Every release, in order.

• Ambient Acme panel — summon the platform helper from anywhere in the web UI; it floats instead of shifting layout, resumes the last session, and shows an active state in the sidebar. Chat engine extracted into a shared useChatSession so the panel reuses attachments + streaming. CurrentViewContext lets routes publish the focused entity, surfaced to Acme as a data-ui-context part (latest-only materialization) with a UI-context chip.
• Per-agent email — each agent gets its own identity and isolated mailbox, with global IMAP/SMTP connection defaults, a Settings → Email tab, live config, and an editable OAuth redirect-URI override.
• /skill mentions — reference skills inline in the composer via a data-skill-ref part with per-message materialization.
• Live config reload — on-disk edits to config.yaml / AGENTS.md / skills apply live by evicting cached agents; global settings (model/email) reload without a restart. Replaces the config watcher with an explicit Acme-only reload_config tool; effective rw-path grants let Acme edit platform surfaces. Only server.host / server.port still need a restart.
• Fixes — web_search no longer shows "0 results" for content-blob providers; no redirect to /login on a 401 while already on an auth page.
• Refreshed the shipped Acme platform agent; added blog + changelog to the docs site, community health files, and color brand logos.

• Local installs are frictionless again. On a loopback bind (the default) the daemon auto-establishes a real session for loopback requests, so opening http://127.0.0.1:3456 lands straight in the app — no claim page, no login form. Auth stays on under the hood (a real local@localhost operator + session). The per-request Host header is the gate: a non-loopback Host (DNS-rebind, or a tunnel/proxy) is never auto-trusted.
• Sharing requires a login. openacme expose rewrites config to bind the network + require an account, then restarts — one command covers tunnels, reverse proxies, and direct IP. openacme expose --off reverses it. Exposing rotates the local operator out so its loopback cookie can't be replayed remotely. (server.requireAuth forces a login on a loopback bind for the rare tunnel-without-an-open-port case.)
• Web member admin (Settings → Members). For shared installs: signed-in-as + sign out, the member roster + revoke, and one-click invite-link generation — previously CLI-only. Deployment-mode aware: local installs see an "invite your team" explainer with a docs link instead.
• Reachable claim/invite links. openacme claim / invite and the boot log no longer print localhost when shared — they use a detected address (or a clear placeholder) and hint to substitute your domain/public IP.
• A branded 404 page, and a docs link in the sidebar.

#35 1582351 Thanks @ukanwat! - Fix a crash when upgrading from 0.8.x: the WASM SQLite engine can't open a database left in WAL mode by the previous (better-sqlite3) engine, so the daemon failed on first query with unable to open database file.

createDatabase now detects a WAL-mode file on open and converts it in pure JS — it checkpoints the committed WAL frames into the main database, switches the header to rollback journal mode, and removes the -wal/-shm sidecars. No native code and no data loss; verified byte-for-byte against better-sqlite3's own checkpoint. The conversion is a one-time no-op for databases the WASM engine already owns.

• SQLite: replace better-sqlite3 (native) with node-sqlite3-wasm (pure WASM). The stores keep driving the DB through a better-sqlite3-shaped adapter + a vendored drizzle session over drizzle-orm/sqlite-core, so query logic is unchanged. Migrations reuse drizzle's own readMigrationFiles + dialect.migrate, so existing databases are recognised and only pending migrations apply. WAL is unavailable on the WASM VFS; the connection uses busy_timeout instead (immaterial for a single local daemon).
• Image processing: replace sharp (native) with jimp (pure JS) for screenshot downscaling.
• Browser: move camoufox-js (pulls native better-sqlite3/impit) to optionalDependencies; it is already lazy-loaded with a graceful fallback, so a failed install no longer breaks the CLI.
• CLI: openacme start no longer crashes on headless boxes when no browser opener (xdg-open) is present — browser auto-open is now best-effort.

#29 f30a4c0 Thanks @ukanwat! - Release 0.8.0 — a large batch of work since 0.7.0 across the whole workforce platform. All @openacme/* packages move together (fixed group).

Highlights:

• Per-member authentication. Replaces the single shared access-secret with email/password accounts (flat-role admins), stateful sessions, first-run claim via a one-time setup link, and out-of-band invite links. No loopback bypass; /api/health stays unauthenticated. CLI: claim, invite, members list|revoke. Breaking: existing installs re-onboard once (the old secret file is ignored; first boot prints a claim link).
• Usage ledger — per-call token/cost metering, /api/usage, and a /usage page (overview, breakdown, activity, forecast).
• Teams org chart and a tasks-board UX overhaul (searchable filters, smoother kanban drag, humanized times, board-mode detail sheet).
• Inline markdown editing (TipTap) across web panes.
• Landing page + docs site (Fumadocs static export), including the remote-access and self-hosting guides.
• Latest model presets refresh.
• Real end-to-end test suite (HTTP + browser) behind a no-mock model seam.
• Default server port changed 3210 → 3456.
• Tool-host scaling fixes: lazy worker spawn, cached stdio MCP discovery, bounded per-session shells.

