A features-and-polish release. Image input now works for vision-capable models with your own provider keys, plus a long list of reliability fixes responding to user feedback since 6.0.

* Image input. The paperclip and macOS drag-drop now accept images, and vision-capable models (Claude, GPT-5, Gemini etc) receive them through your configured BYOK provider keys. The icon greys out with an instant tooltip when no vision-capable model is in your active context.
* Drag and drop also handles documents. Drop a PDF, text file, Markdown, or DOCX onto the message area in any of the three chat windows (Council, Direct Chat, Compare) and it attaches via the same pipeline as the paperclip, instead of inserting the file path as text.
* Image attachments render as thumbnails in conversation history. Click to open in Preview. Image bytes now persist to disk so thumbnails survive a restart.
* Document attachment chips persist in conversation history across all three chat areas now, not only Council.
* Context-size banner above the composer turns amber at 80% of your character cap and red at 100%, with a “New chat” shortcut from the Council view so you can split a long thread before it overflows.
* Input character caps raised significantly: messages up to 500K characters (was 50K), default total cap doubled to 200K, optional cap up to 1M for power users loading multiple long files.
* The Chair (Synthesiser) picker is now strict: it lists only models that are actually in your current council, never persists an empty selection, and excludes Apple Intelligence (which contributes at Stage 1 only). Health-probe at launch skips Apple’s on-device model so it no longer logs spurious errors.
* Model routing accuracy: false-positive “model unavailable” reports through OpenRouter have been fixed, and provider matching is now exact instead of prefix-based.
* iCloud sync: a CloudKit semaphore that could deadlock on slow connections has been removed; sync runs more reliably and never blocks the main thread.
* Direct Chat: a double-tap race that could send the same message twice in rapid succession has been fixed.
* Accessibility labels added to icon-only controls across the chat surface.
* Privacy and security: privacy manifest updated for disk-space API access, debug logging stripped from Release builds, and API keys no longer appear in URL query strings.

If anything is not quite right, tap Feedback inside the app.

A headline features release with reliability polish throughout.

- Web Search arrives across eight providers (Tavily, Brave, Exa, You.com, Kagi, SerpAPI, Perplexity, Google) with source citations presented as clickable chips below each response. Auto mode tries your configured backends in order and falls through cleanly on failure.
- Apple Intelligence now joins the Council on Apple silicon Macs running the latest macOS. It runs entirely on device, needs no API key, and can take part in multi-model councils alongside your cloud providers.
- Chair synthesis streams live tokens as the chair thinks, and automatically retries on transient network errors. Long sessions no longer hang on a single flaky response.
- Per-provider error banners: "Rate Limited (anthropic)" rather than a generic "Rate Limited", so it is instantly clear which provider needs attention.
- The Web Search settings card has been redesigned with clearer status, per-row Show / Hide toggles on API keys, and a three-state indicator so iCloud-synced keys no longer appear as un-configured.
- Dozens of small refinements across Prompt Library, Compare, Direct Chat, and model analytics.
- Apple Intelligence shows a blue "Local" dot in the Council list, with no spurious "Get Key" prompt. Routing and validation now treat it as a first-class on-device provider.
- The Chair (Synthesiser) dropdown is now strict: it lists only models that are actually in your council. Any stale chair from an earlier configuration is reset automatically when you open the Council tab.
- Hidden local models no longer appear in the Add Model picker. The hidden-models filter now applies consistently across every picker that sources models from Ollama or linked GGUF files.
- Historical conversations now explain when a referenced model is no longer available: a per-response footnote plus a banner at the top of the conversation listing each unavailable model with a reason (hidden by you, provider key missing, or no longer in the catalogue).
- Comparison model picker no longer pegs the CPU when scrolling through a long list.

If anything is not quite right, tap Feedback inside the app.

•	Switching between conversations now reliably shows the right content. A bug where a recently completed conversation's synthesis could appear inside a different conversation has been fixed.
	•	The chair retry button ("Generate Chair's Report") now surfaces clear error messages when the chairman model fails, instead of appearing unresponsive.
	•	Network requests fail faster when a provider is unreachable, so one slow model no longer stalls the full council for minutes.

Local models, plus reliability fixes:

GGUF files and Ollama models now stay in your council reliably.

The “Model Removed” toast that used to appear seconds after adding a local model has been removed.

Local models now appear in every picker, including Chair, Direct Chat, and Compare.

Files stored outside the app’s folder such as ~/.ollama/models remain accessible across restarts, and once a folder is scanned the app remembers it.

Each model now uses its own native prompt format including Gemma, Llama 3, Qwen, and Mistral, so responses are coherent and no longer leak raw template tokens into the output.

A crash when sending a message to a local model has been fixed, and any remaining stray template markers are scrubbed from the output as a safety net.

The Local Models settings tab has been clarified with two distinct sections, “Link Existing Models” for scanning folders and “Downloads Folder” for where new downloads are saved.

Cancelling a streaming response now closes the upstream request rather than letting it finish silently in the background. Google Gemini system prompts now work correctly after fixing an invalid API field being sent.

Transient 503 errors during chairman synthesis now retry once, so a single provider blip no longer kills a long deliberation.

The top toolbar is now fully opaque, so scrolling messages no longer bleed through behind the “New Conversation” button and model selector.

