Skip to content

CHANGELOG.md

Latest commit

 

History

History
164 lines (115 loc) · 7.07 KB

CHANGELOG.md

File metadata and controls

164 lines (115 loc) · 7.07 KB
Raw

Changelog

All notable changes to Dialogram are documented here. Format based on Keep a Changelog; versions are the Dialogram platform API version.

[Unreleased]

[0.7.0] - 2026-09-01

Edge routing, and boundary ports as schematic symbols. (#6-#23)

The exported API surface is unchanged, but the minor bump is deliberate: it marks a release that changes how diagrams are routed and drawn. Compatibility is exact major.minor while pre-1.0, so consumers must raise their expected version to 0.7.0 — EXPECTED_DIALOGRAM_API_VERSION in each shell's extension/main.ts — or the platform will refuse to load for them.

Added

  • Orthogonal edge routing through libavoid, opt-in with WORKFLOW_DIAGRAM_ROUTER=libavoid. On the reference network it cuts crossings 285 -> 239, node overlaps 43 -> 1 and bends 166 -> 94 against the built-in Manhattan router. The dependency is vendored as an eval-free build so the webview needs wasm-unsafe-eval rather than unsafe-eval. (#17)
  • Edges route live in the webview while a node is dragged, using the same router and the same port anchors the server uses on commit — so the committed route replaces the live one without a visible jump. (#18)
  • Feedback edges — the connections that close a loop — are drawn and highlighted as such. (#10)
  • A chat tool for asking which editors can open a file. (#13)
  • The property-panel chrome, field toolkit and CSS are exported for library consumers, parameterized by config rather than hard-coded ids. (#6)
  • docs/proposals/structured-port-types.md: what the graph schema and operation vocabulary need before a boundary port's type can be edited, a product or variant type shown, or a type's definition navigated to. (#22)

Changed

  • Boundary ports draw as schematic port symbols instead of rounded pills: an arrow glyph on the wire's own axis, the name on that line, the type on a second line below. The wire used to meet the pill at its vertical centre, which is the gap between the two lines, so it ran through the middle of the label. Rows are a little over half the height of the old boxes. (#23)
  • Nodes and boundary ports paint over edges rather than under them. The same order decides hit-testing, so a wire passing behind a node no longer takes a click meant for the node. (#23)
  • Property sections that are empty — annotations, parameters, port lists — open collapsed; the header already carries the count. A boundary port no longer offers "Input Ports" and "Output Ports" at all, being itself a port. (#21)
  • The compact layout releases boundary ports from their first/last columns, so ELK can place them beside whatever they connect to. (#19)
  • A creation that touches two files is one undo step. (#14)

Fixed

  • Edges no longer render detached from their ports. Persisted edge routes are re-anchored to the current port anchors on load; the snap existed but had never run, because ports were resolved through a GModelRoot.index property that does not exist.
  • Node sizes are server-authoritative on every path. A fresh open no longer commits the client's measured bounds — which include the port and type-footer overhang — nor re-anchors ports inside the committed width, so fresh-open and persisted-reopen geometry now agree.
  • Diagrams no longer twitch on reload. ELK re-centres FIXED_POS ports on a whole-pixel routing lane, returning ports of odd height half a pixel low; its port geometry is no longer written into the model, and edge endpoints are snapped to the server anchors instead.
  • A double-click no longer nudges the node it lands on. Any pointer movement while the button was down counted as a drag, so the tremor in a double-click committed a real move and rerouted the node's edges. Mouse only — keyboard nudges of any size still commit. (#19)
  • Changing a boundary port's type no longer renames the port. Every label edit was routed to a rename of the nearest entity, which for a type label resolved to the port itself, so the port was renamed to whatever type was typed. (#22)
  • Edges leave a port horizontally rather than on a diagonal, and several edges sharing one port fan out instead of turning at the same point. (#19, #23)
  • A node's height accounts for its type label, so the label no longer overlaps what sits below. (#8)
  • Feedback edges are styled on the element the class actually lands on, the walk starts where data enters the network, and the network is ordered the way a layout engine expects. (#11, #12)
  • Every child process settles, so the loading notice always clears. (#15)
  • The build type-checks what consumers actually compile. (#16)
  • The file attached to a chat session is bound to that session. (#9)

Security

  • Resolved npm audit findings. (#7)

[0.6.0] - 2026-07-29

MCP cutover: one agent surface. (#5, #6)

Added

  • create-task-type as a reversible GLSP operation and custom MCP tool (undoable).
  • create-edges accepts nodeName.portName addressing.
  • Auto-layout after agent structural edits.
  • Shared, parameterized property-panel chrome, field toolkit and CSS for library consumers. (#6)

Changed

  • Agents now get exactly two servers: the in-host GLSP-MCP loopback (undoable diagram editing) and the in-process registry (read tools).
  • Agent-dispatched node creation is headless (no dialog), with actionable errors and port-rich confirmations.
  • Chat UX: named tool chips, scroll respected during streaming, clean new-session transcript, explicit empty-state picker, unique session names, model-loading notice always clears.

Removed

  • Legacy stdio MCP server and its plumbing.

Fixed

  • MCP label-edit renames now apply (were a silent no-op).

[0.5.0] - 2026-07-28

GLSP-MCP adoption (parallel-run). (#4)

Added

  • Loopback HTTP GLSP-MCP server per opted-in diagram profile, advertised to agents alongside the legacy stdio server.
  • Agent diagram edits ride GLSP operation handlers as reversible workspace edits — user-undoable like palette edits.
  • Opt-in via profile.mcp.enabled; per-user rollback dialogram.chat.useGlspMcp (legacy workflow.chat.useGlspMcp honored).

Changed

  • Undo/redo persists the document so the diagram reloads immediately.

Removed

  • Server-stack undo/redo MCP tools (the host owns undo).

Fixed

  • Reversible-edit snapshots read authoritative disk content (no false undo refusals).
  • Undo on an empty stack no longer shows an error toast.

[GLSP 2.7.0 + Node 22] - 2026-07-28

Toolchain upgrade, no intended behavior change. (#3)

Changed

  • Exact GLSP 2.7.0 pins on a Node 22 toolchain.

Fixed

  • Diagram no longer force-reloads on build-artifact (.py) changes; own-source edits still reload.
  • Enrichment/queue-visibility redeliveries take an agent-context-only fast path (less drag-time work).

[0.4.0] - 2026-07-27

One chat panel everywhere. (#2)

Changed

  • Single diagram-profile chat panel across products, with richer chat configuration.

Removed

  • Chat-only API surface.

[0.3.0] - 2026-07-24

Chat runtime convergence. (#1)

Changed

  • Unified the two chat runtimes into one ChatRuntime with shared selection context.