Skip to content

Runs and hypotheses

A run is one job for the root-cause agent: it investigates one or more scenarios and produces a hypothesis for each. This page shows how to start a run, follow it live, read the hypothesis, and manage runs, and explains what the Scenarios library is.

Runs belong to a project: the Runs page and the Dashboard show the runs of the project selected in the project switcher.

Start a run

You need the member role or higher.

  1. Open a signature that is mapped to a scenario (see Signatures and scenario mappings).
  2. Select Run root-cause.

DV Forge opens the new run's page. The run is linked to the signature: the signature shows the agent's progress, and its watchers and assignee are notified when the run finishes.

  1. Open Runs in the sidebar and select New run.
  2. Under Scenarios, select one or more scenarios. Each shows its bug class, expected duration and expected cost.
  3. Under Model, choose Sonnet 4.6 (the default) or Opus 4.7 for scenarios that need deeper reasoning.
  4. Select Submit run.

DV Forge opens the new run's page. With several scenarios selected, the agent investigates them one after another. A run started this way isn't linked to any signature.

Run statuses

Status Meaning
submitted Queued; the agent hasn't started yet
running The agent is working
complete Every scenario finished
failed At least one scenario failed (for example, it timed out)

Each scenario inside a run has its own status: pending, running, complete or failed.

Follow a run live

The run page header shows the run id, its status, when it started and finished, and its total cost and duration (live while it runs). The page updates by itself while the run is submitted or running.

Each scenario has a card with the scenario id, its status, and a title that reads Investigating… while the agent works and Investigation once a hypothesis is in.

Inside the card, Tool-call timeline lists the steps the agent has taken, with a count. It's open while the scenario runs and collapsed afterwards; select it to expand or collapse. Each step shows:

  • the name of what the agent did, and the step number,
  • whether it's still running, succeeded or failed, and how long it took,
  • a short preview of what the step was given and what it returned.

Before the first step, the timeline says "Waiting for the agent to fire its first tool call…".

If a scenario fails, a red card under it shows the error.

Read the hypothesis

When the run is complete or failed, each scenario that produced a result shows a Hypothesis card:

Field What it shows
Confidence high, medium, low or none confidence
Location The suspected file and line (file:line)
Commit The commit the agent believes introduced the bug
Owner The author of that commit
Summary The agent's explanation, in a few sentences
Evidence The references it relied on, each tagged with its kind
Footer The scenario's cost, duration, number of turns, and the model used

Because scenarios have a known answer, the card also shows a row of pass/fail check chips comparing the hypothesis with that answer. Hover a chip for a few seconds to see what it checks.

Treat a hypothesis as a lead to verify, not a verdict. When a run started from a signature finishes, confirm the root cause on the signature; see Confirm the root cause.

The agent's state on a signature

A signature tracks the latest run started from it, separately from its triage status:

Badge Meaning
agent running The root-cause agent is working on this signature
agent done The root-cause agent finished; see its run
agent failed The root-cause agent's last run failed

The badge appears on the Signatures list, on the signature page, and in the Root cause column of a regression's signature table (with view run →). On the signature page, view latest run → opens the run. No badge means the agent was never run on the signature.

When the run finishes, the assignee, the watchers and the person who started it get an in-app notification, and the signature can be confirmed as root-caused. See Confirm the root cause and Notifications. Runs started from New run don't notify anyone.

The Runs list

Runs lists the project's runs, newest first, with the total count at the top. It shows the 50 most recent.

Column What it shows
Run The first 8 characters of the run id (links to the run)
Scenarios The scenario, or the first one and +N more
Status The run status
Cost Total cost, once known
Duration Total time the agent spent
Started When the run was submitted

Delete a run

You need the member role or higher.

  1. On Runs, select the trash icon at the end of the run's row.
  2. In "Delete run id?", select Delete. This can't be undone.

A run that is running can't be deleted; wait for it to finish. Deleting a run that was started from a signature removes the signature's view latest run → link.

Scenarios

Scenarios in the sidebar lists the bug scenarios the agent can investigate today. A scenario is a design with one known, planted bug and a failing simulation that exposes it, so you can see how the agent works and judge its hypotheses against a known answer.

The built-in library is a suite of three bugs in an HMAC design block: an off-by-one digest counter, a missing control-register reset, and a pipeline FIFO write race. Each card shows the bug class, title, scenario id, a short description, and the expected duration and cost. You can't add your own scenarios yet.