Skip to content

Releases

1.3.7

Choose a tag to compare

@github-actions github-actions released this 21 Nov 18:21
59278b8

What's new

  • A cleaner message box: one line inside a chat, a few lines on a new chat, with the mic and a filled send button inside it.
  • Buttons under the box for attaching, commands (/) and mentioning (@), so you don't need to know what to type. On a phone they sit behind a + in the box, and the keyboard closes once you send.
  • Attach a folder from the same attach button, or drag a folder into the chat. It's sent as one zip with its sub-folders kept.
  • Late at night or early in the morning, the greeting changes: "Working late?", "Still up?", "Up early?".
  • A step's name that's too long for its row can now be read in full: hover over it, or open the step (tap it on a phone).
  • When a job asks for your OK on many things at once, the first card shows in full with "N more waiting after this" under it; answering it brings up the next.

For developers

  • chat-view.tsx composer: single-row input with mic and send inline, tool row under it on fine pointers, a + menu on coarse pointers; / and @ buttons put the character in and open the same menus (useMentionMenu returns moveCaret) (#344, #345, #346).
  • zip-folder.ts: zipFolder takes {file, path} pairs; readDropped walks a dropped folder entry (#344).
  • tool-calls.tsx: cut-off step labels get a title and wrap when the row is open (#334); waiting Allow cards show one at a time (#340).

1.3.6

Choose a tag to compare

@github-actions github-actions released this 21 Nov 15:00
1a7e372

What's new

  • When Emplio fills a saved login into a website, it now presses Enter afterwards, so most sign-in pages go straight through without a separate click.

For developers

  • backend/src/logins.ts fill sends Enter (Input.dispatchKeyEvent) to the box typed last, except when a single field was named. Its reply and the agent instructions in sandbox.ts say to press the button only if the page hasn't moved on (#332).

1.3.3

Choose a tag to compare

@github-actions github-actions released this 21 Nov 12:23
53c15aa

What's new

  • Long WhatsApp messages that were quoted before 1.3.2 now reach Emplio in full too, not cut after 300 characters.
  • Emplio knows it can fetch older WhatsApp messages from your phone when a chat has little saved.

For developers

  • whatsapp/account.go fullQuote: quotes saved cut ("…") are read again from the saved raw message. Takes effect once the WhatsApp worker is redeployed.
  • whatsapp_linked list_chats, get_messages and search_messages point to load_older_messages.
  • Next.js 16.4.2 → 16.4.6 (#336).

1.3.2

Choose a tag to compare

@github-actions github-actions released this 21 Nov 08:16
7033842

What's new

  • When a reply fails, the message says why when it's a common reason (the model is busy, the chat got too long, the model isn't available) instead of only "Something went wrong".
  • Replying in your main chat now closes the Allow cards an agent showed there, the same as replying in the agent's own chat.
  • Voice input types into the box your cursor is in, so it works in a reply box too.
  • Notification popups show at most 4 lines.
  • Files up to 100 MB can go to and come from your apps (Google Drive uploads, Gmail attachments), up from 25 MB.
  • Emplio's thinking shows headings, bold text and lists properly.
  • Tables in replies look cleaner, and long ones show in full instead of in a small scrolling box.
  • Attach a whole folder on a computer: it goes as one zip, with its sub-folders kept.
  • A zip too big to import now says so, instead of "Nothing in what you picked".
  • A long WhatsApp message someone replies to reaches Emplio in full, not cut after 300 characters.
  • Background rounds run every 4 hours by default instead of every 2.

For developers

  • approvals.ts moveOn also skips cards shown in the chat (showIn), not only cards asked there (#325).
  • chat.ts replyError maps common provider failures (429, 402, 401/403, 404 and "banned" models, context length, 5xx and timeouts) to plain messages; the chat's error card shows them (#321, #314).
  • voice-input.tsx writes into a focused reply note (data-quote-note) (#323); the in-app toast title is clamped to 4 lines (#320).
  • App file limit 25 → 100 MB in integrations/index.ts and composio.ts (#315).
  • Thinking text renders with MessageResponse (#312); Streamdown tables restyled in globals.css, 300px height cap removed (#313).
  • lib/zip-folder.ts: picked folders become one stored zip (#309); the import page reports ZIP64 zips as too big (#322).
  • whatsapp/text.go QuotedOf keeps up to 4000 characters (#327). Takes effect once the WhatsApp worker is redeployed.
  • Background rounds default to every 4 hours (#308).

