gx changelog: new features, updates, and fixes

October 15, 2026

gx runs on your own key, with no account — There is no gx Cloud any more and nothing to sign in to. gx review and gx enhance talk to api.anthropic.com or api.openai.com using ANTHROPIC_API_KEY or OPENAI_API_KEY, and to nothing else. gx auth login, gx auth status, and gx auth logout are gone.

The GitHub App is gone too — Posting a review comment needs no app and no installation on your account. gx review finds a token from GH_TOKEN, then GITHUB_TOKEN, then the GitHub CLI's own login, and uses whichever it finds. If you are signed in with gh auth login, comments already work.

One install command — curl -fsSL https://download.gx.run/install.sh | sh places gx, gx-mcp, and completions, and verifies the archive against its published checksum. Releases are served from Cloudflare R2 inside the free tier, with the same archives on GitHub Releases. gx update is gone.

The commit trailer is no longer a URL — It was gx: https://gx.run/r/<id>, and the page it named is gone. It is now gx-revision: <id>. The trailer was never decoration: resolving a revision to a commit is a git log --grep for that line, so the marker had to stay. Both forms parse, so history keeps resolving.

The site lives in this repository — The Next.js app moved from a second repository into www/, and is now a static export with no backend, no database, and no server. The landing page is otherwise unchanged.

What was left out — Session context, the code index, and the shared review knowledge base need a vector index to search. With none configured they report themselves disabled, silently, and the review runs on the diff, the working tree, and the project's own linters and test suites. That is most of the value; the recall layer is the part you opt into.

August 16, 2026

gx review is the pre-ship verdict — Two commands that both ended in a ship / no-ship verdict was one too many, so gx gates (briefly gx constraints) is gone and gx review is the single gate. The deterministic checks the gate owned moved into review rather than disappearing with it: the secret scan over added lines, the dependency audits (govulncheck, npm audit, pip-audit) when a manifest or lockfile moved, and pytest on Python projects — all three now run on the command that hooks and CI actually call. The gxc/gxg shorthands and the gx_constraints MCP tool retire with it; gxr is the shorthand, gx_review the tool.

Blocking and advisory lanes — Findings now split by how much agreement is behind them. A finding blocks only when both reviewers raised it independently and the verification model confirmed it; one model's opinion is advisory and will not stop you. --fail-on blocking is the default, so the exit code gates without an opt-in flag, and --fail-on none turns it off.

Intent, not a prompt — The positional argument is now one sentence saying what the change was meant to do — gx review "fix the auth timeout" — rather than a question or a filter. The reviewer holds the diff against it, so a change that quietly does more than its stated intent is itself a finding. Omit it and gx derives one from the branch name and recent commit subjects, and says so in the report.

gx enhance — Prints the single highest-value fix in the current change as a prompt written for a coding model — the problem, the code, the change, and how to know it worked. It is gx review's own fix plan narrowed to one item, so the two never disagree. Pipes cleanly: gx enhance | claude -p "apply this".

Reviews survive the terminal — Every run that inspected code saves its full JSON report to review/review-findings-<timestamp>.json, git-excluded automatically. --md prints the PR-comment shape, and --fast trades the second reviewer and the verification pass for wall clock when someone is waiting.

July 29, 2026

PR summary verdict labels — The summary's triage line now states what the reader must do, not just a color: 🟢 Quick scan (LOW), 🟡 Careful pass (MEDIUM), 🔴 Deep review (HIGH). The label is enforced server-side to always match the Blast Radius level, so a red summary can never read as permission to skim.

A new name — The product is now gx. The CLI is gx (gxr for one-shot review, gx-mcp for agents). Install from download.gx.run; docs live at gx.run/docs.

July 26, 2026

Plain Git is the whole workflow — Capture and publishing are entirely hook-driven: git commit records a revision, git push captures the session and publishes. The porcelain that duplicated Git (gx push, gx sync, the background daemon, and the macOS menu-bar app) has been removed. If you have an older install, gx init cleans up the retired pieces.

PR summaries write to the description — gx Cloud writes the rich summary into the pull request description below your own text, behind a gx-owned marker. Your notes survive every regeneration.

gx review is guaranteed read-only — Review writes nothing: no repo state, no .git/index mutation, no ~/.gx on machines that never ran gx, and no uploads when a gate fails. --fail-on turns a review into a CI gate with distinct exit codes for "findings survived" and "nothing was reviewed". See the CLI reference.

Updated August 16, 2026: review now saves its JSON report to review/ (git-excluded, the one file it writes), gates on blocking findings by default, and exits 5 when the panel ran degraded.

Two-model review panel — gx review runs two independent Anthropic reviewers plus a judge that verifies and dedupes findings. When both reviewers flag the same issue independently, the report says so — the strongest confidence signal a panel produces.

July 2, 2026

Generated GitHub PR summaries — After you publish with plain git push and open a PR, gx Cloud posts a generated review summary highlighting blast radius and high-impact review targets.

Code review history memory — gx review pulls in indexed context from your past reviews and GitHub comment history, and attaches review policy context so findings stay consistent with how your team reviews code.

Prompted gx review — Trigger gx review with a targeted prompt to focus the review on a specific concern instead of the full default sweep.

Updated August 16, 2026: the positional argument is an intent, not a focusing prompt — use --scope and --focus to narrow a review.