A local-first multi-agent app development platform, built on PenguinHarness
Changelog
One brief line per release. Per-release detail lives in changelog/<version>/.
- 0.2.13 — 2026-09-15. Company mode turns a Project's Agents into an organization — a CEO, a reporting tree, a calendar, a five-column ticket board and channels, each a file on disk, off until an admin enables it. A Benchmark became a Project-level peer of an agent, the Workspace browser became a working file manager with a context menu and no size ceilings, the desktop app took a place in the system tray, shell commands can be confined through plugin backends, a running tool call can be sent to the background, and three AgentHub fixes end gateway 4xx that used to stop a Task mid-run. 0.2.12 was tagged for this same set and failed in the first job of its release run, before anything reached npm, Docker Hub or a Release page — so 0.2.12 exists only as a tag, and has no folder here. (details)
- 0.2.11 — 2026-09-10. Rebuilt 0.2.10 with electron-builder 26.16.1 so signed macOS installers build again on the current GitHub runner image; 0.2.10 was tagged and reached npm and Docker Hub but never got installers or a Release page, and its per-change entries are in
changelog/0.2.10/. - 0.2.10 — 2026-09-10. The Files panel becomes a two-pane browser with a lazy directory tree, in-place editing with version-checked saves, and drag-and-drop file drops; the context ring now fills against the actual compaction point, adds a draggable threshold cutter that updates the Agent’s compaction threshold immediately, and ranks the five most context-heavy files. Contexts are now assembled whole from Agent State under three explicit tiers, a floating shortcuts ball fans the workbench into a ring, and scheduled tasks move into the dock via the shared Create with AI kit with sidebar alarms. Hooks (
stop,pre_tool_use,user_prompt) are now loop-level with installable packages, goal mode/continual learning/skill library move to plugins, and Agent Settings gains a Hooks tab with enable switch, zip export/import, and chat import. Added official Docker imagehiyouga/penguinharness(amd64/arm64, one container/volume) and reworked Machines intoweb.dbunder Projects with per-machine IDs, one held SSH session, and same-origin API proxy. New models include DeepSeek V4.1 Flash (deepseek-flash) as default for new Projects, Gemini 3.8 Flash, GPT-6 Astra, three Doubao Seed rows, and a protocol-pinned vLLM group. Also improvesexec_commandto acceptcommand, returns correction guidance for rejected tool calls, folds image reading intoread_file, separates API vs tool time, marks background tasks in the sidebar, streams compaction thinking, applies a silence-based LLM timeout, and uses one update dialog from check to restart. (details) - 0.2.9 — 2026-08-28. The release where the model library starts telling you what it costs. The catalog stores list prices and a promoted row declares a
discountbeside them, so the billed rate is derived rather than pasted in and a lapsed promotion is one field to delete; six TokenDance rows carry one, the card shows the price being billed with the rate as a badge, and TokenDance leads the page as the recommended group. DeepSeek's peak and off-peak tiers are billed per request: a scheduled row keeps its peak price on disk and the cost center decides the tier from each usage record's own timestamp, so a range that straddles a boundary is billed at the rate each request ran at and a finished week reports the same number whenever it is read. WeChat joins as a fourth chat channel, bound only by scanning — there is no console to copy a credential out of, the platform's poll handle never leaves the server, and it is the only channel here carrying text, images and files both ways, with voice arriving as WeChat's own transcription. Alongside that: a Machines page that installs this build onto another host from the server's own~/.ssh/config, one update-all button on every page with updates waiting and a kernel trail that can now be put down, a new Project whose default model reads a pasted screenshot without a vision proxy, a QQ gateway drop the connector recovers from that stops counting as a defect, an owner-only clear on the error table scoped to the filter on screen, a retry ladder that occupies one line instead of one per attempt, a key authorization that reports its outcome where you are looking rather than in a toast fired at another tab, the messaging HTTP surface written once instead of once per channel, and the three price buckets renamed in the Chinese UI. (details) - 0.2.8 — 2026-08-27. The release where a bound chat channel stops being a text pipe. Pictures and documents travel both ways on Feishu and Telegram — inbound as the input the web composer submits, outbound as the files a finished run both mentioned and produced, which is the pair of filters that picks the one output that was the point while keeping a steerable reply from becoming a read primitive. Replies arrive as formatting rather than literal
