Manwe release notes

Manwe 0.9.93 · Skills, local tools and workspace polish

Sat, 31 Oct 2026 12:00:00 GMT

Find and use skills from the composer. Type $ to browse active skills, filter by name and insert one without sending your message. The library now includes 57 ready-to-use skills and links to eight external collections.
Discover useful skills before enabling them. Advisors can find relevant inactive library skills and offer Activate and continue or Not now. Natural-language searches now find matching skills reliably, including exam revision and SEO audits. Your choice and the skill’s scope still apply.
Add tools when a task needs them. Advisors can discover optional local tool packs, offer installation when approval is required and continue afterward. The catalogue shows download and installed sizes so you can review the space needed.
Transcribe recordings locally. Create TXT, SRT, WebVTT and JSON transcripts from audio or video using Apple’s on-device speech recognition where supported. Transcript fragments are joined into readable paragraphs and timed subtitle phrases, with Stop support.
Make scanned PDFs searchable. Create a separate PDF with selectable, searchable OCR text while preserving the original document. Pages without recognized text are reported.
Remove downloaded models. Delete supported Manwe-managed models from Settings even when they are not selected. Removing a link to weights you manage elsewhere keeps those files in place.
Keep your workspace comfortable. Fixed window-size restoration through the real app lifecycle, added a gap between sidebar rows and softened sidebar-resize and composer-suggestion hover feedback.
More consistent modals. Sheets use the selected theme, a quiet border and clearer header, footer and instruction-preview surfaces. Fixed native sheet sizing and dimming cleanup when a modal closes or reopens.
A smaller bundled runtime. Removed rebuildable Python bytecode from the installer while keeping the bundled runtime and its tools available.

Build 31 · Developer ID signed and notarized by Apple.

Manwe 0.9.92 · Reliability, workflows and workspace improvements

Sat, 31 Oct 2026 12:00:00 GMT

Images reach your model. Fixed Claude and Codex image input for attachments, rendered documents and browser screenshots. Models can inspect the pixels through Manwe’s existing scoped tools.
Keep Deep Review progress. Collected evidence and provisional findings are checkpointed during a review. Reaching a time limit preserves useful work instead of discarding the run’s progress.
Task time limits are optional. The overall time limit is now off by default across chats, Deep Reviews and other tasks. Enable one in Settings → General → Task limits when you want it. Individual model-response timeouts remain in place.
Build workflows with their context. Improved workflow setup provides more room for instructions, links and references, plus a working folder. Saved references and World context are used before asking you for missing details; per-run inputs can override them.
More scheduling choices. Run workflows hourly, monthly, at several times in a day, or on selected weekdays. Condition triggers create their linked watch automatically, including supported price-threshold checks. Improved time controls and upcoming-run previews make schedules easier to review.
Find active conversations faster. New activity moves conversations to the top of their group. A Recents section brings conversations from across Worlds and agent chats together.
A more comfortable workspace. Window size and position are remembered. The sidebar has a wider resize target, hover feedback and double-click reset. Sheets can close when you click outside, and common source files—including Astro files—open in Manwe’s code editor.
Updates are easier to spot. An available update appears as an accent-coloured button above your profile, with a direct path to the existing update window.

Build 30 · Developer ID signed and notarized by Apple.

Manwe 0.9.91 · Deep Review fixes and improvements

Fri, 30 Oct 2026 12:00:00 GMT

Review code where it lives. Deep Review can inspect the selected working folder through read-only file and repository tools, including paginated searches across large repositories.
Keep useful findings. When some requested areas remain unresolved, supported findings and evidence are saved as a clearly marked partial review. A final coverage check can send the researcher back to investigate gaps before finishing.
Continue when more research is needed. At the research allowance, Manwe explains what remains and asks whether to continue or save the findings so far. Continuing extends only that review and keeps the evidence already gathered.
Choose your starting allowance. New Deep Reviews start with 512 research calls. Change it in Settings → General → Deep review → Research allowance, or find “research allowance” in the command palette. Options are 128, 256, 512, 1,024 and 2,048 calls. Ordinary chat execution limits remain separate.
Clearer report titles. Partial reviews and their exports use a descriptive subject title. Completion status is shown separately, rather than becoming the report headline.
Respect changing access. Changing or revoking a review’s working folder cancels the affected work and prevents late results from being applied.

Build 29 · Developer ID signed and notarized by Apple.

Manwe 0.9.9 · Meet the new Manwe

Fri, 30 Oct 2026 12:00:00 GMT

Manwe began with a simple idea: important decisions deserve more than a single AI answer.

With 0.9.9, Manwe becomes an AI workspace built around ontology, self-learning, and action. Connected knowledge gives your work context. On-device neural networks help recognize patterns in your feedback. Persistent advisors carry lessons forward. Tools and workflows put that understanding to work on your Mac.

