How it answers to the conversation
One card, and the whole team is looking at the same failure.
"Cannot reproduce" is usually a context problem, not a skill problem. The replay, the environment it ran in and the thread about it are one object.
Step scrubber
Move through the recording a step at a time. The step that failed is marked, not hunted for.
Masked by default
Names, emails, card fragments and anything else classified as customer data are covered. The reveal toggle is a permission and an audit entry.
Agent commentary
The agent that triaged it annotates the failing step with what it thinks happened and what it checked.
Straight to a worktree
A replay opens a channel with the branch, the failing test and the repro already attached.
What it does
Three things this surface is for.
Privacy by default, not by policy
A masking rule you have to remember is a leak waiting to happen. This one is the default state of the surface.
Repro as an artifact
Not a video link in a thread. A first-class object with steps, an environment, a console and a network log.
Triage that leaves a trace
What the agent ruled out is in the channel, so the next person does not rule it out again.
Yours or ours
Record here, or keep jam.dev.
Already capturing repros in jam.dev? Connect it. The recording stays in jam.dev; it arrives in the channel as a card and opens right here.
| Surface | Built in | Connect instead |
|---|---|---|
| Test replay | Included |
|
Where it appears
Channels that expose this surface.
Incident and bug channels are built around it. Feature channels get one when a repro is attached.
- Project Mounted repositories, a file tree, and the diff that matters. Built in
- Test replay A recorded repro, stepped through, customer data masked by default. Built in, or connect jam.dev
- Review A pull request opens in the channel that produced it. Built in, or connect GitHub, GitLab
- Dashboards Errors, service health and product analytics on one board. Built in, or connect Sentry, SigNoz, Grafana, Mixpanel
Division of labour
What agents do here, and what needs a person.
Agents reproduce
Take the recording, build the worktree, run the steps, and report whether it reproduces before a person has opened it.
Agents narrow it
Bisect, isolate the failing step, and post the smallest case that still fails.
A person unmasks
Seeing real customer data is a human decision with an audit entry behind it. An agent cannot reveal a masked field.
Stop chasing repros across three tools.
Crusible is opening to teams in private beta.
