v3.9.0

TL;DR — codeep --yolo -- "<prompt>" starts Codeep for a launcher with nothing left to answer, Omarchy's skills in ~/.agents/skills load, and Codeep's colours follow the Omarchy theme.

Added

• codeep -- <prompt> starts a new session with <prompt> as its first
  message (also -p <prompt> and --prompt <prompt>). Everything after --
  is the prompt, so codeep -- review asks the model and does not run
  codeep review, and a leading / is not run as a slash command either:
  codeep -- /commit sends "/commit" to the model. A missing or doubled
  prompt is an error rather than a guess.
• codeep --yolo starts without stopping to ask, for this launch only:
  agent actions run without confirmation, a vague task gets no clarifying
  questions, the folder is used as the project with read & write access, and
  a new session starts. Nothing is saved, so the next plain codeep asks as
  before; changing Agent Confirmation or Agent Interactive Mode in
  /settings ends --yolo's hold on that setting, and so does loading a
  saved profile for Agent Confirmation. Writes to files that
  decide what runs later are still confirmed, and a workspace's own MCP
  servers still ask to be trusted. The status bar shows YOLO while
  confirmations are off.
• Skills in ~/.agents/skills load, the directory agent tools share and
  where Omarchy links its skills. When two skills have the same name, the
  project's .codeep/skills wins, then ~/.codeep/skills, then
  ~/.agents/skills. /skills bundles names the directory each came from.
• A skill can send the agent to the files beside it. invoke_skill now
  gives the skill's directory and reads a file from it (a guide such as
  reporting.md), which read_file cannot reach for a skill outside the
  project. Only files inside that directory are read.
• Codeep follows the Omarchy theme. On Omarchy, the colours Codeep
  paints itself — the brand red, the greys of the chat, the code colours,
  the warning orange and the YOLO badge — come from the current theme's
  colors.toml and change when the theme does, in a session that is already
  open, the way Claude Code, OpenCode and Pi do. A theme colour too faint to
  read as text is darkened or lightened until it reads, so light themes stay
  legible. Follow Omarchy theme in /settings (On by default, shown
  only on Omarchy) turns it off. Codeep only reads Omarchy's state; it
  writes nothing to your Omarchy config. Elsewhere every colour is exactly
  what it was.

Fixed

• Startup asks its questions one at a time. In a folder that is not a
  repository but has its own MCP servers, the question whether to trust them
  could replace "Set as Project?", leaving that one unanswered and the
  startup stopped. It now comes after the startup questions, and a launch
  prompt is sent after it.
• Folded and literal descriptions. A SKILL.md whose description: was
  written as a > or | block, as Omarchy's are, loaded with ">" as its
  description.
• The agent sees a skill's whole description. The catalog cut every entry
  at 200 characters, mid-word, which dropped the part of a long description
  that says when to use the skill.

Security

• Writing to .agents/skills/ is confirmed in every mode, as
  .codeep/skills/ already was: every Codeep session loads the skills in
  ~/.agents/skills.