FrontonGPT

• Why skills matter — they're how org-specific know-how stops living in one person's head (or one buried Confluence page) and starts running consistently in every chat. Once a workflow is captured as a skill, the next person who asks gets the same in-house answer the expert would give — same conventions, same caveats, same references — without having to know which expert to ping. They turn one-off prompts into reusable, versioned playbooks, and they raise the floor of the assistant's output on the work Frontiers actually does day-to-day.
• Browse the catalog at Settings > Skills — every skill currently available to the assistant is listed there, with descriptions and the commit it's synced from, so you can see what knowledge is loaded and propose edits or new ones.
• Source of truth — GitHub — each skill lives as a folder in your organisation's GitHub repo under skills/. Everything is in version control, peer-reviewed via PRs, and auditable.
• Automatic sync — changes merged to main flow through automatically. The commit shown on the Skills page reflects what's live in the assistant right now — no manual deploy, no waiting for the next release.
• Shared with everyone — all members of the organisation use the same skills. There's no per-user override: if it's in the repo, the assistant has it; if it isn't, nobody does. One canonical source, the whole team aligned by default.
• Default-on, server-side — when the chat panel asks the server for the model list, the server injects skills into each model's effective default tools only for models whose own allowlist already permits it — i.e. Auto, Sonnet, Opus, and the other tool-calling families. GPT, Gemini, and standalone Haiku stay untouched, so no model gets handed a tool it can't reliably use.
• Opt-out switch in Settings > Extension — a new Skills tool toggle persists on your member profile (options_skills_tool_enabled). Flip it off and Skills stops getting auto-enabled on every new thread; flip it back on and the next picker refresh restores the behaviour. Per-thread overrides from the composer's tools menu still win — the toggle only governs the default.
• "Skills enabled" pill on the welcome screen — when you open a fresh thread, a small blue pill sits in the top-right of the welcome screen, on the same baseline as the date, the moment Skills is active for the next message. Disappears the second the first user message lands (the welcome view does), or if you turn Skills off in the composer or Settings.

• Durable, not just resumable — durable workspaces (v2.7.4) protect the VM; volumes protect the data. Mount the same volume in a fresh sandbox tomorrow on a different node and your files are right where you left them.
• Shared across sandboxes — mount the same volume in several sandboxes at once for one shared dataset with many readers and writers: parallel jobs, shared working corpora, cross-thread caches.
• Self-hosted, private — the file system, the orchestrator, the sandboxes, the encryption keys, and the network all live on our own infrastructure. Files written to a volume never leave Frontiers: not E2B's hosted service, not OpenAI, not any third party. The privacy guarantee FrontonGPT already gives chat (your prompts and tool outputs stay in our cluster) now extends to anything you persist across sessions — important for sensitive manuscripts, peer-review correspondence, embargoed data, and anything else that simply must not leave our perimeter.
• Cross-product foundation — the same plumbing isn't just for the assistant; it's the durable file layer other Frontiers products can build on top of. Notably AIRA: anywhere it needs to keep a working corpus, a vector index, peer-review artefacts, or per-user scratch space across sessions and machines, it can mount the same kind of volume and stop reinventing storage. One shared substrate, multiple products.

• What's indexed — ~1.2M Frontiers articles, ~126K researcher profiles, ~58K Research Topics, ~2.1K taxonomy labels, and a dedicated ~2.7M researcher insights index for high-recall reviewer / topic-editor shortlisting.
• What it unlocks — reviewer & topic-editor shortlisting (with COI rules), near-duplicate detection on incoming manuscripts, scope-fit checks against a journal section, Research Topic landscaping and adjacency queries when commissioning new topics or journals.
• How it works — PINECONE_API_KEY (read-only) is in every sandbox's env, the pinecone Python package is pre-installed, and the assistant's system prompt knows the catalog so it picks the right index, model, and namespace without you spelling it out.
• New Settings > Embeddings page — browse the indexes, namespaces, models, hosts, and record counts in one screen. Useful when scoping a prompt or sharing what's available.
• New shared integration in Settings > Credentials — Pinecone shows up as an org-managed integration so it's clear the access is shared and read-only.