Mobile-ready PWA with web push notifications, and editorial workflow refinements.

This release wires ping_user agent events through to native mobile notifications and lands a full mobile-responsive UI pass so the operator can run the workforce from their phone.

Push pipeline (@openacme/db, @openacme/server)

• New push_subscriptions table + PushStore for per-device endpoints (single-operator deployment, unique endpoint upsert).
• PushDispatcher fan-outs every ping_user event to subscribed devices via web-push, with 404/410 endpoint cleanup. VAPID keys auto-generate to <dataDir>/push-vapid.json (mode 0600) on first boot.
• New routes: GET /api/push/vapid-public-key, POST|DELETE /api/push/subscribe, GET|DELETE /api/push/subscriptions, POST /api/push/test.
• Auth middleware whitelists /sw.js, /manifest.webmanifest, and PWA icons pre-login so iOS can fetch the manifest before a session cookie exists.
• VAPID subject defaults to a valid mailto URI (Apple's push service rejects .local domains with 403).
• Service worker uses renotify: true so same-tag pushes still alert; test pings use a unique tag per fire.

Web app: mobile responsive + PWA shell

• Bottom tab bar replaces the hamburger drawer on mobile; sidebar is desktop-only.
• Manifest, hand-rolled service worker (push event + notificationclick with includeUncontrolled: true), generated icons, apple-touch-icon.
• Master/detail layouts on /agents, /tasks, /skills, /settings column-stack on mobile with a back-to-list pill.
• Task dialog goes full-takeover above the tab bar on mobile.
• iOS standalone-PWA auth fallback: secret is also stored in localStorage and injected as Authorization: Bearer on every API call, so cookie eviction between PWA launches doesn't force re-login. Login page silently re-authenticates from the stored token.
• Service worker auto re-subscribes to push on every launch when permission is already granted (handles iOS subscription eviction).
• One-tap "Enable notifications" prompt on first PWA launch.

Memory (@openacme/memory)

• DEFAULT_MEMORY_CHAR_LIMIT raised from 2200 to 4000 — accommodates ~60-80 tight one-liner index entries before consolidation pressure kicks in. Per-agent override via memoryCharLimit frontmatter unchanged.

Tools (@openacme/tools)

• Removed web_upload built-in. It only served one workflow (catbox → URL for Buffer's createPost). Agents that need catbox upload should configure a small stdio MCP server via per-agent mcpServers — keeps the third-party host boundary visible in the agent's frontmatter rather than bundled platform-wide.

@openacme/* → 0.6.0

Highlights since 0.5.3:

• Multimodal read_file — images render inline in chat; screenshots from browser_take_screenshot flow through the same path.
• Browser overhaul — pluggable providers (local Chrome, Browserbase, Browser-Use, Firecrawl), per-agent sessions, auto-provisioned Browserbase contexts, tool-result spill to attachments.
• Agent-scoped session_search — full-text search now scoped to the caller's agent; no cross-agent leakage.
• Rename-swap compaction — preflight + UX fixes; dead fork bookkeeping removed.
• Web design pass — Cmd-K palette, workforce status, signal-blue meta, bounded search + FTS5 endpoint, agent filter polish.
• Auth picker with provider-availability gating; upstream provider errors surfaced in chat UI.
• Software Engineer agent template rebuilt with a real SWE persona.
• Fixes: ChatGPT OAuth (two fixes), Browser-Use /api/v2 profile auto-create, context-1m beta dropped on OAuth path, web behind reverse proxy.

Make camoufox-js an optional dependency and whitelist native builds for pnpm 10+.

Two install-blocking bugs in 0.5.0:

1. camoufox-js → impit@0.13.1 carries "preinstall": "npx only-allow pnpm", which breaks every npm-based global install of @openacme/cli. Camoufox is one of several browser providers (chromium / browserbase / browser-use / firecrawl all work without it) and the browser code already lazy-imports camoufox-js with a try/catch (packages/browser/src/binaries.ts). Moved to optionalDependencies so failed installs don't fail the whole tree.
2. pnpm 10's strict build-script policy silently skips native module builds, so better-sqlite3 never compiles → @openacme/db crashes on import. Added pnpm.onlyBuiltDependencies: ["better-sqlite3", "impit", "protobufjs"] to @openacme/cli's manifest so pnpm honors the build at install time without pnpm approve-builds -g.

