| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-09-11 | 7.6 kB | |
| v2.5.0 source code.tar.gz | 2026-09-11 | 3.6 MB | |
| v2.5.0 source code.zip | 2026-09-11 | 4.5 MB | |
| Totals: 3 Items | 8.0 MB | 8 | |
- Agent runtime: Expanded Agent mode with dynamic Tool and Skill discovery, task plans, structured user questions, scoped Memory, workspace and Artifact operations, MCP resources/prompts, persistent execution records, resumable runs, and effect-aware confirmation. Tools loaded by the host are explicitly authorized for later model rounds.
- Chat generation reliability: Stopping before a conversation switch now captures and persists the originating stream before navigation. Continuing an interrupted response retains existing document, search, image, plan, and Tool blocks while recording new structured output and Tool state.
- Deep Research: Added model-knowledge-first planning, bounded knowledge and public lookup, reviewable plans, adaptive research rounds, durable checkpoints, claim-to-evidence auditing, local citations, partial reports, manual resume, report versions, and the responsive Research workbench. Planning keeps the originally selected model through recovery, and approved workspace failures pause visibly instead of silently removing sources.
- Deep Research templates: Added four built-in research templates for competitive analysis, literature review, due diligence, and technical evaluation, plus local custom template management, revisioned Profile snapshots, explicit inheritance or no-template states, and immutable per-task template snapshots.
- Specialized read-only sources: Added arXiv, PubMed, EPO OPS, and SEC EDGAR search/read adapters with fixed official endpoints, source-scope and definition-fingerprint checks, encrypted local credentials, bounded parsing, provider-aware throttling, and discovery results that become evidence only after a source read is committed.
- Frontier steering: Added durable safe-wave commands to promote, demote, restore, or add a research node. Task Web Locks serialize execution; queued changes are applied after a completed wave and checkpoint, while late or already scheduled changes receive an explicit rejection.
- Evidence conversations: Added report-version-bound evidence snapshots and multiple local Q&A topics per report. Answers stream with cancellation and retry, use only the frozen report/evidence/claim ledger, and validate every citation against that snapshot without tools, network, Memory, or Research execution.
- Research extension storage: Added the local IndexedDB
neo-chat-research-extensionsv1 sidecar for templates, source contracts, steering records, evidence snapshots, and threads. The core Research store and existing ZIP format remain unchanged; task deletion, session copies, and orphan cleanup maintain the sidecar lifecycle. These extension records remain browser-local and are not yet included in ZIP backup or encrypted sync. - Research reliability and presentation: Research searches share a serial queue with two-second spacing, preserve partial batches, respect cooldowns, and allow 90 seconds for Tavily Research and planning summaries. Evidence recovery preserves committed excerpts and stable source identity; report headings, dates, numbers, appendices, quality details, and downloads follow the selected English, Chinese, or Japanese locale. Search images retain provenance without becoming verified evidence merely by discovery.
- Research rollback compatibility: New reconnaissance records can contain
timeoutMs: 90000; older readers capped at 30000 need the validator update before opening these records. Existing reports are not rewritten. - Encrypted synchronization: Added opt-in WebDAV and S3/MinIO synchronization using per-domain Automerge documents, browser-side HKDF/AES-GCM encryption, opaque remote names, recovery codes, conflict visibility, transactional local apply, and encrypted OPFS chunks. A local edit made during remote blob download now forces a bounded remerge instead of being overwritten by a stale result.
- Conversation controls: Added customizable page-scoped keyboard shortcuts, shared sidebar/titlebar actions, and icon-only temporary text chats that stay out of saved history. Empty conversations hide titlebar actions, and shared dialogs retain keyboard focus and expose consistent close behavior.
- Read-only sharing: Redis snapshots include the selected branch, images, and
Research reports with configurable expiry, explicit updates, and revocation
before local deletion. Sharing defaults off behind
SHARING_ENABLED; disabling publication preserves authenticated revocation while preventing public reads. - Rendering and documents: CommonMark renders immediately while syntax extensions load independently, preserving streaming fences and source placeholders. Long model output can become persistent document blocks with preview, editing, and export controls.
- Charts and visual export: Added ECharts-backed
chartMarkdown fences with a strict JSONversion: 1/renderer: "echarts"envelope, inline-only datasets, theme-aware defaults, and themarkdown-chartcompatibility alias. Charts provide a chart-only view, fullscreen rendering, localized loading/error/retry states, Markdown table copying, and PNG saving. Chart-aware image and PDF exports wait for rendering, and Mermaid diagrams and mind maps can also be saved as PNG images. - Conversation startup: Reloads now open the welcome view while preserving durable conversations, and sync rehydration keeps an explicitly selected in-memory conversation and its loaded messages instead of replacing them with stale persisted selection state.
- Mermaid mindmap presentation: Scoped the article palette to Mermaid
sources detected as
mindmap, with opaque borderless circular nodes, SVG text labels, and curved links while preserving explicit shapes, rich labels, copied source, and the existing appearance of other diagram renderers. - Mermaid diagram palettes: Corrected abnormal dark-color collapse in Git graphs, timelines, kanban boards, radar charts, and treemaps with derived light/dark palettes, readable text and labels, and stronger boundary and stroke contrast. The overrides are scoped to those diagram types, embedded in rendered SVG for fullscreen and exports, and preserve source and layout.
- Providers and media: Added direct browser requests for configured custom model providers, multi-password deployment access, HEIC/HEIF conversion, and staged image compression with file-backed native OpenAI, Google, and Anthropic multimodal uploads. Loading and shared UI presentation use consistent states.
- Web reading reliability: Jina failures can fall back once to a bounded direct public-page read. Challenge pages and application errors are rejected before evidence creation, and asynchronous plugin failures use route-level error handling.
- Deployment and release safety: Standardized local, CI, Docker, and Worker builds on Node 24. Docker quick starts require an access password and verify API availability. Hosted HTML receives per-request CSP nonces and bounded WebAssembly permission. Compatible framework, provider, sanitizer, validator, and MCP dependency updates clear the production audit; tag releases validate the package version and pass source, Next.js, Worker, and dependency checks before publication.
- Dependency security: Updated vulnerable root and MCP Bridge dependency resolutions for Sharp, JS-YAML, Vitest, baseline-browser-mapping, brace-expansion, Hono, qs, fast-uri, ip-address, and the Hono Node adapter.