• Same assistant as the extension — Teams turns now run through the same pipeline as the chat panel, so you get the full toolset (web, profile, workspace, automations, image, frontiers/publishing connectors, …), all your agents, projects, and the same model picker. Whatever you can ask in chat, you can now ask in Teams.
• Real memory across turns — every Teams conversation is pinned to its own FrontonGPT thread, so the model remembers what you said three messages ago — it doesn't every time. A reply-in-channel gets its own thread; a fresh DM continues the same one until you reset it.
• Type /new, /reset, or /clear to start over — the live thread is archived and the next message starts fresh. The bot confirms with "Started a new conversation."
• Streaming in 1:1 DMs — replies stream token-by-token into a single message that updates in place, the same UX as the chat panel. Group chats and channels fall back to a "…is typing" indicator and a single final message — Teams' streaming protocol doesn't behave well in group scopes.
• Identity is your real Microsoft account — every Teams message is verified against the bot's Azure AD app and matched to your FrontonGPT user by email, so the agent answers as you: your personalization, your audit trail, your usage counts. Unknown senders get a polite refusal instead of being auto-provisioned silently.
• Visible in your thread list — Teams conversations show up in the FrontonGPT sidebar like any other thread, so you can reopen a Teams chat in the extension to keep working in a richer surface (file uploads, the workspace sandbox, …) and the next Teams message picks up where you left off.
• Generated images, exports & QR codes render inline — when the assistant produces an image (image.generate), exports a workspace file (export_file), or generates a QR code, Teams shows the actual image or download link directly in the message — no Auth0 login round-trip, no broken-image placeholder. Behind the scenes the Auth0-protected file URL is swapped for a 7-day pre-signed S3 URL just before the reply lands in Teams, so Microsoft's image proxy can fetch it anonymously. Live artifact / app wrappers (serve_artifact / serve_app) still prompt a one-time Microsoft sign-in per browser — they iframe a live sandbox tied to your member identity, so we keep the auth gate there.

• You vs the org — your share of total messages, your multiplier vs the typical active user, and your rank (e.g. "#42, top 12% of 248").
• Where you fall in the org — a density curve of every active user's message count, log-scaled so the long tail reads as a bell. Median dashed, you marked.
• When the org prompts — a 7×24 day-of-week × hour heatmap (cool blue → hot orange) in your local timezone. Shows the rhythm of the working day at a glance — when the morning ramp kicks in, when lunch hits, when the afternoon push lands, when work tapers off.
• Active users (DAU / WAU / MAU) — distinct users over the last 1 / 5 / 20 completed business days. Today is excluded so the numbers don't drift up over the day; weekends are excluded too. Stickiness shown as both DAU/WAU (the honest B2B anchor) and DAU/MAU (the consumer convention). Public holidays aren't excluded yet — a bank holiday inside the window will dilute slightly.
• Usage concentration (Lorenz curve) — top 1% / top 10% message share, "top X% of users send 50%", and the Gini coefficient. Tells you whether usage is broad or concentrated in a handful of power users.
• Department heatmaps + Org-level trend — the Adoption tab pairs nicely with the new bits: adoption heatmap (% engaged users per department × month), intensity heatmap (messages per active user per department × month), and the Org-level trend chart now plots a third teal line for Threads alongside Users and Messages — with a "vs prev month" delta on hover for all three.

• No more lost workspaces during scale events — retires the red apology from v2.8.4.
• Fails safely: if the drain path is ever unavailable, AWS falls back to a normal shutdown after a 45-minute grace window instead of wedging the cluster.
• With the guarantee in place, the cluster now ranges from 3 to 9 nodes — up to 144 vCPUs and 576 GB of RAM at peak. Quiet weekends are cheaper; busy afternoons still scale up within seconds. More than enough to run whatever unhinged Python you can dream up.

• Attached atomically on the first message — the client stamps the project_uid on the local thread and forwards it with the first completion request, so the thread row is born already grouped under the project rather than briefly appearing outside it and then hopping in.
• Project auto-expands — the target project stays expanded so your new thread slots into place the moment the first assistant reply streams the thread:created broadcast.
• Owned projects only — the + affordance is hidden on shared/read-only projects, matching the existing rename/delete/share permissions.

• Live rename sync — when a fresh thread gets auto-named after your first message (or you rename it manually), every tab pointed at that thread re-titles itself immediately over the same websocket channel the sidebar uses.
• ⏳ while streaming — the title is prefixed with an hourglass for the duration of an assistant reply, so a glance at the tab bar tells you which conversation is still thinking.
• ✅ when the reply lands in a backgrounded tab — if you switched away mid-stream and the answer finished without you, the prefix flips to a green check and stays there until you focus the tab again. Built-in "this one's ready" indicator, no favicon tricks needed.
• Standalone only — the sidepanel shell has no real browser tab to re-title, so the behaviour is scoped to the full-tab surface.