After this release, both npm install -g @openacme/cli and pnpm add -g @openacme/cli produce a working daemon.

Release 0.4.0.

Highlights since 0.3.0:

• Browser tool: managed Chrome via CDP with shared user-data-dir and per-agent tab ownership; ten browser_* tools.
• Tasks v2: comments + events split out of the task body into SQLite; pure event-driven scheduler with debounced wakes, echo suppression, lazy session allocation, recurring self-reset, and mid-turn event injection.
• Agent catalog: bespoke in-tree templates importable into the workforce (CLI agents catalog / agents import, web modal). Ships the Coder and Acme platform templates.
• Skills Hub: install + track skills from GitHub, marketplaces, URLs, .well-known, LobeHub, Skills.sh, ClawHub, local dirs, and a new builtin source.
• AGENTS.md: shared workforce context injected into every agent's prompt; restart-free updates via cache eviction.
• Per-agent workspace + resources: <agentDir>/workspace/ as default cwd with a session-persistent shell, and <agentDir>/resources/ listed in the prompt.
• First-run setup wizard: provider credentials, model seed, agent creation — web + CLI.
• SSE-only streaming for interactive turns: agent runs are server-owned, the originating tab is just another subscriber.
• Pino-backed structured logger with OTel log export.
• LLM-generated session titles via a structured subagent.
• Operator home page with live SSE, plus ping_user / sleep primitives.
• Workforce framing: role + agent_list tool + peer-notes memory; silent OAuth recovery via Claude Code re-import.
• Design refresh: four-color signal system, unified task activity timeline, polished OAuth callback, CLI per-tool rendering with green/red diff backgrounds.

Release 0.4.0. First publish — synchronized with the rest of the workforce.

• Per-agent persistent memory (@openacme/memory) and per-agent task store + scheduler (@openacme/tasks) — both filesystem-backed, scoped by agent id. New task_* and memory built-in tools.
• AI SDK v4 → v6 migration end-to-end. UIMessage is now the canonical shape across DB rows, persistence, agent input, server response, and web render. Web uses @ai-sdk/react's useChat + DefaultChatTransport; server uses createUIMessageStream over Agent.runStream. Zod 3 → 4.
• Anthropic prompt caching for native and OpenRouter Claude paths; cache markers preserve string content as string when marking cacheable.
• MCP: first-class Streamable HTTP transport with OAuth 2.1 client, plus cwd for stdio servers. Web + CLI editors for per-agent MCP server config.
• Web UI: paper-aesthetic redesign, theme toggle (system/light/dark) with light-mode contrast pass, single-URL dev (Hono fronts UI on :3210 via proxy to Next; published serves the bundled static export from the same port), proper date-time picker and searchable assignee in the task modal, inline custom views and git-style diffs in tool rendering, attachments via picker / drag-drop / @-fuzzy file picker.
• CLI: cancellable turns, credential-aware model pickers, windowed session picker that surfaces agent name, background daemon mode with secret-auth for non-loopback access, consolidated pnpm agent <subcommand> aliases.
• Server: RFC 5987 Content-Disposition for non-ASCII attachment filenames; /api/health now reports the actual server package version instead of a hardcoded string. Always-on system tools (agent introspection / self-management) merged into every agent's effective tool set and hidden from the user-facing tool picker.

Anthropic Agent Skills standard + agent-loadable skill bodies.

• @openacme/skills parses canonical top-level frontmatter (tags, related-skills) while still reading legacy metadata.hermes.*. Skill folders are walked at load time so companion files (scripts/*, references/*, …) are recorded as resources without being read until requested. New parseSkillDirectory + Skill.resources/Skill.dirPath.
• @openacme/tools ships a new skill_view built-in (Level 1 progressive disclosure) bound from the server. Returns the SKILL.md body, the on-disk dir path, and the resource list — agents read companion files via the existing read_file / shell tools.
• @openacme/server exposes POST /api/skills/import for multipart folder uploads (path-traversal guards, 200-entry / 10 MB cap, top-prefix stripping) and binds the skill registry into skill_view.
• @openacme/cli adds openacme skills list|view|add|remove and a /skills slash command + read-only overlay in the TUI.
• @openacme/agent-core system prompt now points the model at skill_view.
• @openacme/config adds skill_view to the default agent tools array.

All @openacme/* packages bump together (changeset fixed group) so users always get a uniform version across the workspace.