Archflow 2.4.0

Added

• Archflow runs on hosts other than Claude Code. scripts/build-adapters.mjs generates a
  host-native package from the plugin for OpenAI Codex, GitHub Copilot CLI, Cursor, Gemini CLI,
  OpenCode, and a generic AGENTS.md + Agent Skills package for anything else. Each lands in
  adapters/<host>/ with its own README describing install steps and what that host cannot do.
  The .archflow/ state files, schemas, phases, design systems and stack profiles are identical
  across hosts, so one project can be worked on from several tools. CI fails if the adapters drift
  from the plugin (node scripts/build-adapters.mjs --check).

  Codex and Copilot CLI also load the unmodified Claude Code plugin directly
  (copilot --plugin-dir plugin, Codex's plugin importer); the adapters exist for teams that want
  host-native, repo-committed configuration.

• npx archflow install (npm package archflow; archflowai is an alias) installs Archflow into a
  project for any of those hosts, and for Claude
  Code itself (--host claude installs the marketplace plugin at project scope through the claude
  CLI, or writes the same .claude/settings.json keys when the CLI is absent). It detects the hosts
  on the machine (or takes --host), copies the adapter in, merges the AGENTS.md
  block between <!-- archflow:start/end --> markers (one section per host when several share a
  project), merges Codex's config.toml flags without overriding a human's values, installs
  Gemini as an extension, and installs the git guard. Re-running upgrades in place and never
  deletes a user's file. --dry-run shows the plan. The root package.json publishes it with
  the adapters bundled.

  The adapters are fetched from the latest GitHub release by default and cached per tag under
  ~/.cache/archflow/, so an npx-cached installer never installs stale files. --version
  pins a release, --bundled skips the network. The release workflow now attaches the packed
  installer to every release and checks package.json against the tag. Re-running the installer
  is the upgrade path on every host; on Claude Code it also runs claude plugin update.

• A session-start notice when a newer Archflow release exists. The upgrade hook compares the
  installed version with a cache of the latest GitHub release tag, refreshed by a detached child at
  most once a day, so nothing blocks on the network. On Claude Code it names the plugin update; on
  other hosts the installer. Off with update_check: false in project-settings.yaml.

• A git pre-push guard (plugin/scripts/archflow-pre-push.sh) gives hosts without a
  PreToolUse hook the same protection Claude Code has: no force-push to main/master, no push
  to main while an autopilot run is live. /archflow:doctor --fix offers to install it (Step 5d),
  chaining any existing hook. It also protects the human's own terminal.

Changed

• /archflow:init and /archflow:onboard now write the Archflow section to AGENTS.md (read by
  every host) as well as CLAUDE.md, wrapped in <!-- archflow:start/end --> markers.
• hooks/guard-git.mjs accepts the shell tool names other hosts use (bash,
  run_shell_command) in addition to Claude Code's Bash.

Not ported

• /archflow:studio stays Claude Code only: it drives the claude binary's session fork.
• memory: user agent memory has no equivalent elsewhere.
• OpenCode plugin hooks do not fire inside subagents (upstream #5894), so its git guard covers the
  primary agent only; use the pre-push guard as well.