If you write software and you have not tried this yet, this is the part of the release to read twice.
The plugin turns a DevStride roadmap into an autonomous delivery loop. You plan a real architectural roadmap — solution, capability, epic, story, with the dependencies wired between them — and then the loop walks it for you: selecting the next unblocked item, branching, building, adversarially reviewing its own work, opening the pull request, merging, updating the item, and moving to the next one. Not a code-completion assistant. A planning and delivery loop that runs unattended.
A personal note: our own engineers are seeing well over 20x acceleration working this way. If you have not picked it up yet, you really should.
📚 Get the skills · The planning loop · The delivery loop
Twelve versions shipped since we last wrote about it. The ones worth knowing:
Every item the loop builds gets a second adversarial review beside Claude's own pass — and that reviewer is now explicitly yours to choose. Grok, Gemini, DeepSeek, and open-weight models served locally all qualify, as does anything behind an OpenAI-compatible endpoint. Setup probes for the engines it knows and always offers to name your own, then validates it against a published contract; doctor holds it to the same contract and never reports an engine as wrong merely for being one it does not recognise. Codex remains what setup offers when you do not choose — a default, not a requirement.
Full adversarial review now runs where it pays — at merge boundaries — instead of twice on every story, with routine work getting a bounded risk screen and anything touching authentication, migrations, irreversible state or a deployed contract still verified immediately. Two review cycles is the normal target, with an uncapped continuation for as long as a serious finding stands. This is a major version because one configuration key widened its reach from the local reviewer to every reviewer.
Every follow-up reviewer now receives a cumulative ledger of prior findings, verdicts, dispositions and fixes, so a settled finding is not rediscovered or argued back the other way three cycles later. A passing check can only be reused when the tree and the commands genuinely match what produced it.
/devstride:doctor diagnoses read-only exactly as before, then offers the repairs that are safe to automate as one batch, re-verifying each. What it will touch is a fixed classification, not a judgement made in the moment: files inside your repository's .claude/, or running the command that already owns a fix. It never touches workflow files, git state, DevStride records, or anything on a deployed stage — and run non-interactively it repairs nothing.
/devstride:update updates the exact installation that loaded it, preserving your install name and scope, verifies the installed files against the published tag rather than trusting an exit code, and stops rather than guessing when an install is pinned or administrator-managed. Better still, turn on auto-update once: /plugin → Marketplaces → devstride → Enable auto-update.
Your configuration can now say which stage a checkout deploys to and which stage names mean production. The release skill names the stage at the go-ahead, so you approve a specific consequence rather than a familiar sentence, and the hotfix skill stops and asks before touching a database on a production stage.
Setup can write your repository a status line — model, effort, repository, checkout, branch, stage and pull request, each omitted when it does not apply. Checkout is the one that earns its place: the loop puts work in worktrees, which is exactly when "which checkout am I in" stops being obvious. Because it lives in the repository rather than one person's settings, everyone gets it.
High-risk verification now groups findings by file instead of spawning one agent per finding; a second review round reads only the fixes rather than re-reading the whole diff; and the wait for a cloud reviewer is learned from that reviewer's own history instead of always burning the full timeout.
Branches the loop creates are {user}/{YY-MM-DD}/{item}-{slug}. The year leads so a plain alphabetical listing — git branch, tab-completion, the GitHub branch picker — comes out in date order. A month-first date scatters a year's work across the alphabet, which is exactly what this avoids.
Also fixed: status line values could render in a colour many terminal themes place at or near the background, so a value went invisible while its label stayed (3.1.1).
Every Knowledge Base article now keeps a full version history — who changed what, and when — with a History panel in the editor. You can preview any earlier version and restore it.
Restoring is deliberately narrow: it affects title and body only. Publish state, slug, visibility, audience and category are left exactly as they are, and nothing is deleted — the restore itself becomes a history entry, so the record of what happened stays complete. A writing session is grouped into a single entry rather than one per keystroke.
The MCP server gained first-class Knowledge Base tools, so an agent can search, read, write, publish and archive articles — plus a set for reading version history and restoring a version. Anything you can do to an article in the app, you can now do through the MCP server.
Authentication to the DevStride MCP server is now stored and reused, so connecting is a one-time step rather than something you repeat every session.
Typing a short, common word into search could return nothing at all, because English stop-word processing discarded the term before the prefix match ran. Searching me now finds meetings. The fix covers both halves of the problem: a query made entirely of common words, and a common word sitting among real ones.
When a deploy replaces the front-end bundle, a tab left open from before can request code files that no longer exist and fail in ways that look like random breakage. DevStride now recognises that state and prompts you to reload, replacing an automatic reload that only ever worked in Chromium-based browsers.
Following a comment notification used to open the item and leave you to find the comment yourself. The link now takes you to the specific comment it was about.
2026-08-27
DevStride for Claude Code 1.0.0 → 1.2.x: the plugin now keeps itself current, documentation and release notes run through skills your repository owns, and CI runs once. Update once by hand — then it updates itself.
2026-09-01
Forward a customer's email to your support address and the ticket opens in their name — their subject, their acknowledgement, their reply thread. Plus convert any item into a ticket, published lists as a real table, and scanned attachments on public request forms.