• Off by default — long-time users keep today's reset-to-default behaviour until they explicitly turn this on, so nothing changes under anyone's feet.
• Per-member — the toggle lives on your member profile (not on the organisation), so your teammates can keep the old behaviour while you keep your agent sticky.
• Existing threads unaffected — reopening an old thread still restores whatever agent/model that thread was last using. This only changes what happens when you click New thread from inside an agent or a pinned-model thread.

• Opt-in — off by default. Your request stays anonymous unless you explicitly tick the box.
• Retroactive credits — want credit for a feature that shipped before the checkbox existed? Ping us directly and I'll add you to the release note manually. The pill renders from a static credits field, so adding a missing requester is a one-line change.

• Scaled compute cluster — increased the base node pool and enabled auto-scaling so the cluster adds new nodes automatically when demand grows.
• Auto-scaling compute — the cluster now automatically adds new nodes when CPU or memory utilisation crosses the threshold (scale-out only — nodes are never removed automatically, so running workspaces are never interrupted). No more manual capacity planning during usage spikes. Automatic scale-down with graceful draining is coming in a future release.
• Smarter crash recovery — if a brand-new workspace crashes on startup (e.g. from temporary resource pressure), the system now automatically destroys it and creates a fresh one instead of showing a confusing "paused but did not respond" error. The old error only appears for genuinely long-lived workspaces where your files need protecting.
• Clearer error messages — when a workspace is temporarily unreachable, the message now says exactly that instead of exposing internal sandbox IDs and suggesting you destroy your files.
• Upstream merge (e2b-dev/infra 2026.15) — pulled in 206 upstream commits including a thundering-herd fix for sandbox scheduling, race condition fixes, kernel I/O timeout improvements, and performance optimisations across the board.

Note: During the scaling work, some running workspaces were lost and cannot be recovered. If your workspace is missing files from a previous session, please start a fresh thread — we apologise for the disruption.

• Shared connection (default, no setup) — backed by the organisation's service-account Connected App and granted read-only access on every Salesforce object. Great for "look up this account", "show me my pipeline this quarter", "fetch the case history for opp X" — anything that's a SOQL SELECT. Cannot create, update, or delete records, by design. The credentials row shows "Using shared" so the state is never ambiguous.
• Sign in with Salesforce (your own user) — click Connect with Salesforce, approve the Connected App, and the assistant queries Salesforce as you: exactly the read and write permissions your Salesforce user has, scoped by your role, profile, and sharing rules. Required for any update / insert / delete (create accounts, update opportunities, log tasks, close cases…) and for accessing records the shared service account can't see. Every action shows up in the standard Salesforce audit trail under your name, not a shared bot.
• Per-user, encrypted, auto-refreshed — your access + refresh tokens are stored encrypted alongside your other credentials and refreshed automatically 5 minutes before expiry. The credentials drawer shows the connected account (display name + email) and the Salesforce instance URL.
• Workspace sandbox auto-wired — whichever auth is active (shared or OAuth), every chat exposes SALESFORCE_ACCESS_TOKEN and SALESFORCE_INSTANCE_URL with a fresh token, so simple-salesforce and plain requests calls just work — no token juggling. The tool path (salesforce_soql_query) and the sandbox env vars always agree on the active identity.
• Disconnect at any time — one click on Disconnect falls back to the shared read-only connection; revoke the upstream grant from Salesforce Setup > Connected Apps OAuth Usage if you want to fully cut access.

