Changelog

Added

• Chain templates with role-based participation. Multi-step facilitation pipelines — Delphi panel, Six Thinking Hats, Wardley mapping — where each step can require specific roles. Output of one session feeds the next as context. Auto-creates a project to hold the chain.
• Public chain navigation. Visitors to a session’s results page can walk every step of the chain to see how the collective output evolved from stage to stage.
• Wardley mapping output. The Wardley Mapping chain produces a rendered map as its terminal artifact — anchors and components positioned along visibility/evolution axes, with evolution arrows and strategy decorators. Hosts can correct positions with an edit UI.
• Review my session. Pro-gated AI review of facilitator behavior on every finished session. Surfaces patterns — terse turns, missed follow-ups, compound questions — with evidence excerpts, and proposes additive prompt rules you can apply into Session Design in one click.
• Grounded summaries with verbatim quotes. Summary claims now show the participant quotes that back them, pulled directly from transcripts. Available as a host-triggered backfill on the results page.
• Host notification center. Bell icon + unread badge for chain progression, action-required, and completion events. Async multi-actor flows you don’t have to babysit.
• Chain step-ready emails for participants. When a chain reaches a step that requires their role, assigned participants get a signed-link email with the session topic, project, and a one-click join button. No more pinging people manually when it’s their turn.
• Edit history for session prompts. Every change to the facilitation prompt is tracked — who, when, which UI surface (including which AI-suggestion the change came from).
• Methodology-aware session creation. Refined creation flow with structured methodology pickers and per-field AI-assist.
• Suggested next session card. After a session ends, the results page surfaces a one-click follow-up that pre-fills the same project and brief shape.
• MCP and chat skill catch up to the platform. harmonica-mcp now passes the full PATCH allowlist through update_session (including widgets, results visibility, summary prompt, host context, and template metadata) and accepts widgets_enabled + results_visibility on create_session. harmonica-chat was rewritten as a thin entry file with per-subcommand reference docs loaded on demand, plus design-flow steps for widgets and results visibility.

Changed

• Choose who sees results. A new results_visibility field on session creation lets hosts pick whether the results page is public, link-only, or restricted.
• Role-weighted synthesis on chains. A chain step’s roles can carry a weight tier (normal / elevated / lead). Synthesis leans toward the higher-weighted voices, with a k-anonymity guard so no individual is ever identifiable as the source of the lean.
• Facilitator no longer stalls on terse answers. A guard nudges the facilitator to follow up when participants give short responses early, so conversations stay productive.
• Faster session launch. Welcome message and metadata generation moved off the launch critical path; the launch spinner clears the moment the session is reachable.

Fixed

• iOS Safari message drop. When a participant backgrounded the chat tab while the facilitator was responding, the message sometimes never appeared. The chat now detects the dropped fetch and recovers automatically.
• Enter on mobile inserts a newline instead of sending. Mobile keyboards’ Enter key now adds a line break; sending requires the dedicated send button.
• Sign in with a different account. Auth0 returned an invalid_request when hosts tried to switch accounts. Resolved.
• Sources step no longer blocks Next. Adding a context source ran a prompt-merge that could block the wizard’s Next button until it finished. Now non-blocking.
• AI Suggestion timeout. A hung Suggestion fetch could spin forever; now hard-timed-out with a graceful retry.
• Session-page denied surfaces. Reframed around access (you don’t have permission) instead of chat (you can’t message here), matching the actual state.
• Export modal hanging. Date columns in internal API responses are re-hydrated, so export tabs don’t get stuck on stale state.

Added

• Custom user avatars. Upload your own profile picture instead of being stuck with initials.
• Session Design tab. All the layers that shape a session — HARMONICA.md, platform guidelines, facilitation approach, summary generation — live in one redesigned panel.
• Skippable onboarding. New hosts can skip the /welcome form and re-enter it later from Settings whenever they have the time.

Changed

• Settings consolidated to four tabs. Profile, billing, AI models, and Harmonica API keys.
• Sessions default to Active. The dashboard now opens on the sessions you’re currently running, with Draft and Finished one click away.
• Session creation: grouped picker. Templates and chains in a single grouped list with a compact “what is this?” preview.

Fixed

• Password reset email. Now routed through Auth0’s SMTP relay with the same Loops templates as the rest of the transactional mail.
• View Results link in transcripts. Email transcripts always include a direct link back to the session results page, in every milestone template.

Added

• Widgets in chat (initial). The facilitator can emit polls and ratings inline during a conversation. Phase one establishes the structured-chat substrate; future template tuning will activate widgets per methodology.
• Brief regeneration. When you edit a session’s brief, Harmonica offers to regenerate the facilitator prompt and recap to match — pulling in your project memory and active template.
• Video intro on the landing card. Hosts can paste a video URL into the Session Design Landing tab; it renders as a thumbnail-and-play card above the welcome copy.
• Workspace-level invites. Inviting a collaborator to a session also adds them to the project’s collaborator list. They get follow-up invites on future sessions in the same project automatically.