Research a question, develop an idea, bring in another perspective, and create something from the result. Return later and continue with the context you have built.

An ontology of your world

Your work includes people, projects, goals, constraints, and the relationships between them. Manwe’s ontology gives that connected knowledge a lasting home.

Worlds and Moons organize the context behind your companies, projects, research, and personal life. Setup asks questions suited to each World, helping you establish what matters and how it connects.

Manwe can draw on those relationships during subsequent work. A technology review can consider your company’s requirements. A project conversation can build on established goals and constraints. Personal preferences have their own place, separate from project memory.

The new memory map makes retained knowledge and its relationships visible. Explore what Manwe remembers, inspect its connections, and set memories aside when they should no longer inform your work.

On-device neural networks that learn from feedback

Your corrections can have a role beyond the current conversation.

Manwe uses small on-device neural classifiers to help recognize patterns in corrections and memory choices. These classifiers can train on retained learning signals, while repeated corrections can become preferences that inform future tasks.

This learning works alongside ontology memory: useful context can be retained, feedback can shape future guidance, and your choices can help determine what belongs in memory.

You control these systems. Undo automatic memories from the conversation, set retained context aside, require review before a World keeps new memories, or turn automatic remembering and learning from your work off.

Advisors with continuity—and time to learn

Advisors are persistent identities you can return to, with their own expertise, instructions, model choices, and learning history.

Work with one advisor, assemble a team, ask for second opinions, or commission a Deep Review. Teams can divide work, share findings, and bring results together. Find an advisor by name to continue working with them.

When background learning is enabled, advisors can research between conversations while Manwe is open, idle, and online, within your configured learning allowances. Their retained lessons can inform later work.

Put connected understanding to work

Conversations can now lead directly into work with your files, folders, browser, and connected apps.

Create and edit documents, spreadsheets, and presentations. Build interactive results, research a question, run terminal commands, search a repository, or work with the file or selection open in your editor.

Context from VS Code, Cursor, and Xcode, local Docker execution, and file access requested during a task extend what you can accomplish inside Manwe. Approval modes can also change as a conversation develops.

Choose local models, supported model subscriptions, or API providers. Your Worlds, saved work, and conversation history remain in Manwe as you switch models.

Turn useful processes into repeatable workflows

Save a process as a workflow, write one yourself, start from an example, or build one with an agent.

Run workflows on demand or on a schedule, and carry results into subsequent work. With the optional background helper enabled and approved in macOS, supported scheduled workflows can run after you quit Manwe while your Mac remains awake. Local-model work waits until Manwe is open.

Deep Review has also been rebuilt around your research question and its relevant context, with clearer progress and improved recovery from individual call failures.

Also arriving in 0.9.9

Skills: 48 bundled starting skills, plus your own reusable instructions, imports, and skills prepared from an answer or through MCP.
Saved Work: reopen, reuse, and revise generated files and interactive results, with retained versions.
Apple app integrations: work with Calendar, Reminders, Contacts, Shortcuts, Notes, Mail drafts, Maps, Photos, and Music, with their respective macOS permissions.
Manwe on your phone: pair a private Telegram bot to continue conversations and receive updates while your Mac is awake and Manwe is open.
Sharing and Spaces: styled exports, slide decks, shared pages and sites, and collections of a World's decisions and answers. Live links are served from your awake Mac.
Images and voice: supported image input and configured image generation, composer dictation, answer read-aloud, and translation.
A redesigned workspace: themes, fonts, text sizing, responsive panels, and a command palette for conversations, advisors, settings, and handbook guidance.

Updating from 0.9.7

Manwe creates an automatic recovery archive before upgrading your existing data. Conversations, reports, Worlds, and promoted memory are preserved.

The legacy reusable advisor roster is reset. Advisor participants recorded in historical reports remain preserved. You will be asked to choose your model and complete onboarding again.

To roll back, follow the instructions included with the recovery archive. Do not open upgraded data with Manwe 0.9.7.

Requirements

Requires an Apple Silicon Mac running macOS 14 or later. Some local intelligence features require a newer macOS version or separately downloaded models.

Cloud providers use your configured accounts or API keys. VS Code and Cursor context require Manwe’s editor extension; Xcode context requires Accessibility permission. Container execution requires a running local Docker Desktop engine.

Build 28 · Developer ID signed and notarized by Apple.

