The dashboard's primary debugging screen — 'Why did my agent do that?' — with context assembly breakdown, per-memory score components, tool call timeline, and raw JSON view.
Milestone 4.D.2 — Trace Inspector
Status: Planned
Goal: Track D — Operator Dashboard
Phase: 4 — Operator Platform & Multi-Provider
Estimated effort: 3 days
Track: Track D — Operator Dashboard
Why This Milestone Exists
U1 in the UX spec: "Debuggability is the primary UX goal." The trace inspector is the screen every other page links into — from traces to sessions to agents, from traces to memories used. It renders the full context assembly breakdown with per-memory score components (relevance/recency/usefulness/confidence weighted per category), making the scoring redesign auditable in production.
Six-Section Layout (per UX spec)
| Section | Data Source | Component |
|---|---|---|
| Summary header | ClickHouse trace + AssemblyMetrics | TraceSummaryHeader.tsx |
| Context assembly | ibex.context.v1 trace payload | ContextAssemblyPanel.tsx |
| Conversation | Session archive (MinIO) or ClickHouse | ConversationView.tsx |
| Tool calls | ATP/tool-call tracking | ToolCallTimeline.tsx |
| Raw JSON | Full trace payload (permission-gated) | RawJsonViewer.tsx |
| Cross-links | Memory detail, directive diff, session replay | Inline links |
Orientation (indicative)
Named paths, package layouts, libraries, schemas, env vars, and commands anywhere on this page are rough sketches for orientation — inspiration and a baseline, not a required change list.
During implementation, expect to:
- open the live tree and follow existing patterns before inventing new ones
- research current constraints (latency, tenancy, deploy shape, libraries) more deeply than this page can
- advance the design beyond the sketch where measurement or code reality says so
- land work in different filenames, merged packages, deferred docs, or new surfaces when the situation calls for it
Prefer outcomes over matching any particular file tree or command sequence.
Areas that may be involved (situational — not a checklist):
- Operator dashboard
- Management API
Success signals
Outcome-oriented signals that the milestone is in good shape. Exact filenames, package layouts, and commands may differ from any sketches above.
- All 6 UX-spec sections render with real data end-to-end for a synthetic trace
- Score breakdown component shows per-category weight vectors, not a flattened number
- Raw JSON panel gated by
trace:read_rawpermission; unauthorized users see "request access" state - Golden-trace fixture test: known input → exact expected rendered summary
- P95 page load < 2s measured against real payload
Prerequisites
- Milestone 4.D.1 (Dashboard foundation)
- ClickHouse trace ingestion operational
- Context assembly service emitting
AssemblyMetrics
Last updated on