**bold**: a per-bindingrenderMarkdown, on by default, renders each reply in the channel's own markup — Telegram HTML, a Feishu card's JSON 2.0 rich-text component, QQmsg_type: 2— transforming what a channel cannot draw instead of leaking it, chunking without splitting an entity or a code fence, and retrying as plain text when a platform refuses the formatting, so a reply is never lost to it. Two more per-binding switches decide how a run reaches the chat: one message per non-blank line, and hold the working notes to deliver only the run's last reply. QQ joins as a third channel, bound by scanning a QR code, over a passive-reply budget the connector rations because the platform will not let a bot speak first. Alongside the messaging work: an extension seam (activate(ctx)plus two typed events, re-delivered on every App creation),penguin authfor terminal sign-in with the fixed seed password retired in favour of a printed first-login link, red-dot to-dos for Skill and preset-model updates and unexpected errors that can each be put down, LaTeX on every Markdown surface, a bot account that is no longer owned by one Session forever, a binding panel that separates the three different causes of "nothing happened", Qwen 3.8 Flash in the TokenDance group, and CI fanned out into parallel shards. (details) - 0.2.7 — 2026-08-27. A fix release for what 0.2.6 shipped. The messaging bindings it introduced now work from the desktop app — whose bundles declared
requirebut not__filenameor__dirname, so the Feishu SDK's read of its ownpackage.jsonfailed every connect with__dirname is not defined— survive a Telegram poller conflict, whose recovery was gated on agetMethat never conflicts and so re-polled once a second forever instead of backing off, and run a redelivered message once, keyed on the channel's own message id rather than on the text. The desktop app also puts its bundledpenguincommand on PATH by itself and repairs it at every launch, never touching apenguinit did not write and never installing from a location that will not persist. A command whose Workspace has been deleted now names the directory instead of reportingspawn bash ENOENT, and the POSIX shell resolver gained the fallback chain the Windows branch already had, so a machine without bash no longer fails every command before it runs. Plus a queued follow-up that is recallable however it was queued, landing and README copy corrected where it had stopped being true with the Skill lists now pinned by tests, CI verification of the packed CLI, and tests pinning bothweb.dbupgrade directions. (details) - 0.2.6 — 2026-08-27. Two Flash models (
glm-5.3-flash, listed once per route that sells it, andqwen3.8-flash) beside a new TokenDance gateway group whose Authorize key action mints a key onto your account without a trip to a console, plus app attribution for OpenRouter and TokenDance; subagents you talk to instead of watch — mid-run steering from the model and from the panel's own composer, a per-runabortreplacingkill_subagent, and child approvals that escalate to you; thepenguinCLI rebuilt as an HTTP/SSE client of the running server with a read-and-move command family, control-env injection into every Agent command and thepenguin-orchestrationSkill over it; a Session bindable to a self-built Feishu app or a Telegram bot with no public URL on either side; a clickable context ring that breaks the context into six parts; update badges that lead from the outermost menu to the control performing the update; draggable sidebar and provider groups with per-group sidebar paging; the standalone Trace page removed in favour of the conversation's Trace panel, with import moved into System settings; Agents created from library or project-directory Skills or straight from a snapshot package; a kernel update whose unit is a settings tab; and atomic replacement for every Harness state file; plus the Windows signing identity corrected toNaisNet Technology Co., Ltd.and recorded as a list that keeps the previous identity beside it, which an installed 0.2.4 Windows desktop client still has to be reinstalled to follow. 0.2.5 was tagged for this same set: it published the npm packages and then failed on the Windows desktop job, before any GitHub Release or installer existed — so 0.2.5 is on npm only, and has no folder here. (details) - 0.2.4 — 2026-08-21. A DeepSeek vision model (
