Skip to content

Complete feature index ​

中文 · Guide map

This is a scan of current product areas. The feature guides explain how to use them; Feature status separates current behavior from experimental directions.

Canvas and navigation ​

  • Multiple canvases with create, rename, switch, archive, and delete flows.
  • Infinite canvas: the wheel or a pinch zooms, middle/right-drag pans, nodes move, a drag box selects. On a trackpad, ⋯ → Wheel pans the canvas makes a two-finger scroll pan instead (zoom by pinch or ⌘/Ctrl + wheel).
  • Three semantic zoom levels: full cards, takeaway plaques, and icon-level markers.
  • Minimap, zoom controls, frame navigator, node search, structural arrow-key navigation, and root-path highlighting.
  • Tidy layout, local Align, frames, annotations, undo, and redo.

Guide: Canvas, navigation, and frames

Conversation nodes and floating panel ​

  • One node stores a question, answer versions, model provenance, role, highlights, and attachments.
  • Continue a path, Explore from selected text, edit questions and answers, or regenerate in place or as a sibling node.
  • Floating panel for reading, version switching, attachments, highlights, merge order, grouped context tree, and follow-up input.
  • Right-click node menu for panel/reading access, regeneration, summarizing to a note, copying, CLI continuation, duplication, archiving, and deletion.

Guide: Conversation nodes and panel

Wires and model-visible context ​

  • Purple solid conversation paths, orange structural exploration branches, purple dashed summary references, and red dashed reviewer relationships.
  • Will send preview grouped into materials, references, and conversation turns.
  • Delete or convert a wire without deleting its nodes; archive content to exclude it while retaining it.
  • Context fingerprints, stale markers, answer versions, and dependency-ordered replay.
  • Controlled comparison by regenerating the same question and model after a context change.

Guide: Control context · Versions, staleness, and replay

Material nodes and reader ​

  • PDF, image, HTML, DOCX, Markdown, text, code, and link-snapshot materials within current input boundaries.
  • Original/text/digest views where available; document recognition and editable extracted text.
  • Ask from selected text or a selected visual region, save a passage or image as its own node, and preserve page provenance.
  • Reading rail for source-grounded question chains, with jumps between source, answer, and canvas node.
  • Material overview and inherited attachment controls.

Guide: Material nodes and reader

Organize and review ​

  • Merge summary, Merge & delete, Weave highlights, structured highlights, and multiple downstream highlight modes.
  • Condensed copy of consecutive unbranched runs while preserving the original path and protected decisions/pivots.
  • Sliding reviewer nodes, graph diagnostics, and manual or automatic refresh of dependent work.

Guide: Merge, highlight, and condense

Models, tools, roles, and memory ​

  • Runtime provider presets, OpenRouter sign-in, Ollama, custom OpenAI-compatible endpoints, .env providers, and supported plan connections.
  • Global model selection and node-level model pinning; model provenance stored per answer version.
  • Web search, scholarly search, vision routing, and local/remote MCP tools.
  • Inherited/set/reset node roles and editable role library.
  • Memory as documents:
    • A preferences and an identity document rewritten as facts arrive (with changelog and undo).
    • One dossier per topic written from the labelled conversations (what it is, decisions, where it stands, still open; every line sourced).
    • An inbox for project facts no topic claimed.
    • The other agents' memory files as read-only sources.
  • Recall on an ask:
    • The question's topics bring their dossiers in whole.
    • A detail question adds verbatim excerpts ranked by relevance.
    • Every item listed with its cost, removable.
    • Three amounts: lean, standard, generous.
  • Related conversations per node: the node's terms and topics as chips, model-proposed expansion, cards with open and cite.
  • Decision model (fast thinking): one switch, on by default. (Jev-class models implement the System One contract.)
    • A Jev-class System One endpoint through an aggregator like OpenRouter, the official API, Cloudflare Workers AI, a self-hosted server, or the chat model as an uncalibrated stand-in.
    • A timeout trips a one-minute breaker; every decision falls back to a rule.
  • Multi-model branch scores:
    • Once a multi-model comparison has answered, the decision model marks each branch with whether it answers, how well, and which is best, as a hint that enters no context.
    • Any answer can be answered again with another model as a new version, agent turns excepted.
  • Topics: a table the person names or the model proposes; labelling by the decision model in the host, in the background; weft topics on the command line.

Guide: Memory · Models, tools, and memory · Connect a model · Decision model

Session Atlas ​

  • Discover supported local Codex, Claude Code, DeepSeek Harness, and Pi sessions by project folder.
  • Read-only graph mirrors with conversation nodes and inspectable tool-result attachments.
  • Incremental append, changed-session filters, multi-session canvas mounting, re-mirroring, and source-status indicators.
  • Sources management and optional one-command handoff from supported agents.

Guide: Session Atlas

Toolbar, menus, and data ​

  • Toolbar access to models, replay, frames, search, language, backup, overflow actions, undo, and redo.
  • Overflow menu for annotation visibility, layout, highlight/material overviews, condense, diagnostics, sharing, exports, appearance, memory, and tutorial.
  • IndexedDB persistence, automatic folder backup, complete JSON import/export, selected/context-chain Markdown, event CSV, and read-only URL sharing.

Guide: Interface overview · Data, backup, and sharing

Weft is open source under the MIT License.