Thanks for using AI Council. If you spot anything odd, use the in-app Feedback button and we will take a look.

Bug fixes and polish:
- Council responses no longer blank out during peer review or synthesis. The redesigned renderer keeps every response visible from start to finish, even on long sessions.
- Compare now has a proper welcome screen showing which models you have picked, matching the style of the Council and Simple Chat welcomes.
- The model picker sheet has a cleaner single-row layout with clear Cancel and Done buttons.
- Error banners now include the specific model that failed, so you know straight away whether it is a rate limit, a missing key, or a stale model name.
- Local models (Ollama, LM Studio) are handled more reliably when you add them to your council.
- Faster, lighter markdown rendering in every chat surface, reducing memory use.

If anything is not quite right, tap Feedback inside the app to let us know.

Auto Model Migration - when providers release newer versions, stale model selections are detected and migrated automatically across council members, chairman, direct chat, and comparison. A notification explains what changed.
Session expiry notifications replace silent failures.
Network timeouts prevent hangs.
Bounded notification caches.
Comfort settings - new Appearance controls let you adjust colour warmth and contrast to reduce eye strain.
Background completion badges - conversations that finish processing while you're elsewhere now show a notification badge.
Non-chat model filtering - audio, search, and other non-chat model variants no longer appear in model pickers.
HTML title decoding - conversation titles with special characters now display correctly.
Dynamic model pricing - pricing data is fetched live from providers rather than using static values.
Bug fixes and stability improvements.

Managed Plans
Use AI Council without bringing your own API keys. Three tiers: Starter (150 credits/month) for budget models, Plus (500) for mid-tier, and Pro (1,400) for every model including GPT-5.2, Claude Opus 4.6, and Grok 4. Already have your own keys? BYOK always takes priority.

Credit Top-Up
Purchase 200 bonus credits any time. They never expire and carry over between months.

Auto Model Migration
When providers release newer versions, stale model selections are detected and migrated automatically across council members, chairman, direct chat, and comparison. A notification explains what changed.

Reliability
Session expiry notifications replace silent failures. Credit balances refresh on return from background. Network timeouts prevent hangs. Comprehensive audit of 50 issues resolved. Bounded notification caches. Fixed stale subscription data after plan expiry.

Five major new features:

Direct Chat - Talk one-on-one with any model from your configured providers. Quick answers without the full council process. Pick your model, start chatting.

Smart Escalation - Started a direct chat but realise the question needs more perspectives? Escalate to the full council with one tap. Your conversation context carries over.

Comparison View - See 2-3 models respond side by side or stacked on iPhone. Compare answers in real time to see where models agree and where they differ.

Prompt Library - Save your best prompts and reuse them. Organise by category, mark favourites, search, sort by most-used. Edit and refine over time.

Document Workspace - Attach reference documents PDF, DOCX, TXT, Markdown to any conversation. Models draw on your documents when answering, with relevance scoring and configurable context injection.

Other improvements:
- Conversations now sort by last activity, not creation date
- Prompt Enhancer loads models from all configured providers, not just council members
- Comprehensive help sections for every feature
- Refined typography and tighter list spacing across both platforms
- Animated sidebar reordering when conversations update

Stability and performance:
- Full codebase audit: 80 issues identified and resolved
- Consolidated LLM service layer for more reliable model routing
- Fixed iCloud sync not running on app startup
- Fixed streaming response race conditions
- Improved PDF and document extraction performance
- Reduced memory usage during model downloads
- Fixed numerous layout and overflow issues

New: Prompt Enhancer - Enhanced

The Prompt Enhancer has been completely rebuilt as a stepped wizard. Instead of presenting all clarification questions at once, it now asks one focused question at a time with a clear progress indicator. Choose which AI model drives the enhancement, and optionally target a specific model so the output is optimised for its prompting style. The enhancer draws on over 10 prompt engineering frameworks - including RISEN, RADAR, Chain-of-Thought, and ReAct, to select the best approach for your prompt. Access it instantly from the new wand button in the message input bar.

If something goes wrong during enhancement, you can now retry with the same model or switch to a different one without losing your progress.

 Bug Fixes
 - Fixed Anthropic API queries failing due to incorrect system message handling
 - Fixed Together AI API misconfiguration

Bug Fixes:
Fixed an issue where some providers could not be used even when a valid API key was entered. API keys  were saved correctly but the app failed to recognise these providers as configured, preventing model discovery and query execution. All 20 built-in providers now work as expected.

iCloud Sync
Your conversations, projects, and settings now sync seamlessly across all your Mac and iOS devices. All data is encrypted end-to-end in your private iCloud - we never have access to your conversations.

Bug Fixes
  - Fixed council configuration not syncing between devices
  - Fixed accent colour changes not applying until app restart
  - Many more under the hood fixes

New Features
  - Splash screen with animated logo on app launch
  - Improved conversation title generation with better keyword extraction

  Bug Fixes
  - Fixed light/dark theme not being applied correctly across all windows
  - Fixed setup guide not updating to show completion status when API keys are configured
  - Fixed "Documentation" button text wrapping in provider settings          
  - Several layout fixes, API and bug fixes

  Improvements
  - Updated app icon
  - Refined UI throughout
  - Better visual consistency across all windows and dialogs