Manwe 0.9.7 · Radar & Vault

Sun, 05 Jul 2026 12:00:00 GMT

Adds Radar as a live research surface with model selection, saved watches, source lanes, and persisted signal history.
Adds Vault for uploaded and generated sources, including full-text reading, attach-menu reuse, and safe removal.
Adds light and dark appearance settings and broadens theme coverage across the app.
Polishes Diagnostic Trial reading surfaces, including reliable mouse-wheel scrolling and clearer live positioning.
Improves Scenario Editor spacing, hidden-card handling, and canvas readability.

Manwe 0.9.6 · Model Refresh

Thu, 02 Jul 2026 12:00:00 GMT

Removes Gemini CLI support from the model picker and runtime.
Adds Claude Sonnet 5 while keeping Claude Sonnet 4.6 available.
Restores Claude Fable 5 routing with the current Claude model identifier.

Manwe 0.9.5 · Diagnostic Trials

Mon, 15 Jun 2026 12:00:00 GMT

Manwe v0.9.5 adds Diagnostic Trials: multi-model trial rooms that let several Manwe teams attack the same question, cross-examine each other, and produce one pressure-tested decision record.

Diagnostic Trials

Multi-model trial rooms - run GPT, Claude, Gemini, Qwen, local MLX, or configured API models in the same trial instead of manually repeating the same question across models.
Teams, not solo answers - each model fields its own advisors, prepares evidence, argues rounds, writes a closing brief, and has its answer audited under pressure.
Forum owners - Auditor, Contrarian, and Depth Chair intervene as courtroom roles instead of ordinary advisors, forcing evidence checks, hidden-assumption attacks, and deeper-story synthesis.
Appeal existing records - start a trial from History to challenge a completed Manwe run, with the original advisors defending their record while challenger teams prosecute alternatives.
Decision-pressure matrix - compare branches by verdict, assumptions, evidence, unsupported claims, movement after audit, and falsifier probes. This is not a model leaderboard; it is a pressure map for the decision.

Trial Records

Dedicated Trials page - trials now have their own navigation, history, setup flow, running state, and completed-session view.
Live war-room recording - watch the trial unfold through preparation, debate rounds, forum interventions, closing briefs, audits, rulings, verdict sections, and final artifacts.
Evidence ledger - see prepared, cited, shared, single-team, disputed, and unsupported evidence in one place.
Watchpoints and contributions - final records include what the trial added, what would change the verdict, and accountability items for later review.
Unified exports - trial Markdown, HTML, and PDF exports now use the same visual language as normal Manwe records.

Memory And Advisors

Advisor-pool integration - trials can use the existing advisor pool instead of creating throwaway names, with category-aware selection and no memory mutation until a trial succeeds.
Original-record defense - appeal trials seed the defending team from the advisors and record that produced the original answer.
World context in trials - selected-world ontology context is carried into trial prompts with explicit provenance so teams can use known personal or project context without treating it as generic web evidence.

Models And Providers

Codex controls - added Codex model and reasoning controls for runs and trials.
Fable compatibility - improved direct Anthropic API handling for Claude Fable-style models.
Safer CLI behavior - hardened Codex CLI invocation and model-specific prompt handling.

Stability

Trials now persist early, resume without duplicating prepared artifacts, survive partial team failures when enough teams completed, and repair common malformed JSON responses instead of wasting an entire run.
Cancellation preserves completed trial work.
Trial setup rejects multiple local MLX teams when memory pressure would make the run unsafe.
Scenario canvas responsiveness is improved.
Existing users keep the same bundle identity, Sparkle feed, Application Support data path, fonts, local models, and database location.

The doctrine for 0.9.5 is simple: do not ask one model to be wise in private. Put the answer on trial, make the strongest versions fight, and keep the verdict useful.

Download the DMG below and drag Manwe to Applications. Existing users will receive this update automatically via Sparkle.

Manwe 0.9.0 · Ontology & Worlds

Thu, 14 May 2026 12:00:00 GMT

Manwe v0.9.0 adds the ontology layer: Worlds, governed memory, a readable context inspector, lifelines, and sandbox scenario runs.

Ontology & Worlds

World-scoped context - create reusable worlds for different parts of your life or work, then run Manwe through the selected context instead of one giant blended profile.
Governed memory lifecycle - suggestions stay inert, drafts stay inert, and only promoted Memory becomes context Manwe can use.
Selected-world isolation - promoted memory is only visible when one of its selected worlds is in context.

What Manwe Sees

Context inspector - inspect the exact profile, world, active memory, and relationship context Manwe can use.
Run preflight - Quick Run and Pro Run show what saved context will enter the prompt before the run starts.
Clear exclusions - suggestions and drafts are visible in review, but remain outside prompt context until promoted.

Memory Review

Memory inbox - review new suggestions from the home screen one card at a time.
Draft and promote flow - save promising candidates as drafts, then promote reviewed memory from the world review board or run results.
Relationship quarantine - cross-world relationship suggestions can be captured and reviewed without silently becoming graph canon.

