Run detail¶

/runs/:id is the deepest review surface in Studio. It has two display modes:

Live stream (/runs/new-:jobId) — Server-Sent-Events from the worker. Auto-redirects to /runs/:id once run.created fires.
Historical (/runs/:id) — Persisted run record. The default view once a run has finished.

Both modes share the same tabbed layout once the run has emitted its first persisted row.

While a worker is still alive, a sticky banner at the top of the page shows:

Elapsed timer — wall-clock seconds since the run started.
Current activity — e.g. Processing public.users (2 / 156 columns).
Batch progress — column N of M.
Live cost — USD spent so far, updating per batch.
Cancel button — sets the cancel token; the worker stops between rows (not mid-row).

The banner disappears once the run reaches a terminal status (Success / Failed / Cancelled).

Tabs¶

Summary¶

Run metadata: command, status, duration, started-at. Below:

Scope card — JSON representation of what was submitted.
Tokens & cost card — input tokens, output tokens, total tokens, cost at run time, cost at current rates.
LLM reasoning — when the provider returned a reasoning trace (Anthropic extended thinking, GPT-5 / o-series, DeepSeek-reasoner, Kimi K2.x), it is rendered here instead of being discarded.
Confidence distribution — pie / bar chart of high / medium / low confidence counts when the run produced more than a handful of rows.

Results¶

The main review surface. Paginated at 50 rows per page.

A ResultsFilterBar above the table exposes:

Control	Effect
Search box	Debounced 300 ms substring match on schema / table / column / comment
Sort dropdown	Natural order, confidence asc/desc, logprob asc/desc, name A→Z, status (unreviewed first)
Group dropdown	None, Schema, Table
Status chips	All, Unreviewed, Accepted, Skipped — with per-status counts
Review presets	Low confidence (<0.7), Has citations, Table-level only

The DataTable shows one row per asset (schema.table.column). Each row exposes:

Checkbox for multi-row selection (used by the ReRun action)
Asset path — schema.table.column, mono
Asset kind — column / table / view
Confidence pill — high / medium / low plus a logprob badge when available
Source — rag / provided / function / other
Alternatives count — number of additional drafts generated for this asset
Description — the chosen draft, inline-editable
Status — Unreviewed / Accepted / Skipped
Actions — Pin, Skip, Restore (for re-runs)

Expandable per-row sections show alternatives (with a one-click "promote" button on each), citations with snippet previews, the model's reasoning, and the version history when this row has been re-run.

Production warning chip. When the LLM under-produced the configured n_alternatives and the agent had to top-up retry and / or pad with FALLBACK strings to reach the target count, the row carries a ⚠ chip reading e.g. produced 2 of 3 requested (retry got 0, fallback padded 1). The FALLBACK entries render with placeholder text so you can spot them and either accept the partial set, edit the chosen description inline, or trigger Re-Run with a different model. See Alternatives diversity mode — hard guarantee on N for the mechanism.

Variations from an alternative ✨¶

Every alternative on a multi-alt row carries a small ✨ trigger next to its confidence badge. Click it to open the Variations modal — the chosen alternative is used as a seed, and the new run is anchored on it under either semantic (paraphrase) or lexical (shared vocabulary) mode. The new run appears in /history; the source row carries an audit pointer back to the seed via parent_run_id + seed_alternative_id. The trigger is hidden when the row has fewer than two alternatives — nothing to vary around.

Scope¶

JSON editor showing the scope the run was submitted with — useful when auditing why a particular asset wasn't covered.

Settings¶

JSON editor showing the run's effective LLM settings — n_alternatives, temperature, prompt_detail, verbosity, batch_size, max_tokens. Frozen at run time so re-runs from this page reproduce the original conditions unless explicitly overridden.

ReRun¶

Single-item: click the ↻ button on a row. Multi-select: tick one or more rows → click Re-Run in the toolbar. The dialog opens with:

Additional instructions — a free-text addendum appended to the original prompt so the re-run sees the original DB / docs / code inputs plus your guidance.
Advanced LLM settings (collapsed by default) — the same override panel that RunNew exposes. Edit any field to override the active LLM profile for this re-run only; leave a field at the profile default to inherit it. Fields exposed:
Generation: temperature, max output tokens, alternatives per column, column batch size, prompt detail, description verbosity, confidence signal, alternatives diversity mode (disabled when alternatives per column is 1), thinking budget.
Confidence thresholds: high, medium.
Cost overrides: input USD / 1M, output USD / 1M.

See Alternatives mode and Confidence signals for what each knob does.

When N > 1 items are selected the modal notes that defaults reflect your active LLM profile and overrides apply uniformly to all selected items.

The Reset to profile defaults link (visible once you've changed at least one field) rewinds every Advanced field to the profile's saved value in one click.

The original DB scope, database / catalog, and the cached first-run table profile are all reused, so the re-run is comparable to its source rather than a fresh shot. Cost amortises across re-runs because the profiling step doesn't repeat. The saved LLM profile on disk is never mutated — the override only affects this single re-run job.

Pending queue handoff¶

The new v2/v3 row lands in the pending queue automatically with its top alternative pre-picked, so a click on any of its alternative buttons routes through the existing patch / apply path:

Auto-seed. Top alternative becomes the chosen description. Swap to another with one click.
Per-asset supersede. The new row replaces any prior pending entry on the same (schema, table, column, asset_kind) — the queue never holds two competing picks for the same column.
Cross-version mutual exclusion. Earlier versions of the same asset go non-clickable while the latest holds the queue entry; the tooltip points you at where to deselect first.
Apply count. The Apply pending queue CTA and the N queued header tally entries on v2/v3 rows too, so a pick on any descendant updates the count immediately.

Variations behaves identically; the two paths share the same queue helper. See Variations — Pending queue integration and Pending queue for the full lifecycle.

CLI parity exception. The CLI does not expose a per-run /rerun --alternatives-mode (or --confidence-signal, etc.). The single tactical exception is /rerun --temperature for diversity nudges. The full override surface on the CLI side lives in the interactive picker on /run; see CLI run-and-apply. This asymmetry is deliberate — scripted CLI invocations stay reproducible from config.yml alone, while Studio's modal-based override is the experimentation surface.

Pinned cells¶

Every row has a Pin button. Pinning persists to localStorage under amx.compare.pinnedCells.<profile> and pushes a custom amx:pinned-cells-changed event so the topbar pin counter and the pinned-cells drawer stay in sync across tabs.

Pinned cells survive page navigation. They're useful for keeping a candidate set in view while you ReRun, compare, or apply individual rows.

CLI equivalents¶

Studio	CLI
Run detail Summary tab	`/history show <run_id>`
Run detail Results tab	`/history results <run_id>`
ReRun	`/rerun <run_id>[.schema.table.column]`
Skip / Restore	`/review <run_id>` interactive picker
Apply pending	`/apply`