v5.15.0

Added

• Add a self-contained Japanese product guide showing what to request and
  what results to expect. Keep model settings in expandable details and make
  the task flow readable on mobile screens.

Changed

• Describe current task execution, role defaults, manual model choices,
  progress reporting, and restart behavior in the English, Japanese, and Codex
  READMEs. Distinguish task-count progress from acceptance results.
• Route Claude deep-planning and advisor defaults to Fable 5.1/high and Codex frontier
  roles to GPT-6 astra. Keep lightweight workers and existing Codex role effort
  separate, and retain the isolated Sonnet security reviewer.
• Calibrate active workflow, agent, embedded CLI, and setup prompts around
  scoped outcomes, recoverable inputs, independent delegation, and observed
  completion evidence. Preserve required checks, review limits, and permissions.

Fixed

• Keep handoff timestamps numeric during session initialization and resume
  on both GNU and BSD systems, so Linux can restore structured handoffs.
• Isolate advisor configuration regression tests from the operator's active
  settings while continuing to verify explicit project model choices.
• Normalize supported working-directory aliases before native companion
  dispatch so reviews inspect the requested project.
• Deliver large plan context and worker status through input streams, avoiding
  operating-system argument limits while retaining full evidence and bounded summaries.
• Bind the managed Codex reviewer profile during setup so an existing default inline
  role cannot silently shadow its model and instructions. Preserve custom
  role bindings and distinguish native permission inheritance from companion
  read-only review execution.
• Preserve explicit Codex model and effort through companion dispatch, including
  Codex ultra, without prompt-based effort calculation replacing role settings.
• Resolve advisor defaults for the target project and preserve its explicit
  model choice instead of reading configuration from the installed plugin.
• Keep task execution and the primary-environment guard on the same working
  directory and write intent; reject unsupported config overrides and extra
  writable roots before dispatch.
• Include the runtime helpers required by advisor and loop execution in the
  Claude and Codex plugin packages.
• Carry the selected plan, completion criteria, concrete failure evidence, and
  prior advice through loop worker and advisor dispatch, including restart.
• Preserve Go worker task descriptions and review refinements, and make the
  opt-in review verdict contract explicit instead of accepting an APPROVE
  substring in contradictory text.
• Include task completion criteria, inferred scope, exclusions, and reviewer
  notes in browser review instructions without changing its schema or routes.