Lifeline

Timeline view - canon memory, predictions, watchpoints, and projected branches now appear as a lifeline instead of disconnected cards.
Time-aware watchpoints - dated predictions can become upcoming, due, or resolved.
Outcome drafts - marking a watchpoint happened, partially happened, or did not happen creates an inert draft before it can become memory.

Scenario Playground

Sandbox runs - add temporary assumptions, ignore canon memory, force projected outcomes, and run a scenario without editing worlds or memory.
Canvas controls - pan, zoom, fit, reset, minimap navigation, and node inspection make world context easier to explore.
No automatic learning - scenario runs are saved in History, but do not create ontology suggestions or advisor learning.

Learn

New product guide - Settings now includes readable guides for Start, Worlds, Memory, Scenarios, Lifeline, Chats, Records, Research, and Models.
Debug empty-state preview - debug builds can preview onboarding empty states without deleting local data.

Stability

Hardened ontology lifecycle behavior across delete, rerun, continuation, multi-world promotion, and selected-world context isolation.
Added relationship endpoint validation and safer graph-promotion guards.
Fixed a Swift 6 concurrency warning in the Gemini backend.
Improved empty states and first-run surfaces for Worlds and the home memory inbox.

The doctrine for 0.9.0 is simple: flight recorder before autopilot. Manwe can reason with memory, but only the memory you let into the room.

Download the DMG below and drag Manwe to Applications. Existing users will receive this update automatically via Sparkle.

Manwe 0.8.0 · Market Intelligence

Sun, 03 May 2026 12:00:00 GMT

Manwe v0.8.0 adds Market Intelligence, Event Radar, a much larger research surface, inspectable computed evidence cards, and the first full Ask Manwe web workflow.

Market Intelligence

Market-aware decision records - market questions now connect fresh signals, filings, earnings context, price/volume history, and source-backed catalysts.
Clearer evidence judgment - reports are better at showing what matters, what is weak evidence, what made it into the visible analysis, and what still needs confirmation.
Computed evidence cards - timing context, earnings checks, historical reaction context, market screens, source-backed shortlists, and deeper evidence summaries are now inspectable.

Event Radar

Fresh signal scanning - Radar scans markets, companies, technology, policy, health, climate, geopolitics, and culture.
Saved watches - keep recurring watches and review recent signals.
Run from signal - open a normal Manwe run when a signal deserves deeper analysis.

30+ Source and Evidence Surfaces

Expanded research layer - Manwe now reaches more scholarly, legal, government, market, software, health, macro, and public-record sources.
Cleaner source handling - cited sources are separated from broader research context, and user-facing source lists avoid more generic, mismatched, or low-signal pages.

Ask Manwe Web

Real web workflow - sign in, run Quick or Pro, attach seed material, review saved records, inspect sources, and manage usage/billing.
Account controls - record deletion, account deletion, cancellation flow, and legal/settings links are now part of the product surface.
Computed source inspection - web records can open Manwe-computed evidence cards too.

Cleaner Decision Records

Openable computed sources - in-app and exported HTML evidence cards open instead of becoming dead links.
Friendlier artifacts - exported artifacts use clearer decision-record naming across HTML, Markdown, and PDF.
Better follow-ups - continuation records better understand whether you are clarifying, adding evidence, or changing the objective.

Quality of Life

Favorite advisors
More visible loading steps
Settings legal links
Improved guest expert and deeper-story behavior

Stability

Fixed continuation CPU spikes.
Fixed earnings evidence pinning.
Hardened web Pro run generation.
Improved web language and research query handling.
Fixed Codex model fallback behavior.
Reduced generic market-page noise in user-facing source lists.

Download the DMG below and drag Manwe to Applications. Existing users will receive this update automatically via Sparkle.

Manwe 0.7.0 · Deeper Reasoning

Sun, 19 Apr 2026 12:00:00 GMT

Pick how hard your models think. Start as a conversation, run when ready. The app now speaks the language of askmanwe.com — advisors, runs, decision records.

Reasoning Control

Per-model effort picker — Low / Medium / High / X-High / Max in Settings. Claude Sonnet, Claude Opus, Codex, Anthropic API, and OpenAI API. Higher effort = deeper deliberation, longer runs, sharper output.
Graceful fallback — backends without extended thinking silently ignore the setting. Claude CLIs older than 2.1.114 fall back transparently, no broken runs.
Reasoning badge on decision records (in-app, PDF, HTML, and Markdown exports) when a non-default effort was used.

