MAIster records each execution attempt as a Run. The active-workspaces rail shows current work; the Run ledger also keeps finished, failed, scheduled, and scratch attempts.

​
Find a Run

Run history with filters Choose Active workspaces → See all or open /runs. Filter the ledger by project, state, source, runner, and date range. The URL stores the filters, so you can send the same view to another project member. Each row shows the task or Run identity, project, status, source, start time, duration, runner, and token count. Flow and platform-agent rows open /runs/{runId}. Scratch rows open /scratch-runs/{runId}. The task page has a narrower history: it lists every Run created for that one task. A retry creates another attempt and keeps the previous revision, cost, evidence, and decisions intact.

​
Read the Run page

Run workspace with graph and current result Start with the header and primary result:
  • task title, project, branch, state, current node, and chosen Flow;
  • pinned package and Flow revision;
  • runner and model snapshots for the sessions that performed the work;
  • total duration, token usage, and recorded cost.
The graph shows the route taken through the Flow. Select a node to inspect its attempts, resolved prompt, structured result, output artifacts, gates, and transition. Rework may create several attempts for one node; MAIster keeps the older attempts and marks superseded results as stale. Selected node and its execution state Recorded prompts are distinguished from template previews. A preview for an older attempt is not proof of what was sent. Standalone agent transcripts show dispatched initial, resume, rework, and message prompts alongside responses. Steered messages appear in the same conversation. Open full artifacts when a node or consensus result only shows a bounded excerpt.

​
Understand why a Run ended

Failed, crashed, or abandoned Runs show a recorded terminal reason when one is available. It distinguishes an operator stop, an agent-process crash, a lost turn or event stream, a budget or duration limit, and an unavailable workspace. Older records may have no further detail; the status alone does not establish the cause. Read the reason with the node attempt and retained work before choosing Recover, relaunch, or review. A result that could not be applied safely is a visible failure requiring investigation, not proof that the agent completed successfully. Host-pressure pauses have their own notice and continuation rules.

​
Open code, evidence, and history

The lower workbench provides these views: The right inspector contains actions that apply to the current state: respond to a human gate, recover an eligible attempt, interrupt a node, take over the workspace, request rework, export, or promote. MAIster checks project rights and Run state again when you submit an action.

​
History after workspace removal

Removing an eligible workspace keeps the Run and its recorded history. Runtime content remains subject to retention; a missing or expired artifact is reported explicitly. Live file and worktree-dependent diff actions require a worktree. The Git panel offers Reattach when a local branch, publication, or archive reference can restore it. Restoration does not restart the agent. Failed worktrees also have a cleanup countdown and are preserved before scheduled removal.

​
Compare attempts

The task page answers “what happened each time?” Use an Evaluation study when you need a recorded comparison across Flow revisions, coding agents, models, methods, quality measures, and price.