v0.9.2

Added

• CI workflow (.github/workflows/ci.yml) — Every PR and push to main now runs a mechanical gate on GitHub Actions: bash syntax check on all tracked shell scripts, shellcheck at warning severity, the 7 guardian self-tests (language-check, policy-lint, check-command-rename, doctor-check, gh-account-guard, session-health, pre-tool-check) plus install.sh --self-test-migration on both macOS (BSD userland) and Ubuntu (GNU userland), the repo-content guardians (language-check, policy-lint, plan-md-contract-check), and the installer smoke (install.sh --dry-run + snapshot.sh list). Previously these guardians only ran when the agent workflow remembered to invoke them in-session; the BSD-sed (v0.7.1) and sort -V (FR-51) incident class is now caught by the dual-userland matrix before merge. Actions pinned to commit SHAs.

Fixed

• shellcheck findings across the script suite (14 warnings → 0) — isparto.sh uninstall block now uses "${ISPARTO_HOME:?}/..." so an unset/empty ISPARTO_HOME aborts instead of expanding to rm -rf /bin-style paths (SC2115); lib/snapshot.sh prune-sort rewritten from array=($(...)) to a glob-safe read-loop compatible with macOS bash 3.2 (SC2207) and bare > truncation given an explicit : no-op (SC2188); intentional glob-matching case patterns in pre-commit-security.sh and the by-design single-element matcher loop in install.sh annotated with scoped shellcheck disable directives rather than silenced globally.

Changed

• Repositioning: "Agent Team framework for Claude Code" → "cross-provider AI dev team" — All active positioning surfaces now lead with the cross-provider review structure (Claude plans and reviews what GPT implements; a zero-context GPT pass blind-reviews the plan against the product spec) instead of categorizing iSparto as a Claude-Code-specific framework. Claude Code + Codex CLI remain named as the runtime. Updated: README.md / README.zh-CN.md lead sections, comparison table (new "Who reviews the work" row), and a new dogfooding-evidence section linking both review directions to session-log/case-study entries; docs/product-spec.md Product Positioning + Competitive Differentiation (new paragraph contrasting single-vendor multi-agent fleets); docs/zh/quick-start.md opening; CLAUDE.md Project Overview; GitHub repo description and topics. Terminology standardized on "cross-provider" across active files; historical Tier 4 entries keep the old wording.