Chat First, Run When Ready

Casual chat flow — think out loud with Manwe, invite up to 3 advisors from your pool, hit Run when you want a structured debate. The advisors you chatted with become the must-have panel.
Three chat modes — Smart (one advisor per turn, picked by relevance), Panel (every invited advisor responds in sequence), Let them talk (advisors continue the thread on their own).
Name autocomplete — type an @ to direct a question to an advisor, or to Manwe for the orchestrator's meta view. Keyboard nav with ↑↓ Return/Tab/Esc.
Chats page in nav rail — active chat sessions before they become runs.

Local Models

Qwen 3.6 35B MoE — latest Qwen3.6 Mixture-of-Experts on mlx-community. 35B total, 3B active per token, dynamic 4-bit. Sharper reasoning than 3.5. Needs 36GB+ RAM.
Qwen 3.5 27B Opus removed — experimental distilled variant dropped.

Decision Rooms

Advisors, runs, decision records — full terminology alignment with askmanwe.com. The output is now a decision record, not a report.
Guide rewritten from scratch — covers the Contrarian and Auditor roles, casual chat flow, reasoning effort, and every backend option. Surfaced in Settings → Learn.
Inline tooltips on the reasoning picker and the Contrarian/Auditor description.

CLI Sign-In

One-tap auth — "Sign in with Claude / Qwen / Codex" buttons in Settings open Terminal with the right auth command pre-typed. No more copying subcommands.
Retry Detection now resets auth cache too. Install the CLI, log in, tap Retry once — done.

Stability

Codex no longer hangs at agent rounds — stdout/stderr drain continuously so the pipe buffer never deadlocks.
Claude Code 2.1 auth detection fixed — the minimal environment we passed broke login detection for the new native binary.
Round-summary JSON leak fixed — when a model's response had unescaped inner quotes, the full JSON blob used to get dumped as the summary. Now the summary value is recovered or we fall back cleanly.

Quality of Life

⌘↵ Quick Run, ⌘⇧↵ Pro Run shortcuts on home. Plain Enter still starts chat.
Focus states on every main input — accent-tinted border when typing.
Jump-to-latest button appears when scrolled up in chat.
Onboarding copy refreshed.

Download the DMG below and drag Manwe to Applications. Existing users will receive this update automatically via Sparkle.

Manwe 0.6.5 · Your Data, Their Debate

Thu, 16 Apr 2026 12:00:00 GMT

Upload your documents. Every advisor reads them. The Auditor fact-checks claims against your actual data — not just web search.

Seed Data

Upload documents (PDF, TXT, MD) in the scenario editor. Your data flows through every stage of the pipeline — agent generation, debate rounds, report, predictions, and future paths.
Your facts override web research. If your document says "CPF is irrelevant to my situation," no advisor will recommend checking CPF records. The prompt system enforces it.
The Auditor fact-checks against your data. When an agent cites a number, the Auditor cross-references it against what you provided — not just what the web returned.

Multi-Backend

Codex CLI — run OpenAI models via Codex. Auto-detected, verified in Settings.
API keys — connect directly to Anthropic, OpenAI, or Alibaba with your own API key. No CLI needed. Keychain-secured.
Custom providers — add OpenRouter, Groq, Together, Fireworks, Ollama, LM Studio, or any OpenAI-compatible endpoint. One tap to add, models fetched live from the provider.
CLI auth checks — Claude and Qwen show "Not logged in" in Settings instead of silently failing. Unauthenticated CLIs hidden from the model picker.

Quality of Life

Onboarding — scrollable model list, skip button for users who want CLI or API instead of downloading a local model.
No-model guard — clear message when no model is configured instead of a cryptic error.
FlowState fixes — resolved Accelerate framework deprecation warnings and pointer lifetime issues.

Download the DMG below and drag Manwe to Applications. Existing users will receive this update automatically via Sparkle.

Manwe 0.6.0 · The Contrarian

Tue, 14 Apr 2026 12:00:00 GMT

The Contrarian rework changes how Manwe reasons. A/B tested on the "car wash test" — a logic puzzle every major LLM fails — Manwe now gets it right.

Reasoning Engine

Contrarian rework — method-based frame-checker that traces proposed actions to physical outcomes. No domain persona, no hallucinated expertise. Conviction starts at 50%, holds ground if never logically refuted.
Conviction trajectory — evidence sorted by which positions strengthened during debate. Verdict sees "Strongest: X (50→72%), Weakest: Y (70→35%)" before writing.
Overall confidence now reflects domain expert consensus, excluding synthetic agents.

Multi-Language

Debate, summaries, and reports respond in the language you type. No configuration — detected automatically.
Supported on all cloud models and Qwen 9B+. Qwen 8B stays English for reliability.

Qwen 3.6 Cloud

New cloud backend via Qwen Code CLI. First non-Claude model to hit 10/10 in report analysis.

