New TabKpis component shows scope-aware metrics with trend arrows
aligned right in the content tab bar. Each metric shows current value
and an arrow indicating change vs previous period (green=good, red=bad).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Row click now navigates directly to the split view with diagram.
Removed: DetailPanel, inspect column, unused imports (ExternalLink,
ProcessorTimeline, RouteFlow, useExecutionDetail, useDiagramLayout,
buildFlowSegments).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Sidebar wrapper gets height:100% to fill window
- Route-scoped Exchanges uses same Dashboard table (not compact ExchangeList)
- 50:50 grid split: table on left, diagram on right when route selected
- ContentTabs gets border-bottom and surface background for visibility
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Thin wrapper pages that conditionally render AgentHealth/AgentInstance
and RoutesMetrics/RouteDetail based on URL params for the nav redesign.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Port alignment BEGIN on DO_TRY compounds makes edges attach at the top
instead of center, keeping the main flow level. Post-processing also
stretches all DO_TRY sections (doFinally, doCatch) to match the widest
section's width for visual consistency.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
setPointerCapture on the SVG redirected click/dblclick events away from
node <g> elements, breaking drill-down (double-click) and potentially
click selection. Now only capture the pointer when clicking on empty SVG
space, preserving normal event flow on nodes while keeping drag-to-pan.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
ELK's partitioning doesn't reliably order disconnected children within
a compound node. Instead, let ELK lay out freely then re-stack sections
in correct order (try_body → doFinally → doCatch) by adjusting Y
positions in the ELK graph before extraction. This propagates correctly
to both node and edge coordinates via getAbsoluteY().
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Layer constraints (FIRST/LAST) don't work for disconnected components
in ELK's layered algorithm. Replace with invisible edges that chain
try_body → doFinally → doCatch to guarantee correct top-to-bottom order.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The previous fix skipped ALL edges from DO_TRY nodes, which also
removed the continuation edge to the next node in the main flow
(causing LOG nodes to appear disconnected). Now checks if the target
is a descendant of the DO_TRY ELK node — only internal edges are
skipped, continuation edges to the next main flow node are kept.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
ELK TB layout places children in insertion order. Now explicitly adds
DO_FINALLY before DO_CATCH so the visual order inside DO_TRY is:
try body (top) → finally → catch blocks (bottom). Also reduces
internal spacing to keep the compound more compact.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Backend: DO_TRY compounds now use a virtual _TRY_BODY wrapper with LR
layout for the try body, while DO_CATCH/DO_FINALLY stack below as
separate sections (TB). Edges from DO_TRY are skipped like route-level
handler edges. Removes ELK-v2 debug logging.
Frontend: _TRY_BODY renders as transparent wrapper, DO_CATCH as red
tinted section, DO_FINALLY as teal section. DO_FINALLY color changed
from red to teal (completion handler, not error).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When pressing Enter in the command palette without explicitly selecting
a result (via arrow keys or mouse), the search query is now applied as
a server-side full-text filter on the Dashboard table. Explicit
selection still navigates to the exchange. Updates design system to
v0.1.18 for the new onSubmit prop.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Previously onPointerDown bailed out when the target was inside a node
(data-node-id), blocking pan entirely over nodes and compound groups.
Now panning always starts, and a didPan ref distinguishes drag from
click — node click handlers skip selection when the user was dragging.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds lucide-react and replaces all HTML entity and emoji icons across
the UI with proper SVG icon components. Tree-shaken — only imported
icons are bundled.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The Dashboard was fetching 50 results without a status filter and
filtering client-side, causing fewer matches when filtering by error
compared to route-specific pages that filter server-side. Now passes
statusFilters to the OpenSearch query. Backend supports comma-separated
status values for multi-select filters.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Root cause: graph.getNodes() is a flat list with duplicates — handler
compound children appear both nested inside their parent AND as
top-level entries. The previous separation tried to filter the flat
list but missed the duplicates, leaving handler children in rootNode.
New approach: walk from graph.getRoot() following non-ERROR edges to
discover main flow nodes. Edges targeting handler compounds (ON_EXCEPTION,
ON_COMPLETION) are not followed. This cleanly separates main flow from
handler sections using the graph's own structure.
Falls back to flat list filtering (old behavior) when graph.getRoot()
is null (legacy/test graphs).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Root cause found: RouteGraph.getNodes() is a FLAT list that includes
handler compound children (log8, setBody1, etc.) as top-level entries
alongside the main flow nodes. The handler separation only identified
the compound PARENTS (ON_EXCEPTION) but not their children, so 7
handler children leaked into rootNode as main flow nodes, causing
ELK to place the real main flow at wrong Y positions.
Fix: two-pass separation — first identify handler compounds and
collect ALL descendant IDs, then build mainNodes excluding both
handler compounds AND their descendants.
Debug logging left in temporarily for verification.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The root cause of the Y-offset: ELK places main flow nodes at
arbitrary positions (e.g., y=679) within its root graph, and the
frontend rendered them at those raw positions. Handler sections were
already normalized via shiftNodes, but the main section was not.
Now useDiagramData.ts applies the same normalization to the main
section: computes bounding box, shifts nodes and edges so the section
starts at (0,0). This fixes the Y-offset regardless of what ELK
produces internally.
Removed the backend normalizePositions (was ineffective because handler
nodes at y=12 dominated the global minimum, preventing meaningful shift
of main flow nodes at y=679).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Based on thorough code review, fixes all identified issues:
1. **Y-offset root cause**: Added post-layout normalization that shifts
all positioned nodes and edges so the bounding box starts at (0,0).
ELK can place nodes at arbitrary positions within its root graph;
normalizing compensates regardless of what ELK computes internally.
2. **Bounding box**: Compute from recursively flattened node tree +
edge point bounds. Removes double-counting of compound children
(children have absolute coords, not relative to parent).
3. **SVG double-drawing**: Compound children were drawn both inside
drawCompoundContainer and again in the allNodes loop. Now collects
compound child IDs and skips them in the second pass.
4. **findNode**: Now recurses into children for nested compound lookup.
5. **colorForType**: Removed redundant double-check on EIP_TYPES.
6. **Dead code removed**: routeNodeMap/indexNodeRecursive (populated but
never read), MIN_NODE_WIDTH/CHAR_WIDTH/LABEL_PADDING (unused).
7. **Static initialization**: LayoutMetaDataProvider registration moved
from constructor to static block (runs once, not per instance).
8. **Debug logging removed**: Removed diagnostic System.out.println.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. findCommonParent: replaced with correct lowest common ancestor
algorithm using ancestor set intersection (previous version only
walked from node 'a', not a true LCA)
2. Bounding box: compute totalWidth/totalHeight from actual positioned
node coordinates instead of rootNode.getWidth/Height. The rootNode
dimensions don't account for handler sections in separate ELK roots.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