Changed

• Warmer end-of-chat. When a session ends, participants see a conditional Results link with warmer copy instead of a generic “thanks for completing” wall.

Fixed

• Welcome message links. Markdown links render correctly in the welcome card, and bare URLs auto-link so participants don’t have to copy-paste.

Added

• Three new picker models. Gemini 2.5 Flash-Lite (Free), Claude Haiku 4.5 (Pro), and Gemini 2.5 Pro (Pro) join the model picker. Existing defaults are unchanged; pick what fits the session.

Changed

• Live-transcript voice input. The mic button now streams interim transcripts and a waveform as you speak. Useful for accessibility and mobile contexts.
• Bring Your Own Model (BYOM) hardening. Hosts running Harmonica on their own LLM keys get clearer error handling, better fallback behaviour, and a tidier model-selection flow.
• Sharper summaries. Pro summaries now run on the LARGE tier (Claude Sonnet 4.6 by default) instead of MAIN — slower by a couple of seconds, noticeably better at long, multi-thread synthesis.
• Welcome card. Cleaner copy, less clutter, faster path to the first action.

Added

• Organizational context (HARMONICA.md). A markdown file scoped to you as a host is now wired into session creation, so every new session inherits your team’s vocabulary, decision-making norms, and standing context without copy-paste each time.
• Context Sources. Attach reference material to a session — files (chunked and embedded for retrieval), prior Harmonica sessions, or MCP servers. The facilitator draws on them during the conversation.
• AI-drafted welcome message. New sessions ship with a tuned welcome message generated from your brief, plus link-preview metadata for shared URLs.

Changed

• Cross-pollination never blocks responses. If the cross-pollination check takes too long, the facilitator replies anyway. Conversation flow comes first.

Fixed

• Customer billing lockout when a subscription event arrived out of order
• Edge cases in upsert that occasionally reset paid-tier settings on sign-in

Added

• Public results pages. The results page is now a public-facing product page in its own right — collapsible summary sections, a stats bar above the thread, and a layout meant to be shared with people who don’t have a Harmonica account.
• Invitation landing page. When someone invites you to a session as a co-editor or viewer, you now land on a preview of the session with one-click accept or decline — no more cold links into half-loaded results.
• Summary tools in the MCP. harmonica-mcp gained generate_summary and summary_prompt so agents can trigger a synthesis on demand and tune how the recap is written.
• Participant tools in the MCP. New list_participants tool + per-participant filtering on get_responses. Hosts can also set cross_pollination and distribution via update_session.

Changed

• Transactional email through Loops. Welcome, invitation, and milestone emails now flow through a single dedicated provider with consistent templates.
• Manage access shows names and roles. The collaborators panel surfaces email under display name and always shows each member’s role.

Added

• Smarter cross-pollination. Responses are now clustered before being offered across threads, so insights shared between participants are higher-signal and less repetitive.
• Session scratchpad. The facilitator builds and refines its understanding of the conversation as it goes, adapting later questions to what it has already heard.
• Pre-session questions in the MCP and chat skill. harmonica-mcp create_session accepts a questions array; harmonica-chat gained a configurable pre-session questions step.
• Telegram group selection. New list_telegram_groups MCP tool lets agents enumerate connected Telegram destinations when creating a session.
• Lifecycle commands in the chat skill. Added /harmonica-chat edit (update session metadata) and /harmonica-chat review (analyze transcripts for facilitation issues + propose prompt fixes).

Changed

• Resilient MCP client. harmonica-mcp now retries with backoff on rate-limited API responses instead of failing the tool call.

Fixed

• Session prompt edits propagating correctly through the builder
• Form-field handling regressions during session creation

Added

• Public REST API v1. Bearer-token authentication with hm_live_… keys gives integrators programmatic access to session creation, response collection, summaries, and chat. Manage keys from your profile.
• harmonica-mcp. MCP server published to npm. Any AI agent that speaks MCP can create and query Harmonica sessions.
• harmonica-chat v2.0.0. Claude Code companion for the session lifecycle — guided design, accelerated creation, status checks, follow-ups, and invitation drafting.
• help.harmonica.chat. New docs site with API reference, MCP server docs, and host guides.

Changed

• Security updates. Dependencies patched for known vulnerabilities. No behavioural changes for hosts or participants.

Changed

• PDF uploads now accept files up to 10MB. Previously capped at a size that rejected most multi-page reports and recorded transcripts.

Added

• Edit the facilitator prompt from session settings. You can now tune how the AI facilitator opens, probes, and closes a conversation without leaving the session view.
• Edit the summary prompt directly from session results. Rephrase the synthesis question, regenerate, and compare in place.

Changed

• Session results UI redesign. Cleaner hierarchy between the synthesis, the raw responses, and the per-participant threads.
• Templates page revamp. Back navigation, clearer template cards, faster path to “use this template”.

Added

• Session templates for common deliberation types. Retrospectives, brainstorms, stakeholder consultations, hiring debriefs. Each template ships with a tuned facilitator prompt and synthesis prompt, ready to run.
• Improved session creation flow. Faster path from “I have a question to ask a group” to a sharable session link.