Agent Naming

Deterministic names from 9 culture-tagged pools (English, East Asian, South Asian, African, Latin American, Middle Eastern, European, Turkish, Eastern European). Per-panel uniqueness enforced.

Sources

Used and Unused sources shown separately in app, HTML, and Markdown exports. Relevance-ranked. Retroactive.

Reliability

Failed sims keep session for retry. Queue toast. Orphaned session auto-redirect. Topic voice in summaries. Temporal context in surprise events.

Download the DMG below and drag Manwe to Applications. Existing users will receive this update automatically via Sparkle.

Manwe 0.5.5 · Relevance Engine

Sun, 12 Apr 2026 12:00:00 GMT

2,300+ lines across 41 files. The biggest update since launch.

Relevance Engine

BM25 + Entity + Embedding scoring — New three-layer relevance system replaces pure embedding filtering. Sources must pass BM25 keyword match and entity overlap before agents see them. Sharper debates, less noise.
Source relevance filter — Embedding similarity drops irrelevant sources from report footers
Sources modal: Unused tab — See what got filtered out and why. Cached embeddings fix CPU spike on large source sets.

Prompt Architecture (3-phase rewrite)

Phase 1 — Targeted prompt quality fixes, empty response retry logic
Phase 2 — XML-tagged prompt structure, conviction mapping per agent, contrarian identity system, CLA depth fix
Phase 3 — Agent identity pipeline: pool history, past debate context, and specialization flow into every prompt
Anti-repetition system — Agents no longer echo each other across rounds
Three-way topic voice — Reports adapt tone: personal ("Should I…"), clinical (health), or general
Temporal awareness — All report prompts handle dates correctly, no more past events framed as future

Redesigned UI

Homepage redesign — New layout with themed category colors and PP Editorial New font
Resizable nav rail — Drag to resize, collapses gracefully. Right-click context menu on sessions. Hover states across the entire app
Responsive layout — Overlay sidebars on narrow windows, adaptive homepage padding, title wrapping
Live simulation UX — Pipeline progress bars, agent thinking indicators in round cards, progressive report rendering as it's written
Research panel search — Filter events by content and detail text
UI polish — Banner padding, rounded sidebar corners, research panel hover effects, frosted button refinements

Smarter Agent Casting

BM25 pre-filter on pool matching — Prevents cross-specialty casting
Softened pool matching — BM25 checks agent spec text, not raw topic string
FactCheck query extraction — Foundation Models first on Quick Run, LLM fallback on Pro/Claude

New Models

Qwen3.5 27B — Dense Ultra tier for 36GB+ Macs
Qwen3.5 27B Opus (distilled)

Fixes

Medical framing no longer leaks into non-clinical verdicts
Model switching in queue captures correct model at queue time
Wikipedia namespace handling fixed in pipeline
Verdict previews in history/nav rail stripped of markdown artifacts
Agent interview chat trims leading whitespace
MLX memory limit lowered 95% → 90% for safer queued simulations
History shows verdict instead of prediction label

Manwe 0.5.0 · Living Ensemble

Wed, 08 Apr 2026 12:00:00 GMT

Living Ensemble

Your advisors are alive now. They remember, grow, and change through debate.

What's New

Living Agent Pool

Agents develop worldviews across four dimensions — epistemological lens, temporal orientation, agency belief, and optimism
Maturity tiers: Fresh → Seasoned → Veteran → Transformed
Agents accumulate cognitive dissonance when challenged — cross a threshold and they genuinely transform
24 built-in agents with pre-assigned worldviews

Talk to Your Advisors

Chat with any agent directly from their card — they remember every debate
Participation history tracks conviction arcs, dominant actions, rivals, and myth-layer reflections
Memory is on-demand: agents recall what's relevant, not everything

Agents Live Between Simulations

Every simulation saves its research to a shared knowledge store — real articles, papers, and findings from Semantic Scholar, PubMed, CORE, and web sources
After each sim, the top 2-3 agents most relevant to the new knowledge receive personal life events based on real-world findings in their domain — a toxicologist reads about a new drug interaction study, an economist reacts to fresh labor statistics
A continuous background loop keeps evolving agents over time — leave the app open and your advisors develop new experiences from accumulated knowledge across all your past simulations
Idle updates run on Apple's Foundation Models (Neural Engine) — no GPU wake, no cloud API cost, completely silent

Smarter Panel Casting

Three-vector diversity model: different stakes, different ways of knowing, different relationships with time
Category-filtered pool matching — agents are only reused within their domain (no more cross-category noise)
Self-only similarity check — agents can naturally converge without being penalized for agreement

