composer's permission chip, model chip and context ring sit in one segment beside send, the ring printing its fill as a number (#1307); the permission picker is drawn as a sibling of the context card, each mode with its own glyph and colour (#1263, #1288); the rail slides out instead of vanishing (#1282); and the rail's import action moves into a More group while the thing the rail lists is called a session throughout (#1267).

  • Sub-agent work reads like the conversation. A sub-agent's record in the task pane is painted by the transcript's own renderer, and a delivered result that opens a new turn is drawn as a row of its own rather than as a line inside the previous reply (#1248). The floating "N running" chip is gone: once the host turn has finished and delegated work is still live, the turn's own row reads "waiting for subagents" (#1190). A direct chat with a sub-agent says that the exchange stays out of the main agent's context and how many entries will be reported with your next message (#1245).

  • Delivered files carry a download link on the tile, the desk's shelf and the viewer header (#1185). A failed tool call is drawn as ordinary work with a failed chip, and no longer opens its fold (#1188).

  • A dragged composer height stays where it was put, and the turn nav holds still during the drag (#1173). The model picker's Mode row reads the conversation being opened, not the one being left (#1122). A delegated result is drawn once when a reload replays its turn (#1312), and a delivered bubble with a wide table or code block stays inside its turn (#1254).

  • Quiet text meets 4.5:1 contrast in both themes, and the digits on the amber count badges are dark (#1322). Every module page starts with the Task Board's title row and frame, and the connector pages fill a wide window (#1323). The By model table under Settings > Usage fits its card (#1284), the desk palette sits on the Tasks board's ground (#1192), sub-agent record actions no longer introduce sideways scrolling, and settings-row controls stay aligned inside their cards (#1267).

Evolver

  • The evolver's patch guard is tightened: paths that escape the subject tree are rejected (#1222), every path in an applied diff is checked, not only the declared ones (#1273), immutable paths match case-insensitively (#1274, #1324), and a command that writes to the subject checkout through the worktree's .git is detected, undone, and its candidate discarded (#1275).

  • Every git call is bounded by a 60-second timeout (#1223), the failure map and node ledger are written atomically (#1224), and an activation spec with a threshold or max_chars below 1, or an invalid trajectory_regex, is refused at parse time (#1225).

Dependencies, documentation and regression coverage

  • Dependency updates include fast-uri 3.1.9 (#1145), source-map-js 1.2.2 in both front ends (#873, #874), sharp 0.35.5 in the WhatsApp bridge (#875), and soupsieve 2.9 (#482).

  • The changelog is reorganized by shipped release (#1194), and both READMEs add a trajectory walkthrough under What's New (#1170). DAG tool descriptions, prompts and status labels consistently distinguish a decision node from its selection.

  • Regression coverage now pins the existing tree-aware bandit scheduler (#1290) and the two history readers' agreement on CRLF files (#1247). Test-only fixes wait for complete state-file payloads (#1255) and remove a live-writer race from the cancelled-run backfill test (#1230).

Install

New install on Linux, macOS, or WSL2:

curl -fsSL https://raven.evermind.ai/install.sh | bash

New install on native Windows, in PowerShell:

irm https://raven.evermind.ai/install.ps1 | iex

Windows PowerShell 5.1 (the version built into Windows) cannot follow that URL's
redirect and stops with (308) Permanent Redirect. If you see that error, use the
direct one instead:

irm https://raw.githubusercontent.com/EverMind-AI/Raven/refs/heads/main/install.ps1 | iex

The installer puts uv, Node.js 22, Raven and its plugins in place, downloads the
browser runtime and the word dictionary and PDF models the knowledge reader uses
(about 110 MB), installs or offers LibreOffice, and then finishes by opening
Raven in your browser. First-run setup happens on that page. When a download
host cannot be reached, the installer falls back to a mirror and says so, except
for Raven's own release files, which come from GitHub Releases and have no
mirror; RAVEN_NO_MIRROR=1 keeps every download on its default host. The
installer holds the terminal while the page is up; press Ctrl-C to stop it, then
start Raven again with:

raven web

That keeps Raven running in the background and opens the page; raven web --stop
stops it. Bare raven does the same, and opens the TUI only where no browser
can be opened. Prefer the terminal? raven tui runs the same first-run setup and
opens the TUI, and raven onboard stays the explicit way to reconfigure later.
Set RAVEN_MINIMAL=1 to skip the browser, model and LibreOffice downloads, or
RAVEN_NO_LAUNCH=1 to have the installer return without opening the page. The
Windows installer does not download the dictionary and models; there, and after
RAVEN_MINIMAL=1, raven resources fetches them.

Upgrade

Already running Raven? Upgrade in place -- configuration, sessions, and memory
are preserved. On a raven web install whose page is served by 0.2.3 or later,
the update row at the bottom of the sidebar does this for you, and the page
reloads itself, still signed in, once the new version is up. A page served by
0.2.2 or earlier refuses it; from the command line, stop the page first, then
upgrade:

raven web --stop
raven upgrade

On Linux and macOS raven upgrade runs the install in the foreground and
returns when it is done. On native Windows it hands the install to a separate
helper and returns at once; wait for the helper's completion message. Then
start Raven again:

raven web

Neither raven upgrade nor the update row downloads the knowledge reader's
word dictionary and PDF models (about 110 MB). An install that has never
fetched them runs this once:

raven resources

raven upgrade installs the latest stable release together with the plugin
wheels it ships. It never picks up a pre-release unless this install has joined
the beta channel (a ~/.raven/beta.json that only the beta installer writes).
raven upgrade --check reports whether a newer release exists without
installing it. Editable source checkouts are never overwritten, and a page
served by one refuses the update row as well: raven upgrade reports the
checkout path and how far it is ahead of or behind origin/main, and the remedy
is git pull && ./install.sh in the checkout on Linux and macOS, or, on
Windows, raven web --stop, git pull, then
powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1 once the pull
succeeds. Rerunning the one-line installer
also upgrades, and ends on the running page.

Release Status

  • Version: 0.2.8 (#1336)
  • Tag: v0.2.8
  • Release type: patch
  • Assets: the raven wheel and source distribution, four plugin wheels
    (everos_memory, everos_cloud_memory, design_engine, ppt_engine), the
    locked constraints file raven-constraints.txt, and the plugin list
    raven-plugins.txt, which installs every plugin wheel the release carries

Notes

  • Full changelog: v0.2.7...v0.2.8. The notes cover the tagged release and its attached packages.

  • PyPI publishing is not enabled yet; the supported public install path uses the GitHub Release wheel asset.