Gentleman-Programming/gentle-ai v2.2.3 on GitHub

What's New

Gentle AI v2.2.3 makes SDD attempts compact, review recovery exact, and candidate delivery practical at repository scale without weakening release protection.

Reviewers can inspect immutable candidates through provider-bound tooling, recover the exact failed slot after an access failure, and let an unchanged candidate continue under ordinary repository policy when review is declined. Context7, repeated syncs, local-model discovery, and several installer and diagnostic edges are also safer and more predictable.

Changes

Compact orchestration and safer review recovery

Compact SDD attempts: normal apply, verify, and remediation orchestration now uses bounded acquire/settle results instead of carrying the complete attempt history, while preserving idempotency, CAS protection, budgets, immutable history, and explicit-only reset. (commit)
Exact recovery and provider-bound inspection: a pre-inspection access failure returns to the exact reviewer slot without spending another review budget, and reviewers inspect immutable base and candidate trees through gentle-ai review inspect-candidate rather than shell-specific raw Git recipes. (#1986, #2011)
Declined-candidate delivery: declining review creates no lineage or receipt, but the exact unchanged candidate can continue under ordinary repository policy. Candidate drift asks again, and release remains protected. (#2046)
Large-repository scaling: review startup no longer materializes the complete repository path inventory. References are checked against the immutable trees, so small candidates scale independently of repository size while directory and fabricated references still fail safely. (#2061)
Provider-defect handoff: blocked SDD agents now report provider defects with consent first and preserve the consumer repository instead of offering to switch projects and repair Gentle AI in place. (#2062)
Windows review portability: final path, shell, and timeout-sensitive review checks were made portable, and the complete seven-shard Windows suite is green for the shipped commit. (portability, stability)

Agent lifecycle and learning capture

Codex multi-agent v2: generated Codex guidance targets the v2 lifecycle and waits for successful terminal completion before dependent work continues. (#2030)
Delegated learnings survive handoff: SDD phase agents and generic delegations close with structured Key Learnings that Engram passive capture can retain across delegation boundaries. (#1707)

Context7, sync, and model discovery

Context7 registers in the right scope: Claude Code user registration now merges safely into ~/.claude.json, while workspace installation and sync respect --scope workspace. (#1909, #1617)
Repeated syncs converge: managed sections collapse to one copy without disturbing user content, and skill-registry fingerprints include content bytes so same-size edits cannot leave stale cached skills. (#1764, #2019)
Local and aliased models stay visible: LM Studio models are discovered from the local endpoint, and reasoning-effort variants fall back deterministically when providers expose the same model under different IDs. (#1917, #1695)

Safer tools and project policy

Installer and diagnostics fail more safely: GGA source installs resolve exact tag refs, the theme setting matches the installed identifier, doctor deduplicates symlinked PATH entries, and Pi reports which MCP probe phase timed out. (#1751, #1630, #2015, #2020)
Issue creation is repository-aware and privacy-aware: the skill discovers each repository's issue workflow and requires environment-specific names, paths, hosts, and secrets to be sanitized before publication. (#1779, #1981)
Benchmark and contributor guardrails are explicit: empty explicit journey selections fail instead of producing meaningless runs, benchmark validation attestation is part of contribution guidance, and the verified quick-fix handoff is published for contributors. (#2018, #1991, #2010)

Validation

Shipped commit: 919ea3daf2afd8287bb42d86f8d979a9741b4b9e
Required CI run 30626349016 passed all 9 jobs for the exact shipped commit.
Windows Full Suite run 30626349021 passed all 7 shards for the exact shipped commit.
Stable distribution remains four macOS/Linux archives plus a Minisign-signed checksum manifest. Official Windows binaries and Scoop remain withheld pending trusted Authenticode; Windows installation through Go 1.25.10+ is supported.

Contributors

Thanks to the community contributors whose merged work is part of this release:

@dnlrsls - nested-cwd review status, staged-recovery proof, Pi MCP timeout attribution, benchmark selection guard, doctor PATH deduplication, skill-registry fingerprints, LM Studio discovery, repository-aware issue creation, duplicate-section collapse, benchmark attestation docs, and the quick-fix readiness audit. (#2024, #2023, #2020, #2019, #2018, #2015, #2010, #1991, #1917, #1779, #1764)
@decode2 - provider-bound candidate inspection, branch-shadowing-safe GGA installs, effort-variant fallback, theme identifier alignment, and workspace-scoped Context7 injection. (#2011, #1751, #1695, #1630, #1617)
@Denver2828 - moved Claude Code Context7 registration to the ~/.claude.json user registry with a non-destructive merge. (#1909)
@ardelperal - added the mandatory pre-submission privacy review to the issue-creation skill. (#1981)
@pablontiv - closed delegated agent reports with Key Learnings for Engram passive capture. (#1707)
@MzaGuille - updated generated Codex guidance for the multi-agent v2 lifecycle. (#2030)

Upgrade

Homebrew:

brew upgrade gentle-ai
gentle-ai sync

Go 1.25.10+ on any supported platform, including Windows:

go install github.com/gentleman-programming/gentle-ai/v2/cmd/gentle-ai@v2.2.3
gentle-ai sync