Specrew v0.40.0-beta41

Feature 202, the Beta4.1 UX stabilization. BETA - this prerelease is not a stable promotion. Gallery prerelease
labels may not contain a dot, so Beta4.1 is published as 0.40.0-beta41.

Beta4.1 changes what you see and how often you are stopped, not what governance requires. Validated on Claude
Code, Codex and Copilot CLI; Cursor and Antigravity are experimental. The full notes, including how this release was
validated, are in docs/release-notes-v0.40.0-beta41.md.

Fixed

• Your approval is recorded once, keyed by the host's own message ID, so no spare approval is left behind. On
  Copilot, an approval that was not recorded is reported.
• Copilot no longer records everything twice by also running Claude Code's hooks. Passed live on Copilot CLI
  1.0.101.
• A sub-agent's task text is never taken as your approval or your answer.
• Codex workshop answers are recorded. A brief journal lock is retried, a lasting failure is reported in the same
  turn, and specrew status leads with a health line while it lasts.
• The packet is shown once per session, and again only after a resume, a compaction, a long gap or a change to
  what it shows. A side question gets an answer.
• The Stop hook no longer blocks a turn over wording, and a missing profile line is filled in for you.
• The orientation comes before the first question, and the adaptation sentence reads the same on every host.
• Status, orientation, resume and specrew start name the same feature and say which record decided it.
• A second feature in a worktree can reach specify, and readiness no longer reports a false READY.
• Refusals name the step that fixes them; the design-analysis refusal lists every gap at once.
• specrew init inside another repository's folder no longer commits to it or writes .github/ at its root.
• A closeout is recorded only on your approval, and two sessions no longer see each other's refusals.

Changed

• The approval form on Claude Code: a phrase on its own line counts even beside a paste; text inside a paste never
  counts; a message quoting more than one offered reply is refused as a quoted menu. Other hosts keep the one-line
  rule.
• "the remaining lenses" delegates every open workshop lens; the reviewer is chosen once, at the agenda.
• specrew update removes the retired .specrew/runtime/workshop-receipt-outcomes.jsonl and says so.

Known issues

• Rarely, a typed approval records nothing and nobody is told (B4F-151). If Specrew cannot read the form of an
  approval, it records none, and the notice saying so is lost. Nothing unsafe is recorded. If the work does not
  move on after you approve, type the approval again. The fix is in Beta5.
• On Copilot, a workshop question can open Copilot's question box, and an answer typed there is not recorded
  (B4F-131). Copilot's recovery is a second question box, and an answer there is not recorded either. Answer in
  the chat as a normal message, the recovery included; if an answer went into a box, type it again in the chat.
• On Copilot, pressing Esc on the question box can swallow your next message (B4F-153). Copilot folds it into the
  running turn, and no hook sees it. If Copilot carries on as if you had not answered, type the answer again.
• On Copilot, a boundary packet runs its reply options together and shows its hidden tag (B4F-156). Type one
  reply on its own (approved for <stage>, changes needed: <what to change>, ...); the tag needs nothing from you.
• A workshop turn that waits for you can end without saying what to type (B5F-011). If a workshop turn ends and
  nothing follows, it is waiting for you: look in its last paragraphs for the reply it asks for, such as move on.
• The guidance specrew init prints after setup is out of date (B5F-008). It names a guide that is not in your
  project and describes steps the crew does. Launch your host in the project; the crew drives the lifecycle.
• After the review allowance is used up, a review-coverage block can stop every turn end, even in a retro or a
  closeout (B5F-012). It counts records and test data as changed source and offers implementation choices that do
  not apply there. Claude Code shows it as "Stop hook error"; it is not a crash, your work is safe, and outside
  implementation there is nothing to decide.
• Beta4's two known issues are unchanged: lens confirmations bind the next typed reply, and a scaffolded feature
  you never started keeps announcing its workshop.