Research Pipeline

Agents read full research abstracts from Semantic Scholar, CORE, PubMed, and ClinicalTrials — not truncated snippets
Per-agent sentence ranking using NL embeddings — each advisor gets findings most relevant to their expertise
All search prompts produce targeted research queries (8-15 words) instead of bare keywords
Claim-triggered verification: when agents cite statistics, the system auto-searches and regenerates with real evidence
Circuit breaker pattern for rate-limited APIs — try once, disable on failure, no more mid-sim timeouts

Foundation Models Integration

Knowledge extraction and context classification run on Apple's Neural Engine via @Generable structured output
Unified routing: Quick Run → Foundation Models, Pro Run → local LLM, Claude → always Claude
Automatic fallback to LLM if Foundation Models hits a safety filter or is unavailable
Idle agent updates run entirely on Neural Engine

FlowState

Extracts numeric time series from user-provided seed data (patient vitals, financial projections, market data)
Short series (3-7 points) get trend descriptions; long series (8+) get full statistical forecasts
User-provided data now reaches agents as quantitative context alongside external API data

Reports

Overall confidence computed from median agent conviction — no more hardcoded 50%
Model-specific prompt modifiers for anti-hallucination guardrails
Stated facts extracted from user input to prevent agents from inventing details
Auditor challenges assumptions about the user that weren't explicitly stated

Performance

4-bit KV cache quantization via GenerateParameters.kvBits
Removed 20+ LLM search-decision calls per sim (~150s saved)
Per-agent briefing ranked once and cached — no recomputation across rounds
Briefing depth scales by run type: Quick Run 10 sources, Pro Run 20

Models

Qwen3 8B — Standard (16GB+)
Qwen3.5 9B — Pro (24GB+)
Qwen3.5 35B MoE (A3B) — Ultra (36GB+), 35B quality at 3B inference speed
Claude Sonnet / Opus — Cloud

Manwe 0.2.6 · Patch

Mon, 06 Apr 2026 12:00:00 GMT

Patch release:

Auditor now moderates debate relevance (redirects off-topic drift)
Future paths salvaged from truncated JSON instead of discarding
Fix raw JSON summary display on Qwen 9B (thinking tag stripping)

Download the DMG below. Existing users: update via the app.

Manwe 0.2.5 · The Deeper Story

Sun, 05 Apr 2026 12:00:00 GMT

The Deeper Story

After the debate ends, Manwe now synthesizes a meta-narrative — the single story underneath all the stories. It names the invisible drama driving your decision.

CLA Overhaul

Complete rewrite of Causal Layered Analysis layers, validated by Professor Sohail Inayatullah's digital twin (UNESCO Chair in Futures Studies, creator of CLA).

Sensory-first myth generation — each agent anchors to a concrete detail
Independent myth generation — no groupthink, like writing on paper before sharing
Worldview link-back — agents connect assumptions to their earlier advice
Adaptive CLA gating — fires for decisions, skips for diagnostics unless debate stalls

New Intelligence Sources

FlowState quantitative forecasting engine (IBM, 9.1M params, pure Swift/Accelerate)
Polymarket prediction market intelligence — smart money positions

Quality of Life

Markdown rendering in all reports (bold, italic)
Model name shown on reports (Qwen3 8B, Claude Sonnet, etc.)
Divergence ratio tracking per round (data collection for future features)

Download the DMG below. Existing users: update via the app (Check for Updates in Settings).

Manwe 0.2.0 · Future Paths & CLA

Wed, 01 Apr 2026 12:00:00 GMT

What's new

Future Paths — after every simulation, Manwe generates 2-3 divergent timelines showing what happens if you take different forks. Each path is a month-by-month causal chain grounded in evidence from the debate. "Month 3: You sign a lease in Mission District, burning $38K/month. Month 12: You discover enterprise PMF with 6 months runway left." Then you can talk to your "future self" from each path — ask them what surprised them, what they'd change, what happened next. The paths are built from debate evidence, not your stated preferences, so they can challenge your assumptions.

Deeper debates — debates now descend through layers of depth instead of circling at the same level. Early rounds hash out arguments. Later rounds push agents to surface the deeper beliefs driving their positions, then synthesize the whole situation into something more meaningful. Inspired by Sohail Inayatullah's Causal Layered Analysis framework. The final rounds produce genuinely different content: "Gold is the ancient hearth, Bitcoin is the flickering match" (Qwen 9B), or five distinct perspectives on LASIK from five agents: a mortgage, a one-way flight, a tattoo artist, a face tattoo at 25, and selling your childhood home (Claude Sonnet).

Pre-sim clarification — the chat refinement flow now asks structured foresight questions before assembling the panel. Instead of improvising, Manwe probes deeper to understand your assumptions, concerns, and what a good outcome looks like — so the debate is grounded in your situation rather than generic.