1.3.1

Choose a tag to compare

@github-actions github-actions released this 20 Nov 10:40
ec23f26

What's new

  • Emplio can do everything WhatsApp lets a linked device do. Reply privately to someone's group message in their own chat, send locations and contact cards, vote in polls, star messages, and fetch older messages from your phone. It can mute, archive, pin, mark unread and delete chats, turn on disappearing messages, and use your WhatsApp Business labels. In groups it can change the photo, approve or turn down join requests, check an invite link before joining, and run communities. It can look up anyone's profile (with a business's address and hours), block and unblock people, change your privacy settings, about text, name and photo, share your contact link, post a status, and follow, read and post in channels.
  • WhatsApp replies go faster: Emplio can answer people who wrote in the last day up to 60 times an hour (300 a day). Messages to chats that have been quiet stay at 20 an hour, so your number stays safe.

For developers

  • whatsapp/more.go: the rest of whatsmeow's client, as HTTP routes (reply-private, location, contact-card, vote, star, history, chat/{mute,archive,pin,unread,delete,disappearing,label}, labels, label, message/label, group/{photo,requests,preview}, communities, community/{groups,link}, profile, blocked, block, privacy, me, contact-link, status, channels, channel, channel/{messages,follow,mute,react,post}). Chat-list changes go through app-state patches; labels are kept in the account's kv from LabelEdit/LabelAssociationChat events (one full regular-patch fetch for older accounts); on-demand history leaves unread counts alone; Store.Oldest and Store.DeleteChat; group and profile photos become square JPEGs (stdlib). create_group takes a community, update_group_settings adds only_admins_add and join_approval, send_file sends .webp as a sticker.
  • Send limits: replies (the other side wrote in the chat in the last 24 h) get 60/hour and 300/day counting every send; other sends stay at 20/hour and 150/day counting only non-replies (sends.reply, Store.HeardFrom).
  • whatsapp_linked: 44 new actions (74 in all); COVERAGE.md lists them and what's still left out (calls, presence, others' statuses).

1.3.0

Choose a tag to compare

@github-actions github-actions released this 20 Nov 06:40
e238f83

What's new

  • WhatsApp is here. Link your own WhatsApp or WhatsApp Business: connect WhatsApp in Integrations and scan the code from your phone (WhatsApp → Linked devices), like WhatsApp Web. Emplio can then read and search your chats and groups, reply to messages, send files and polls, forward, edit, delete and pin messages, run your groups (make them, add and remove people, admins, settings, invite links), check which numbers have WhatsApp, and run automations on new messages, edits, deletes, reactions, poll votes, people joining or leaving groups and new chats. WhatsApp automations come with paid plans: checked every hour on Starter and Growth, and the moment it happens on Pro. It only writes to people and groups you already chat with, and keeps it slow so your number stays safe.
  • The official WhatsApp Business API connection is now called "WhatsApp Business API".

For developers

  • New whatsapp/ service in Go with whatsmeow (pure-Go SQLite: the login in whatsmeow.db, chats and messages with FTS5 in data.db, one folder per account) behind a token-protected HTTP API. Nobody stays connected: accounts connect when used, drop after 10 idle minutes, and catch up every 4 hours. Sends: known chats only, 20/hour and 150/day. About 29 MB with an account linking (Baileys was about 126 MB idle).
  • whatsapp_linked own app: 30 actions (Periskope's WhatsApp-level API as the checklist; a separate reply_to_message because small models miss send_message's reply_to), linking through the OAuth plumbing with our own QR page (/whatsapp/link), and a polled new_message trigger (polls/whatsapp_linked.ts, every 60 minutes, at least 60) over the new list_recent_messages action, plus 7 event triggers (edits, deletes, reactions, poll votes, group members, new chats, added to a group) over list_events. The server records those events, delivery ticks, reactions and poll votes (events, reactions, votes tables). Plans: new OwnApp.plans ({ triggers, instant }) checked by triggerPlan in startTrigger (refused with an upgrade message, polled, or pushed), cloud.plan(workspaceId) (null where nothing is sold: everything allowed), and an hourly recheckPlans that sets automations up again when a plan changes. Only whatsapp_linked sets it: Pro gets push through a watch on the WhatsApp server (PUT /accounts/{id}/watch: the account stays connected and posts each message and event, HMAC-signed, to /webhooks/app/whatsapp_linked/<connection>). Needs WHATSAPP_URL and WHATSAPP_TOKEN; Docker installs get a whatsapp service and the token from setup.sh.