Added

• Voice typing in the chat input. Participants can speak their responses instead of typing them. Falls back gracefully when the browser doesn’t expose microphone access.

Changed

• LLM model is configurable per tier. Hosts can route the small/main/large tiers to different providers, with Gemini as a fallback when the primary is overloaded.
• Cross-pollination defaults to off. Hosts opt in per session rather than out, after participants told us they expected stricter isolation by default.

Added

• Multi-step session creation flow. The first step asks what you’re trying to learn, before anything about format. New users land in a guided intro instead of a blank form.

Added

• Complete chat overhaul. Mobile-first redesign of the participant experience, with cleaner threading and better keyboard handling.
• Persistent chat state. Conversations survive page reloads and reconnections instead of restarting.
• Share settings. Configure who can view, edit, and participate per session.
• Real-time summary updates. Hosts see a colored indicator while the summary is regenerating, with debounce so frequent edits don’t thrash the model.
• Exclude participants from summary. Toggle a participant on or off without losing their thread.
• Quick-create from dashboard. Skip the templates page when you know what you want.
• Email invitations via Gmail SMTP. Replaces the previous third-party sender.

Changed

• Projects sort by last modified. Recent work surfaces first.

Fixed

• Mobile chat autofocus removed (was zooming the page on iOS)
• Session invite encryption
• Chat scroll behaviour on long threads
• Response export for sessions with many participants
• Several form-state and styling regressions

Added

• Session rating. Participants can rate a session at the end, giving hosts a quick read on how it landed.
• Editor role can add sessions to projects. Project owners no longer have to do this themselves.
• Owner permissions for viewing and editing sessions. Cleaner separation between who can see what.

Changed

• Summary prompt editing moved to a dedicated dialog. Less crowded session settings, more space to compose.

Fixed

• 200-session cap on the dashboard
• Renaming projects from the profile page
• AI Suggestion regression

Added

• Cross-pollination intelligence is live. When relevant, Harmonica surfaces what one participant said to another, in their own thread, without naming them. The facilitator decides moment-by-moment whether sharing helps.
• Session knowledge system. Foundational version. Harmonica retains context and learnings from previous sessions in the same project.
• Improved project handling. Better continuity when running a series of workshops with the same group.

Added

• Cross-pollination between threads. Harmonica identifies opportunities to share insights across discussions and generates context-aware follow-ups. Early version, paving the way for the production launch in v0.9.
• AI Suggestion button. A lightweight, PM-style response generator next to the input field, powered by GPT-4o-mini.

Changed

• Gemini 2.0 as primary engine. Stronger reasoning, larger context windows, faster long-thread handling.
• Cross-pollination timing controls. Capped at most once every three minutes per participant to avoid disrupting the conversation.

Added

• Projects. A workspace view for organising and managing multiple sessions in one place. Migrated existing sessions over with cleaner organisation and improved access to historical workshops.

Changed

• Ignored-response handling. Better detection of off-topic or non-substantive replies so the facilitator can move on rather than chase them.
• Chat scroll and navigation. Smoother keyboard and scroll behaviour during long sessions.
• Chat results on the session page. No more jumping between tabs to see what each participant said.
• Results page polish. Cleaner layout, faster path from session to insight.

Added

• Data collection forms. Streamline pre-session input gathering — collect emails, contexts, or anything else before the conversation starts.
• Advanced user analytics. Engagement and completion insights for workshop hosts.
• Session timestamps and sort by creation date. Find recent sessions faster.
• Start-new-session links on results pages. Run a follow-up without going back to the dashboard.

Changed

• More resilient chat. Better error handling and higher timeout limits for slower models.
• Smarter templates. When you edit a session’s creation fields, the template recreates instead of going stale.
• Cleaner page titles and social metadata. Better link previews when sharing a session.
• Privacy controls. New session visibility limits.

Fixed

• Session loading and data fetching across pages
• Message ordering in chat
• Anonymous-user handling
• Chat export for multi-participant sessions
• Session-count accuracy on the dashboard

Added

• Personal dashboard. Secure cookie-based session memory so you don’t have to log in to come back to your own work.
• Role-aware session controls. Hosts see management options; participants see a streamlined chat. Same URL, different experience.
• Individual session result links. Per-participant pages for sharing specific responses.
• Real-time summary creation. Watch the summary form as participants respond — no waiting for the session to end.
• Session completion handling. Helpful prompts and a one-click “Create New Session” when a conversation wraps.
• Logged-out test mode. “Create Test Session” lets people try Harmonica without signing up.
• Direct result links. Share specific session outputs with one URL.

Changed

• Single-card summary display. Cleaner, less cluttered.
• Hidden empty transcripts. Participants don’t see noise where there’s nothing to show.
• Visible participant names. Where appropriate, surfaced in lists for context.
• Chat-style transcript formatting. Reads like the conversation instead of a database dump.