Plan the work

The plan lists what the agent will do and the questions it has, with the evidence for each.

On this page

Every Implement session starts with a plan. The agent reads the ticket, the code, the data and the reference documents that help with the ticket (it says which it used), and says what it will build, where, and where it thinks the story is wrong or unclear. Limn checks every piece of evidence in the plan before you see it. Nothing is built until you approve the plan.

#What the plan holds

Illustration: The plan: each criterion, the points to settle with their evidence, the pull requests, then Approve and start.
  • The repositories it will change.
  • Each acceptance criterion, in the ticket's order: Planned, Already met or Not planned.
  • Points to settle: each place where the agent says the story is wrong, unclear, or does not match the code, the docs or the data. Each says what the story says, what the agent suggests, and its evidence.
  • The pull requests: the agent splits the work so each pull request is smaller to review, at most three per repository and six in all. Each one must build and pass its tests on its own. In one repository, each builds on the one before it.
  • A ticket update (draft), when the story needs work.

#Evidence Limn checks

Limn checks every piece of evidence itself. A point it cannot check is left out. Evidence can be lines in the code as they were when the session started, a section of a reference document, a data file (a fixture, seed data, a schema or a migration) or a query the agent ran against a sandbox service and Limn saw run. Nothing comes from production. When Limn removes something, one line says how much, for example "Limn left out 1 point it could not check."

#Settle the points

For each point, press one of:

  • Agree: build what the agent suggests.
  • Keep the story: build what the story says.
  • Reply: tell the agent what you think. It revises the plan.

The heading counts your answers, for example Points to settle (2 of 3 answered). You can also talk to the agent freely in the conversation: Ask a question, or tell the agent what to change in the plan. When a reply changes the plan, Limn says The plan changed. Look at the new revision. Earlier revisions stay readable. Your answers carry to a new revision while the point is unchanged.

#Update the ticket

After planning, The ticket shows the whole ticket in two boxes: the description and the acceptance criteria. Each box starts with the agent’s draft, or with the ticket’s current text when the plan does not change it. The agent must write a draft when it raises a point to settle or does not plan a criterion, and it keeps the ticket’s layout: the same headings in the same order, the same list style, and GIVEN / WHEN / THEN when the ticket uses it.

Change the text as you like. In each box, the words that differ from the ticket are highlighted, in the same colour as a quote from your reference docs. Under each box, What changes shows the lines that differ from the ticket, and a note says where the text no longer keeps the ticket’s layout. The notes are advice; you decide.

#Approve and start

  1. Answer every point to settle.
  2. Press Approve and start. Limn reads the ticket again and the agent switches to building.
  3. Watch the build. The agent then builds to the end. You can stop it or message it.

Approve automatically

Tick Approve the plan automatically at Start, or in the build tab before the plan is ready. When the plan is ready, Limn takes every point as Agree, approves it and starts the build. The agent builds the plan as revised, its acceptance criteria included.

  • The ticket is not changed.
  • When the pull request is open, the build tab shows the ticket changes the plan suggested. Press Update ticket to write them to Jira, or No update needed.
  • When the AI spend limit is reached, or another build is running, the plan waits for you to press Approve and start.

What the build reports

When the build ends, The agent says gives, per pull request, Done, Partly done or Not done, a summary, and the checks it ran (Passed, Failed or Not run), and per criterion whether it is met. This is the agent's own account, not a gate: read the diff. If the agent ran out of steps with work unfinished, Limn asks it to carry on, at most twice.

Then push the result: see Push the result.