maxritter/pilot-shell v12.0.0-beta.28

on GitHub

New

• A task now takes seven steps: Discuss, Research, Plan, Outline, Implement,
  Verify and Review. Each step writes one Markdown file, from 01-discuss.md
  to 07-review.md, that holds everything about it, and there is no separate
  folder of agent notes. Discuss, Research and the Outline show their document
  the way Files does, with Sections and Copy; open questions take
  its place while you answer them. A shared link on qualitylayer.dev shows the
  same document, diagrams included. You still approve twice: the Plan before
  any code, and the finished change.
• Tell me more asks what would help: I don't get the question, Explain
  the choices or your own words. The card shows your agent picking it up and
  rewriting the question in plainer words, with the earlier wording one click
  away. A question can let you pick more than one choice.
• Teams follow each other's tasks. Follow a teammate's shared task to hear
  its milestones in Slack. Slack messages are grouped per person, and each one
  opens the exact step or ask in the App. Everyone sets a morning digest, quiet
  hours and a time zone.
• For you on Home lists everything waiting on you, and the sidebar badge
  and the bell count the same things. The task header says who an approval
  waits on.
• A teammate's shared task uses the same header and all seven steps. Teammates
  can ask and comment on any step's passage.
• Research is its own step. Agents read the code without your request in front
  of them, so the Plan rests on how the code works today. The findings are
  written up in 02-research.md (for a bug, Why it breaks), and Research
  starts while your Discuss questions are still open.
• Questions come in batches, only in Discuss and Research. A batch has three to
  six questions, each with its facts side by side, and a step with nothing to
  ask has no batch. A strip lists the whole batch, and Use the
  recommendations for all answers the rest at once. The Plan asks no batch;
  a choice that only appears once a design is drawn sits inside it as Your
  call, at most two per task.
• The Plan is reviewed before you are told. The review starts with the first
  full draft, and your agent folds the findings in before the Plan reaches you.
  With one agent installed, a separate reviewer of that agent reads the Plan
  instead. Reviewed by shows who read it and what happened to each finding.
  The contracts between the parts show on the Plan, folded to one headline
  each.
• The Outline cuts the approved design into slices while you read the Plan.
  It needs no approval, and you can start the build as soon as you approve.
  A command that pushes, deploys or reaches an outside address is refused
  unless the Plan's Before the build granted it.
• Verify runs six named checks in a fixed order: Polish, Security review,
  Project checks, Independent review, Second opinion and Fix rounds. The same
  names appear in the App, the commands and the docs.
• Settings has two Second opinion selects: On the Plan (Always, When risky)
  and On the built change (When risky, Always).

Fixed

• The App on macOS no longer quits unexpectedly when its output closes, so
  macOS stops asking to reopen its windows at the next start.
• On Windows, a hook stays silent when the App or the CLI is missing or blocked
  (for example by a Defender rule), instead of showing a hook error on every
  tool call. qualitylayer doctor reports it. install.ps1 -Uninstall removes
  QualityLayer without running any of its programs and keeps a copy of each
  agent settings file it changes.
• /ql inside pasted text still starts QualityLayer in Claude Code, Codex and
  the other agents, also when the paste starts with another slash command.
• When the Plan's review cannot run, the agent tries once more and then says
  why, instead of counting the review as done.
• Browser checks no longer fail when the terminal forces coloured output.
• After an update the App always shows the new version: a server from another
  build is replaced, an install stops the old servers, and a server nothing
  points at ends by itself.
• A long question can be scrolled to its Send button. While a batch is open it
  fills the page, and its counts agree with the strip.
• Your agent wakes once when a batch is answered, not once per answer, and at
  once when you ask for more.
• The task page has one sidebar toggle, in the header.
• On a shared seven-step task, a teammate's ask holds the right approval, and
  teammates see the Review conversation and the Plan's picture.
• The Plan page shows when the Outline's check is still running beside
  Approve, and lists what changed in the Plan after you approved it.
• The Plan's reviewer looks at the rendered mockups, not only their source.
• The Implement page folds its checks into one line, and opens it when a check
  failed.
• The Verify page opens with how the change is checked. The commands the
  project checks ran or reused, and how far a running check is, sit in the
  Project checks lane instead of in two blocks above it.
• A message from another agent session that mentions /ql no longer starts
  QualityLayer.

Good to know

• New commands: qualitylayer plan draft checks the Plan and starts its
  review, and outline amend records a change to the Outline made while
  building. question ask --batch refuses a batch the step may not ask, one of
  more than seven questions, and a fourth batch in a step. question more
  answers one Tell me more without resending the batch, question show prints
  a batch, question link ties answers to Done means points, and helper result keeps a research helper's report in the task.
• Tasks you started before keep the files and the five steps they began with.