Liner Code 0.1.6 (alpha)

Update from 0.1.5 if you have it. Liner Code does it by itself (Restart to update), or
download the new DMG.

New

• Memory across sessions. Agents save what they learn about a project with memory_add, and
  every new session there starts with it; sessions_search reads earlier sessions. Settings ›
  Projects shows the memory and lets you edit it.
• Agents outside Liner Code can work with your sessions through its MCP server: list, read,
  start, message and wait for them. It's off until you turn it on in Settings › Safety, with a
  key only you can read.
• Agents check pages when the session isn't on screen. The app keeps a hidden Browser for
  it, where scripts run and the console is read, and a screenshot still works.
• The Browser opens only what's the session's. A dev server is known by the session that
  started it, plain HTML opens from the session's folder with no server, and archiving or
  removing a session stops the servers it started.
• Carrying on at a usage limit: another account of the same tool first, else another
  family's everyday model, from the limit card or by itself (Settings › General).
• Reviewer decides asks the Reviewer's quickest model about each step and brings you only
  what it won't allow. Auto is the tool's own auto mode.
• Routed work: the model for each role (everyday, hard jobs, judge, critics) is yours to pick.
• Codex's command output streams into its running step, and every running step counts up.
• A turn that has said nothing for 4 minutes (12 while a command runs) says so once, on the
  timeline and as a notification.
• A tool's update shows its progress as a bar.
• Agents get a ripgrep when the Mac has none, and a failed command says how it ended (exit 127:
  not installed).

Changed

• A plan writes quietly. While the lead writes its plan, one line says so, with its steps so
  far, and opens for a look; the plan card takes its place when it's done. Claude Code's own plan
  mode gets the same line.
• Edits in a row to one file are one step on the trail ("7 edits · ✓ You allowed ×4"), and
  thinking too short to show leaves no gap.
• The sidebar names the lead while it plans and while its plan waits on you.
• Esc closes the / or $ list first; it stops the turn only after.

Fixed

• The context ring read 100% in long sessions: it added up every request of a turn. It reads the
  last one now, for Claude Code and Codex.
• A message sent as a turn ended could start a second turn beside it, and one sent during a
  review never went. What you send while work is in flight now waits and goes when it ends; Stop
  keeps it for you.
• Auto and "don't ask again" took ls & rm …, a command on a second line, rg --pre or a sed
  that writes as reading. A routine's trust lasts only for its own step, and its steps never
  push, merge or post.
• Stop pressed while a tool was still opening didn't stop it, and a request left from a turn
  that had ended kept the session in Needs you.
• Claude Code: a steer's answer could be lost, spend was counted again after a restart, extra
  usage read as a usage limit, and a plan you edited before approving wasn't the one it built.
  /usage and other local commands now show what they print.
• Codex showed an error twice, and a stopped command could go back to running. An ACP agent's
  text either side of a step ran together, and its cost was counted more than once.
• A failed send lost your prompt and files. Editing a plan step lost focus after each letter.
  ⌘⇧D and ⌘⇧P could settle or pin the open session. A Browser step's line ended in a stray dot.
• Keeping the default branch pulled no longer moves a checkout under a session working in it.
• First run said "1 tool and 0 keys is ready", times could read "1m 60s", and settings switches
  had no names for VoiceOver.