Workflow · Start a ticketDraft
A ticket goes in. A verified root cause and a reviewed plan come out.
The first thing you run in a fresh worktree, before any code is written. It reads the ticket, checks the ticket’s diagnosis against the actual code, finds the root cause with a team of investigators, and stops with a plan for you to approve.
/start 12312 4435One or more ticket numbers. With no argument it reads the number from the branch name.Open source, free: skills/start/SKILL.md. The step numbers below match the file.
Step by step
- 0
Resolve and claim the ticket
Ticket numbers come from the argument or the branch name. If your tracker has a claimed convention, the ticket moves to in progress with a comment naming your branch, so a teammate scanning the board does not start the same work.
- 1
Fetch the tickets
Full ticket content plus recent history, for every ticket, in one parallel batch.
- 1.5
Already shipped?Blocking gate
Mandatory, and it runs before any investigation. In a shared repo someone else can ship the same fix while you are reading the ticket. The check is cheap; a full cycle on work that already merged costs hours and leaves a duplicate branch.
- 2
Read for symptoms
The ticket is read for what the reporter observed, kept separate from what the reporter believes caused it. Keywords are extracted for the investigators.
- 2.9
Is the premise true?Blocking gate
Mandatory for tickets that add to or harden an existing feature. "X has no retry" assumes X works and only lacks retry. That assumption is checked before a plan is built on it. A plain bug report skips this gate.
- 3
Root-cause investigation
Three kinds of read-only agents are dispatched together. By default they run as a team and message each other, so a cause that spans two domains surfaces. For a narrow ticket they run isolated, which is cheaper.
- codebase-navigator: a fast search agent maps the modules and call sites involved
- domain investigators: one per domain the ticket touches, from a routing table you define
- fresh-eyes: gets only the raw ticket text, no hypothesis and no tools, and says what it thinks the cause is
- No agent applies a fix during investigation, whatever tools it carries
- 4
Synthesize the root cause
The lead reads every report, and in team mode the conversation between the agents as well. If the cause cannot be confirmed it says so, states what would settle it, and marks the hypothesis unconfirmed.
- 4.5
Fix the class, not the instance
For bugs only. A separate read-only agent takes the confirmed cause and looks for sibling sites with the same shape of bug. Skipped for features and for one-line mechanical fixes.
- 5
Write the plan
The plan is saved to your plans directory as a file you can open, with a fixed structure.
- Root cause and the fix
- Implementation steps, in phases
- Testing, including why this was not caught and how it will be prevented
- The bug class and its guard
- Risks and adjacent code, and what is out of scope
- 6
Plan review
The plan goes to the plan-review skill. Blocking findings are folded into the plan before you see it. Without that skill installed, the plan comes straight to you for sign-off.
- 7
Your reviewYou decide
A summary with the full path to the plan file. Nothing is implemented until you say go.
Why it is shaped this way
Orchestrate, don’t investigate
The lead agent frames the question and judges the answers. Subagents do the reading and tracing, so the lead’s context is spent on judgment.
The ticket is right about the symptom
Tickets are usually right about what broke and often wrong about why. The stated cause is verified against the code before anything is planned on top of it.
Check it is not already done
Investigation and planning are expensive. Checking whether the fix already shipped is cheap, so it runs first.
Fail conservative
If the premise or the root cause cannot be confirmed, the skill stops and asks. It does not plan on an unconfirmed foundation.
What it will not do
- Write or change code. Everything before your approval is read-only.
- Trust the ticket’s diagnosis without checking it.
- Present a guess as a confirmed root cause.
Questions
What is maxx-start, then?
A launcher. It creates the worktree for the tickets, opens your agent in it on the model you name, and runs /start. /start is the skill; the launcher saves you the setup.
Can it take several tickets at once?
Yes. Related tickets are investigated together, which saves time and tokens when they touch the same code.
Does it need a particular tracker?
No. The command that fetches a ticket is a configuration slot. The worked example uses GitHub issues.