Operator CLI UX: FormRunner, huh pickers, Lipgloss chrome

[maxBRT]

Summary

Lands the Charm/huh operator UX slice: a FormRunner seam with shared Lipgloss theme tokens, huh Agent picker on Init, huh Queue multi-select with Shift+↑/↓ reorder, styled ship --help, and Lipgloss Run chrome (header, denser Phase observe strip, louder Abort). Color/ANSI for Run chrome, throbber, and observe is gated on stderr being a TTY so redirected logs stay plain.

Linked issues

• FormRunner seam + shared Lipgloss theme #67: FormRunner seam + shared Lipgloss theme
• Huh Agent picker on Init #68: Huh Agent picker on Init
• Huh Queue multi-select + Shift+↑/↓ reorder #69: Huh Queue multi-select + Shift+↑/↓ reorder
• Styled ship --help (Lipgloss) #70: Styled ship --help (Lipgloss)
• Run chrome + denser phase/observe strip #71: Run chrome + denser phase/observe strip

RISK

• TTY / non-TTY mismatch (medium). Init and Queue confirmation now depend on huh detecting a real terminal. CI pipes, script, or weird PTYs can mis-detect and either skip the picker (ErrNonInteractive / default Agent) or hang waiting for input. Operators piping ship output or running under automation are most affected.
• Queue order surprises (medium). Selected order is Run order; Shift+↑/↓ only reorders the focused selected row. Operators who expect list position (not selection order) to drive Iterations can process Tickets in the wrong sequence.
• Cancel semantics (low–medium). Esc/cancel on Init must not write .ship/config.yaml; cancel on Queue confirmation must abort the Run before Claim. A regression would leave a half-written config or claim Tickets the operator never confirmed.
• ANSI in help when redirected (low). WriteUsage / ship --help still emits Lipgloss color unconditionally; only Run chrome/observe/throbber are TTY-gated. Log scrapers or screenshots of --help may show escape codes.
• Charm dependency surface (low). New direct deps (huh, lipgloss) can break builds if module resolve fails or Charm APIs change; limited to this CLI binary, not agent Phase adapters.

QA — stress-test plan for a human

1. Happy path

1. In a clean checkout with gh auth and at least two GitHub issues labeled ship, run ship init on a real TTY: confirm the huh Agent single-select (blurbs + keyhints), pick an Agent kind, and verify .ship/config.yaml gets that agent:.
2. Run ship (or ship --help then a full Run): confirm styled Usage/Flags/Config sections and aligned flag columns.
3. On Queue confirmation, multi-select ≥2 Tickets, use Shift+↓ / Shift+↑ to reorder the focused selected row, confirm. Verify the Run processes Iterations in that selected order (Implement → Review per Ticket).
4. Watch Run chrome: once-per-Run header with Agent kind + branch; throbber shows dim remaining-queue hint while Tickets remain; successful Phase lines get a compact tools/tokens strip; finish through Final / PR open if you have a short queue.

2. Edges / failure cases

1. Non-TTY Init: ship init </dev/null (or pipe stdin closed) — should default Agent to cursor without a TUI and still write config (or fail cleanly per existing non-TTY Init rules); must not hang on huh.
2. Non-TTY Queue: ship </dev/null with candidates present — expect ErrNonInteractive / no Claim; no hang.
3. Cancel Init: open ship init, Esc/cancel before accept — no .ship/config.yaml write (or no overwrite of agent if you staged a dry run carefully).
4. Cancel Queue: start ship, Esc on Queue confirmation — Run stops; Tickets must not be Claimed / moved to In Progress.
5. Empty selection: try confirming Queue with zero Tickets selected — must refuse / stay in picker per form validation, not start an empty Run.
6. Abort chrome: force a Phase failure (bad Agent binary or short --timeout) — stderr shows loud red Abort, report path under .ship/runs/…, last tools; redirected stderr (ship 2>run.log) should be plain text without ANSI.
7. Reorder focus: with 3+ selected Tickets, move focus to an unselected row and hit Shift+↑/↓ — confirm only selected rows reorder and Run order matches selection order, not visual list index alone.

3. Out of scope

• Agent adapter correctness inside Implement/Review/Final (cursor/pi/codex/claude prompts).
• GitHub label ensure / Claim / Done / Abort tracker edge cases beyond “cancel did not Claim”.
• install.sh / GoReleaser / update paths.
• Replacing typed-line TypedLines FormRunner beyond the production huh path (tests may still stub it).

Made with Cursor