Dashboard and run report

Overview — run KPIs (status, elapsed, agent runs, peak agents), the completion bar computed from actual ticket state (a running ticket is never counted as complete), per-status glowing tiles, agent/model activity and the error/feedback log.

Workflow — the hierarchical ticket graph: one card per ticket, arrows for real dependencies only (never invented), stage columns from the dependency waves, zoom / pan / fit-to-view, click a card for details (model + provider, attempts, timestamps, feedback, per-ticket event history, upstream/downstream links).

Tasks — every ticket as a row: status, attempts, model, scope, acceptance command, feedback.

Agents & models — which model ran which tickets, per attempt, from the recorded state.model.

Plugins — the agent plugins the run’s agents actually load (project config, global config, local plugin directories), with not-installed expected plugins and an install hint.

The start-time × duration diagram: one bar per ticket/attempt, placed by the real wall-clock started_at / finished_at / last_attempt_at timestamps recorded in state.json (never reconstructed). Parallel work shows as overlapping bars (T2/T3 in the showcase, peak agents: 2); zoom in to read the sequence. Events (status transitions, attempts, operator replies, run termination) are plotted on the same time axis.

Status colours: merged = green, running = blue, pending = amber, blocked = red, failed = grey. The failed status is rendered for completeness — the current orchestrator emits blocked at exhausted attempts; nothing in production code sets failed, and the three terminal-ish states are distinct and never inferred from each other.

The run-status tile shows RUNNING / FINISHED / STALE / NO_RUN (run.status); a fresh watch heartbeat adds the "watch standing by" badge (mail-loop.adoc#watch).

orch.py run (re)generates out/run_report.html at start, on every state change, and at the end. It is a self-contained document — no CDN, no network, works offline and straight from file:// — with the same workflow graph, KPI cards, a sortable/filterable task table, agent/model activity, timeline and errors section, plus the full Markdown of all tickets embedded at the end (a real internal hyperlink in the PDF). In the static report the details panel pre-selects the most actionable task (blocked/failed first, then pending with blocking feedback).

out/run_report.pdf (A4, print-optimised CSS, interactive controls hidden) is produced by Playwright’s headless Chromium — the report is client-side rendered, so only a real browser engine prints it faithfully.