1.2.102

Choose a tag to compare

@github-actions github-actions released this 20 Nov 06:25
c997618

What's new

  • Automations can use the steps Emplio saved for reuse. The ones that already exist are allowed once, on the same card as the automation's other actions; ones made later just run.
  • An automation run that hit a problem no longer passes as "nothing to do": it shows as failed and you're told, so Emplio's check-in can fix it.
  • Big Gmail searches come back in full instead of failing as too large.
  • After a job's report, its chat shows a "Saving steps to reuse next time" line instead of "Nothing to keep."
  • Agents keep seeing every step they can reuse after Emplio's daily tidy-up moves some to another file.
  • Listing your GitHub repositories works again, and a few confusing error messages are clearer.
  • On a Mac running several chats at once, Emplio's helper files can no longer be read half-written.

For developers

  • sandbox.ts: kernel.py, the emplio library and the emplio binary are uploaded under a per-reply suffix and moved into place in the install command (#295).
  • GitHub list_repositories_for_the_authenticated_user never sends type with visibility/affiliation (422); emplio threads show with no ref shows the latest job; store-skill errors with no message say what failed (#296).
  • composio.ts fetchInPages: gmail.fetch_emails past Composio's payload limit is re-asked in halving pages by page_token and joined; an email too big alone comes without its payload, with a note (#298).
  • Automations list saved tools (tools.<area>.<name>, risk changes/spends) named in the prompt or on the agent's list as actions; approve lets an automation's run use a saved tool it doesn't list (listedActions) (#299).
  • toolDefined (agents.ts) follows from tools.<area> import re-exports (3 hops, aliases) for agents' tool lists and addTools (#300).
  • outcome: a run with a problem reported in its chat or its jobs' chats that ends [SILENT] fails (#301).
  • KEEP_TOOLS_PREFIX / keptNothing (shared): the tool-saving nudge renders as a divider and its "Nothing to keep." reply is hidden (#302).

1.2.100

Choose a tag to compare

@github-actions github-actions released this 19 Nov 16:04
927bf2a

What's new

  • Emplio gets website work done in fewer steps: it does what it already knows in one go (signing in and moving on, filling and saving a whole form), goes straight to a record's own page, and makes the same change to many records in one pass. In our tests a website task cost about 17% less, and the same change on 8 orders about 27% less.

For developers

  • browserGuide (sandbox.ts): rules 2 to 5 on steps: everything already known in one run or command, a loop by find for the same change on several records, a record's own address over searching or paging (snapshot -i -u), and looking for what's needed (snapshot -i | grep) after a first look. Main's 1.2.99 guide changes kept as they are.
  • Browser benchmark: new bulk task (8 orders to Delivered); /work in commands maps to the run's folder; each run gets its own agent-browser socket folder (runs side by side once shared a browser); the first browser result gets the command list, as firstInChat gives it. Main vs this, 6 tasks × 4 runs on Haiku: 23/24 → 24/24, 11.8 → 9.5 steps, $0.0041 → $0.0034 a task.

1.2.99

Choose a tag to compare

@github-actions github-actions released this 19 Nov 14:26
be0f54f

What's new

  • Emplio reads much shorter instructions before each reply, so replies cost less and it gets commands right the first time more often: it rarely has to look up how a command works.
  • It does independent things at the same time (several look-ups in one go) instead of one after another.

For developers

  • Code mode: the system prompt carries a Python signature list of the emplio library (commandList in sandbox.ts, one for Emplio and one for agents), replacing per-command sentences; a test checks every listed command and flag exists. agentRules, MANAGER_PROMPT and the sandbox instructions compressed: the manager prompt is about 8,900 → 7,200 tokens. In a 36-reply probe, help() calls went 16 → 6 and wrong flags 1 → 0; hand-off, manager, skills, background-round and workday evals hold (workday input per chat 136k → 115k).
  • New rule in agentRules: commands that don't need each other's results go in the same step (they already run side by side, up to 4).
  • Browser: every command as a Python list (BROWSER_COMMANDS), sent with the first browser result once per chat (again after the chat is folded into notes, Session.notesUpto); emplio browser help prints it without starting a browser. Without code mode, the first result and help carry BROWSER_SKILL instead.

1.2.98

Choose a tag to compare

@github-actions github-actions released this 19 Nov 12:50
b0595cf

What's new

  • Agents learn from every piece of work you give them, not only the first. When you send more work to an agent that's already done a job, it now also keeps what it worked out as tools, so the next time is quicker.

For developers

  • The tool-saving nudge (SAVE_TOOL) now goes after each finished run of a job, not once per job. nudged() skips the nudge's own run (also when a restart or the step limit splits it) and runs with 2 commands or fewer (SMALL_WORK). The other conditions stay: completed, not interrupted, not asking, done, and no tool saved in that run.