v1.0.0

The first stable release. Everything a player touches was audited end-to-end
for production — the six capture paths (Claude Code CLI + IDE panel, Codex CLI

• IDE extension, Cursor, Copilot Chat), every CLI command, scoring parity with
  the server, the installers, and the docs — and the sharp edges found were
  fixed. No protocol change: a v0.4.x install upgrades cleanly with
  promptly update.

Fixed

• promptly init --force (and play --force) empties the folder before
  unpacking — keeping .git and .promptly/ (your reset backups and the
  solve clock survive). Unpacking over leftover files failed the post-unpack
  baseline check with a misleading "corrupt download" error that no retry
  could ever fix.
• A no-arg promptly play outside a level workspace fails cleanly instead
  of first ending the active capture session and rescoping the daemon to the
  wrong folder — the same wrong-directory guard start already had.
• promptly start <level> accepts the short level names every other command
  takes (lru, 7, stage-1-01). It falsely refused them in the correct
  workspace — including the exact form submit's own hint tells you to type.
• The signed prompt count P no longer inflates on agentic runs.
  Task-subagent transcript lines (isSidechain), meta records (isMeta), and
  slash-command lines in the Claude Code JSONL no longer count as user prompts
  or misgroup the turns that follow them.
• Copilot thinking text is counted once, in the thinking bucket. It was
  estimated into the output count and (when rounds record real counts) summed
  again as thinking; without recorded counts it now estimates thinking instead
  of inflating output.
• An unresolvable Cursor per-prompt model pick no longer inherits the
  previous prompt's — turns fall to the composer's configured model or
  degrade to estimated, never to a stale price.
• promptlyd status reports a stopped session as idle, matching what
  promptly status already said.
• Pairing no longer times out after a single poll if the server omits the
  device-flow window — a missing expires_in defaults to the usual 15 minutes.
• promptlyd install on macOS no longer fights promptly down and
  promptly update: the launchd agent relaunches only after a crash
  (KeepAlive.SuccessfulExit=false), the same policy as systemd's
  Restart=on-failure. And a service install now runs with the idle
  auto-shutdown disabled (--idle-timeout-secs 0) on every platform — an
  always-on agent used to exit cleanly after 15 idle minutes and, under an
  on-failure-only restart policy, stay down.
• promptly doctor's closing verdict counts warnings honestly ("4 of 7
  checks need attention" instead of calling warnings failures).
• promptly score's live context line shows the run time it actually
  assumed (--seconds included) instead of a hardcoded "2s".
• Internal build-plan numbers no longer leak into promptly --help, and the
  --api-url help names the real default (the production site, not localhost).

Changed

• Local projections score estimated captures exactly as grading does.
  Editor-adapter captures (Cursor / Codex / Copilot) floor the effort base at
  anchor parity — the server's fairness rule since 2026-07-15 — and the scored
  harness is the weight-dominant one over the signed sources, so watch,
  score, and the submit preview no longer project higher than the graded
  result on estimated runs. The breakdown says when the floor applied.
• claude-fable-5 spellings resolve like the other Anthropic tiers —
  datestamped or reordered forms find the priced row instead of the floor.
• A Codex session can never be re-scoped by a later metadata line: the
  session's original cwd stays authoritative, so a future Codex version
  persisting cwd-carrying events mid-session can't silently drop turns.
• A crashed adapter scan reports itself on /health / promptly doctor
  (unsupported, with a restart hint) instead of dying silently behind a stale
  detected status.
• The CLI notes a version-mismatched daemon when it reuses one that
  survived an upgrade (a service-managed or non-default-port daemon), with the
  promptly down hint to relaunch it on the new build.
• The installers upgrade a live install safely. Both stop or step around a
  running daemon instead of dying halfway, install.ps1 keeps the old binary
  as a rollback until the new one has landed, runs in its own scope (no
  variables leak into your shell via irm | iex), preserves a REG_EXPAND_SZ
  user PATH instead of flattening %VAR% entries, broadcasts the PATH change,
  and works on Windows PowerShell 5.1.
• promptly init --force refuses a folder that isn't a Promptly workspace
  — now that --force wipes before unpacking, a mistyped --dir must never
  point it at an arbitrary folder.
• Every release asset ships with a .sha256 checksum file.
• README correctness pass: the signed chain is v4 (not v3), all six capture
  paths are named — including the Claude Code IDE panel's JSONL-only caveat —
  and the install, idle-shutdown, and command-list copy matches the code.

Install / upgrade

• macOS / Linux: curl -fsSL https://raw.githubusercontent.com/AidanHT/promptly-daemon/main/install.sh | sh
• Windows: irm https://raw.githubusercontent.com/AidanHT/promptly-daemon/main/install.ps1 | iex
• Already installed: promptly update

Full Changelog: v0.4.8...v1.0.0