Apple-style sidebar — replaced the icon rail with a proper sidebar. Navigation items at top (Ask Manwe, History, Agents, Settings) with line/filled icon states. Recent sessions grouped by time (Today, Yesterday, Last 7 Days) showing topic + truncated verdict snippet. Feature highlight banner and Check for Updates at the bottom. Sidebar hides when viewing a simulation for full-screen content.

Time estimates on Pro Run — the Run Simulation button now shows estimated duration based on your model, round count, and hardware. Updates live as you drag the round slider. No more guessing whether to wait or make coffee.

Cancel simulation fix — cancelling a running simulation now returns to the homepage cleanly. Previously could leave you stuck on a "Session not found" screen with a dead Back button.

Improved classification — decision classification is now more reliable across all models. Questions like "Should I get LASIK?" are correctly identified as medical decisions and get appropriate specialist panels.

Longer outputs — increased generation limits across the board for improved output quality, especially on smaller models. Reports and analysis are more thorough with less truncation.

UI polish — round cards are fully tappable (not just the tiny chevron), history category tabs have proper hit targets, Pro Run back button shows "Back" text, Settings page layout matches Agent Pool width and header style, sidebar toggle buttons hidden on chat and configure pages where they serve no purpose.

PDF export — disclaimer text contrast improved for readability on white backgrounds and printed copies.

Manwe 0.1.5 · Smarter local models, bug fixes, memory management

Mon, 30 Mar 2026 12:00:00 GMT

What's new

Grounding search for local models — Qwen models now do a quick web search (~1 second) before the debate starts to understand ambiguous topics. Previously, "Is it worth paying for X Premium?" would generate generic SaaS advice; now agents correctly understand it's Twitter's paid subscription and debate accordingly.
Model memory unload — models are freed from GPU memory when the simulation queue is empty. Previously held ~5-6 GB of RAM indefinitely after completing. On a 24GB machine, that's 70% of RAM freed.
AI disclaimer — "This report was generated by AI. AI can make mistakes. This is not financial, legal, or medical advice." now appears in all reports and exports.
Auto-scroll fix — scrolling up during a live simulation no longer glitches back to the bottom. Auto-scroll pauses when you scroll up, resumes when you return to the bottom.
Empty predictions handled — reports without predictions show a subtle explanation instead of an empty section label.
Predictions label consistency — fallback default changed from "Things to Consider" to "Predictions."
State contamination fix — viewing a completed simulation while another is running no longer overwrites the running simulation's data.
History context menu — running simulations show "Cancel Simulation" (with confirmation), completed simulations show "Delete" (with confirmation). Previously a single "Delete" button could silently kill a running sim.
Model selection persistence — selecting a model on the homepage or scenario editor now persists across app restarts. Previously reset to default on relaunch.
Delete running sim safety — deleting a running simulation from history now properly cancels the engine first.

Manwe 0.1.4 · Adaptive speed, End Debate, 13 data sources

Sun, 29 Mar 2026 12:00:00 GMT

What's new

Adaptive simulation speed — Quick Run rounds scale with your hardware
"End Debate" button — gracefully wrap up a simulation anytime after the early rounds. Agents give final positions and the report generates.
Cancel confirmation — destructive cancel now requires confirmation to prevent accidental data loss
DOAJ data source — 13th source added: 12.5M peer-reviewed open access articles covering humanities, social sciences, law, and education
Improved API resilience — better handling of failed data source calls, saves time when sources are temporarily down
Faster simulations — optimized late-round processing for shorter wait times
Pro Run rebalanced — tuned default round count for better quality-to-time ratio
Events in reports — surprise, planned, and injected events now visible in completed reports and all exports (HTML, PDF, Markdown, Rich Text)
Updated AI engine — mlx-swift-lm, mlx-swift, swift-transformers updated to latest versions with native Qwen3.5 support and performance optimizations

Manwe 0.1.3 · Report Redesign & Export

Fri, 27 Mar 2026 12:00:00 GMT

What's New

Redesigned report layout — 3-layer read: verdict card → predictions → action plan → evidence → risks → debate rounds
Export formats — Save as PDF, Save as Web Page (HTML), Copy as Rich Text, Copy as Markdown
Category-themed verdict cards — medical (red), finance (orange/gold), tech (cyan), lifestyle (teal), and more
Brand color — new Manwe indigo #4636FC replaces generic AI purple
Save Report button — full-width glass button with category color, right in the verdict card
Branding on exports — Manwe header + "Powered by Manwe" CTA footer on all formats
Smart filenames — exports named from topic + date (e.g. should-i-switch-from-react-2026-03-28.html)
Interview buttons — frosted white glass for better contrast
Panel at top — agent cards first on completed simulations, interview is a key feature

Download Manwe-0.1.3.dmg