Observability and storage

The dashboard, run traces, the experience timeline and the sandbox file storage.

#The dashboard

The Dashboard (/dashboard) is a health summary over a chosen time window (hour/day/week):

• the health strip: runs, in flight, model failures, pending approvals;
• providers and models: calls, p50/p95, retries, tokens, cost;
• projects grouped by org unit;
• the quality loop: gates, blocks, the judge and its agreement;
• sagas and tasks: budget, tokens, time, human approvals (HAB/HWT — hover a metric for the explanation);
• human factors: attention (HAB), idle wait (HWT), flow efficiency — on the same row as storage and server;
• the sandboxes widget: live containers, integral load (CPU/memory/network), hover — who and which run, click — jump to the trace.

Fault tolerance and scale bounds (replay, workflow history, Continue-As-New, quotas) — see docs/resilience-and-scale.md.

#The Supervisor tab and the /supervisor page

The supervisor model's monitoring surface (docs/supervisor-model.md): failed-run diagnoses collapsed into same-root failure clusters, health anomaly explanations, and the wave-B/C cases — knowledge contradiction triage and hygiene, evolution proposals, skill candidates, user friction, intake audit.

The queue works like this:

• the tab badge shows how many items need attention;
• knowledge conflicts expand into a side-by-side comparison of the two formulations (differences highlighted, cosine quoted); actions live on each side: Invalidate retires that knowledge; Skip closes the item unchanged; Generalize… replaces both with one broader statement (editable before it applies);
• bulk actions: row checkboxes, select-all, then 'Skip selected' or (when every selected row is a conflict) bulk invalidations;
• filter by case and project; resolved rows toggle;
• HAB coefficients and models are configured in Agents → Quality settings.

The run trace (Trajectory tab) shows the session's human factors next to cost: ⏳ human wait (HWT, measured), 👁 human attention (HAB, estimate), ✋ human decisions.

#The run trace

The trace is a tree of turns, newest first. An expanded turn contains everything that happened inside it: model calls, tool calls with arguments and results, approvals, knowledge events (knowledge/recall/hints collapse into a knowledge group with the aggregate "N learned · M extracted · K hints"). The window is 50 events; older ones load as you scroll down. Uncertain operations are flagged and resolved through reconcile; cross-agent knowledge disputes are shown as a banner.

A long model call is the first line of the trace: elapsed, tokens, speed and the channel state ("the token channel did not connect" — a transport problem up to the provider, not the model; "the provider does not stream (a single chunk)" — a gateway is buffering the response). It can be nudged with the ↻ button (the request is cancelled and the turn restarted) or cancelled. A running run_command shows a live console tail right in its line; after completion the output is available in the row details. Detached commands (background: true) never fall out of the trace window: a pinned "Live background operations" panel at the top streams each of them.

A stuck run is stopped with the ■ Stop action in the list (click twice to confirm): a forced workflow terminate that journals the terminal event — it lands even when a graceful cancel cannot reach the stuck workflow.

#The experience timeline

The Timeline (/experience) is a temporal unfolding of knowledge and runs: which knowledge appeared/changed, in which runs it was applied, where it worked and where it misled. Grouped by scopes, with retrospective audit: "what did the system know at the moment of that decision". An incident retrospective and repeat-mistake tool.

#Sandbox storage

The Storage page (/storage) is a browser for the project's persistent filesystem: the workspace (/workspace of the project — what agents see and what survives runs) and the toolchain cache. Capabilities:

• a lazy tree with sizes and modification dates;
• preview: markdown with formatting, code with syntax highlighting, inline images;
• download, upload, delete of files and folders (deleting root areas is left to the background sweeper).

Project quotas and disk usage live on the dashboard, the "Storage" widget. The full storage model — docs/sandbox-storage.md.

#Export and audit

All Journal events are append-only with idempotent ids; they reconstruct the full picture of any run. Access audit (who read secrets, who changed resources) lives in the corresponding administration sections.