• Per-member, isolated — one database + one role per (user, organisation), provisioned on demand. Your role can only touch your database; the admin role used to provision it is not a superuser, so the blast radius of anything going wrong is bounded to your own DB.
• Shared across all your workspaces — every chat and sandbox you open connects to the same database, so data you write in one conversation is available in the next. No per-thread forks, no manual wiring.
• Wired into the sandbox automatically — once provisioned, the standard PGHOST / PGUSER / PGPASSWORD env vars plus a full DATABASE_URL are exposed in every workspace. psycopg, sqlalchemy, and pandas.read_sql all pick them up with zero setup. Tell the assistant to "save these rows to my Postgres" and it just works.
• Batteries-included extensions — every provisioned (and every reset) database comes with five extensions pre-installed: vector for pgvector embeddings (cosine / L2 / inner-product search over OpenAI/Anthropic vectors), pgcrypto for gen_random_uuid() + bytea hashing, citext for case-insensitive text columns, pg_trgm for fuzzy text search (% / similarity() + GIN indexes), and unaccent for accent-insensitive matching on non-ASCII text. Core Postgres already provides tsvector/tsquery full-text search and jsonb key/value storage, so no extension is needed for those. Heavy or sensitive extensions (PostGIS, FDWs, untrusted PLs) are deliberately left out.
• Reset in one click — bad schema experiment? Click Reset database and we drop and recreate the DB with the same credentials, empty. Your sandbox's DATABASE_URL keeps pointing at it — no reconfig.
• Email me the password — the password is never shown in the UI after provisioning, but a single click on the DB card emails your current credentials (host, port, user, password, and a ready-to-paste connection string) to your registered address, so you can hook up psql, DBeaver, TablePlus, Metabase, or your favourite BI tool. Rate-limited to once a minute per member.
• Rotate password — if you suspect a credential has leaked, click Rotate password to generate a fresh random password without touching your data. The new password is written into the encrypted blob immediately and picked up by every new sandbox you open; currently-running sandboxes keep the old password in memory until you start a new chat. Follow it with Email me to retrieve the new credential.
• Durable and backed up — data persists across sessions and automated snapshots of the shared instance cover every member's database for free.
• Encrypted at rest — your connection password is stored encrypted alongside your other credentials and never shown to the UI after provisioning.

• Permanent URLs — the link you share (or bookmark) is a branded assistant-ui URL, not an E2B host. No more 24-hour expiry sneaking up on Monday-morning reopens of a Friday-evening dashboard.
• Auto-resume on click — if the workspace is paused, opening the link transparently wakes it and the dashboard loads 1–3 seconds later. No manual resume, no "Workspace paused" error page.
• Org-scoped sharing — anyone in the same organisation as the thread owner can open the link. Non-members still get a clean "Access denied" message instead of a broken page, and thread-owner identity is shown up-top with name + avatar so viewers know whose workspace they're looking at.
• Stays fresh while you watch — leave a dashboard tab open all afternoon and it keeps working, even across multiple idle-pause cycles. A background refresh silently rotates the underlying signed URL every few minutes.
• Short-lived tokens under the hood — the signed sandbox URL lives only inside the iframe and rotates every 5 minutes, so leaks from browser history, referrer headers, or shoulder-surfing the address bar are worthless. Copy-link always gives you the permanent, auth-protected URL.

• Retrospective nudge — when a premium ($$$) turn finishes short (under ~6,000 chars in, under 500 tokens out, no tool calls), a small amber banner appears under the reply showing what this turn cost and what a cheaper model would have cost for the same token shape, with a one-click Switch button that flips the rest of the thread. The cheaper suggestion stays in-family: Opus and Sonnet 1M nudge to Sonnet, GPT-5 Pro nudges to GPT-5 Thinking. Dismissible per message.
• Responsibility warning — picking a $$$ model as your thread-start default now shows a red card right under the model dropdown, spelling out the cost commitment and suggesting a $$ model as the default with premium reserved for hard turns via the sidebar picker. A ($$$) suffix is also added to premium entries in the dropdown so the tier is visible before you commit.

• Caveman — drops articles, filler, pleasantries, and hedging. Fragment-friendly. "Inline obj prop. New ref. Re-render. useMemo fix."
• Caveman Ultra — telegraphic. Abbreviates common terms (DB, auth, fn, conn, repo, deps…), strips conjunctions, replaces "causes / leads to / then" with → arrows. One word per idea where possible.
• Caveman 文言 (Wenyan-Ultra) — responds in ultra-compressed Classical Chinese, the most token-efficient written language humans ever invented, with code / commands / URLs / error messages kept verbatim in their original language and mixed inline. Roughly 80–90% character reduction versus English.

• Permanent (default) — flips the thread's picker too, so follow-ups stay on the new model.
• One-off — routes this one message only; the picker snaps back afterwards.

• Pick the workspace image — Standard (1 CPU / 1 GB) for most tasks, or Heavy (2 CPU / 2 GB) for bigger dataframes, ML models, and multi-process pipelines.
• Flip durable vs ephemeral independently of your global default — some threads benefit from a clean slate each time; others want files and state to stick around.
• Destroy the current sandbox inline — the trash icon next to the refresh button on the workspace block in the chat kills it cleanly. The picker unlocks immediately so you can switch image or lifecycle before the next turn spawns a fresh sandbox.