deepseek-v4-flash-vision-exp, and DeepSeek moved onto the Responses protocol) with a directclaude-opus-5preset and an add-group dialog that imports an endpoint's whole model listing, the chat page rebuilt around two tabbed docks holding every side panel and any number of server-side terminals whose screens survive a reload, background execution with harness-injected completion reports, kill tools and service links on the process list, a Project-level sandbox command policy that denies destructive shell commands under every approval mode, usage charts rebuilt as time series with range and precision controls, Memory groups that export and import whole plus in-chat memory-change lists and a prompt that names when to save, one System settings dialog with explanations behind a circled "?", draft-screen examples that fill the composer beside a folder of the user's own saved prompts, per-server MCP permission, and desktop installers built from bundles that import the login shell's environment on GUI launches. (details) - 0.2.3 — 2026-08-19. Conversations you can branch, take back and watch: forking a Session from a completed reply, recalling queued steering and follow-up messages into the composer, live per-row Session status in the sidebar, a compaction rework (tool results kept, interrupted compactions discarded, a summary that streams into a collapsed row, and compaction working on a Session resumed after a restart), thinking levels selectable wherever the work happens with a mid-chat switch that states its cost and a new
maxtier from AgentHub 0.4.4, custom-model protocol and vision detection alongside a per-model fast mode, the session row's hover pair back beside a right-click menu, drag-and-drop composer attachments, and in-place hot updates of a running harness. (details) - 0.2.2 — 2026-08-11. Long-term Memory across sessions (per-agent user and Workspace scopes, index-only injection, a Memory tab), MCP Server support over stdio / Streamable HTTP / SSE with in-app management, Skills/Vault/Schedules behind Memory-style editable prompts and enable switches, dated config kernel versions with a smart no-loss Update beside restore, a hardened core runtime (unlimited default turns, context-window-derived caps, atomic Trace appends), desktop and proxy polish, a web UX batch, and a clean dependency audit. (details)
- 0.2.1 — 2026-08-04. The desktop app: an Electron shell over the unchanged server and Web App (signed-in on open, shared data root, per-root single-instance lock, three-platform installers mirrored to OSS), a landing download page with mirror-aware static links, download source selection for standalone installers and
penguin update, compaction that survives malformed summaries with retries and costs finally visible, a random seeded admin password with login throttling, skills manageable from agent settings, per-Project new-chat defaults, a Traces page scaled onto a SQLite trace index, and the server's last IO hotspots closed. (details) - 0.2.0 — 2026-08-03. One sealed installer bundle per target serving online and offline installs alike — now mirrored to Alibaba Cloud OSS with automatic fallback — truncated tool output recovered from the Session scratchpad, chat navigation for long conversations (shell-style ↑ input recall, stacked sticky run headers, a minimap tick rail with hover previews), steering that survives reloads and carries file attachments, a DB-served session list with paged sidebar groups, and the refreshed qianwenai / OpenRouter model catalog defaulting to deepseek-v4-flash. (details)
- 0.1.5 — 2026-07-30. Self-contained offline installer bundles for all five platforms (with MinGit bundled on Windows), composer attachments of any file type with images reaching steering and goal objectives, in-run recovery for every LLM failure short of a rejected credential, a tenth-shorter default system prompt, the paper-editorial web-design theme with thinking/image guidance in the SDK skill, and a refreshed README/landing install story. (details)
- 0.1.4 — 2026-07-27. Carries the whole 0.1.3 feature set to npm, which 0.1.3 itself never reached: its publish job failed on a version-endpoint test asserting a hard-coded build date, and only the release jobs stamp one. Plus blog images hosted in the sibling community repo rather than in this clone. (details)
- 0.1.3 — 2026-07-27. Windows support (shell selection,
install.ps1, a win-x64 package), goal mode looping Tasks until an objective completes, a docked subagents panel with a live call graph, LLM-error reconnects with a visible countdown, and in-app version display with update checks and admin self-update. (details) - 0.1.2 — 2026-07-26. Builtin file tools with git-style edit diffs, mid-run steering and queued follow-ups, handoff-style model switching with a per-request thinking level, free OpenRouter models, and isolated-origin Workspace HTML previews. (details)
- 0.1.1 — 2026-07-22. Gemini 3.6 Flash and 3.5 Flash-Lite in the model catalog, a Workspace-grouped chat sidebar, in-place upgrades via
penguin update, and the vLLM / Ollama / LlamaFactory and bento-slides skills. (details) - 0.1.0 — 2026-07-21. First feature release: the Web App, landing site and docs, the model catalog, and the self-improvement skills. (details)
- 0.0.1 — 2026-07-19. First tagged release; changelog history starts after this tag.