• Excel — only the first 10 sheets and first 100 rows per sheet are read; skipped sheets and rows are called out in the preview.
• CSV — the parser stops after 200 rows instead of materialising the entire file just to slice it.
• PowerPoint — capped at 50 slides in the preview, with the rest called out.
• Anything genuinely too large (e.g. a 100 MB PDF, 50 MB CSV, 40 MB PPTX) now shows a clear note telling the assistant to open the file in the workspace with the right tool (pd.read_excel, pd.read_csv, pdfplumber, …) — so big files land in the right tool for the job and the rest of the thread stays snappy.

• If a reply drops mid-stream, the extension quietly reconnects and resumes the same answer — no duplicated tokens, no lost partials.
• If the provider is overloaded (Anthropic 529), the banner now explains what’s happening and tells you to wait a few seconds before clicking Retry, instead of a cryptic "Model error: Overloaded".
• If we push a deploy while you’re in the middle of a turn, you see "Server restarting — reconnecting to a fresh instance…" instead of a connection-lost scare.

• Fixed a race between the sidebar and the service worker that could (rarely) invalidate your session family and force a fresh SSO login. Token refreshes are now serialised through a true atomic lock across all extension contexts, so the sidebar, the standalone page, and background requests can’t trip over each other.
• Your profile is no longer overwritten every time you sign in again. Custom profile pictures, edited display names, and assigned roles are preserved across sign-ins instead of being reset to whatever Auth0 sent.
• Failed or aborted replies no longer leave zombie assistant bubbles in the thread history.
• Fixed the typing animation "ghosting" across messages — after you stopped a reply and started a new one, both bubbles would animate as if still typing. Only the currently-streaming message animates now.
• Fixed a 401 that could hit retries on long-lived sidebars — the chat transport now refreshes its auth token on every request the same way the rest of the app does.

• On a schedule — "every weekday at 9am", "first of the month", whatever cadence fits. The agent wakes up, does the task, and is there waiting for you when you check in.
• From a webhook — POST JSON to a unique URL and the agent reads the payload and acts on it. Great for connecting Zapier, n8n, GitHub Actions, or anything else you already have.
• From an email — every automation gets its own email address. Forward a message there and the agent reads it (subject, body, attachments) and replies.
• From a /command — bind an automation to a shortcut so triggering it is a three-keystroke affair.
• Run history for everything — every fire is recorded with when it happened, how long it took, whether it succeeded, and what the agent produced. Replay, inspect, or pick up a thread to keep chatting with the result.
• Optional email notifications — flip a switch per automation and get an email when a run finishes.

• Manage them from Settings > Commands. Stuck on what to write? Draft a command with AI from a quick description.
• Share commands with your whole organisation so the team stays consistent.
• Off by default. Flip the master switch in the extension sidebar when you're ready — password and one-time-code fields are always skipped.

• Create, rename, and delete projects from the sidebar
• Move any thread into a project via right-click
• Share a project with your entire organisation or specific people
• Shared projects are read-only for viewers — they can browse but not modify
• Visual icons distinguish owned, shared-with-org, and shared-with-user projects

• Projects, pinned threads, and recent history in one unified panel
• Search across everything — projects auto-expand to show matching threads
• Right-click context menus for rename, pin, share, move, fork, and delete
• New threads and name changes appear in real time via WebSocket

• New permissions: Chat.ReadWrite, ChatMessage.Send, ChannelMessage.Send, Mail.ReadWrite
• Re-authorization required — go to Settings > Credentials and reconnect your Microsoft 365 account to grant the new permissions

• Pick up exactly where you left off — your files, data, and progress are still there when you come back
• No more re-uploading files or re-running setup steps when resuming an old thread
• Works seamlessly — just keep chatting and the workspace wakes up on its own
• Pause and resume workspaces manually from settings when you want full control

• Define a name, icon, system prompt, model, and tools
• Choose "Base + custom" mode to inherit formatting, safety, and citation rules automatically — or go fully custom
• Pick a behavioral role (General, Agentic, Writer, Explorer, Researcher, Executive)
• Use "Draft with AI" to generate a first draft of your prompt from a short description
• Auto-generate a description and conversation starters with one click
• Duplicate agents to quickly create variants
• Preview the full combined prompt in real-time
• Your agents appear under "My Agents" in the extension's agent picker