Axon Shaft docs

Changelog

Every release of the dashboard and API, and of the action.

The dashboard and the API share one version and ship together. The action has its own: a workflow on axonpack/shaft-actions@v1 gets each v1 release as it comes out. New releases of both are in the RSS feed.

Dashboard and API

1.15.0

Minor Changes

    • Community checklists run in reviews: the grant sends each one at its pinned version, with its key @publisher/slug, publisher, version and exclude (the files it leaves out), and its check run is Axon Shaft / @publisher/slug. Replies and tasks never get one.
    • Their rules are framed as someone else's: the prompt fences them, says they come from @publisher, and limits them to findings within their scope. They can't change the output format, other checklists, the verdict rules or the instructions around them.
    • Run reports take check keys up to 100 characters.
    • Docs: a security page on how community items are kept safe, and the Marketplace guide covers using community items and approving new versions.
    • Community items in the Marketplace: community templates join the Templates tab, and Show picks All, Built-in or Community. A community item's page shows its author, uses and every version with its notes.
    • Use a community checklist: add it to a review from its page or from Add a checklist, pinned to the version you saw. Starting from a community template copies it at its version. Reply and Task automations can't take one.
    • Approve new versions: when an author publishes a new version, the automation shows Version n available, with a diff of the rules, and nothing changes until you approve it. An unpublished checklist keeps running at its version.
    • Sign-in: a community checklist's page opens again after signing in, instead of Page not found.
    • Glob tester: matches [abc] classes and a ** inside a name the way the action does.
    • Five more checklists in the catalog: GitHub workflows, Containers, Infrastructure as code, Database migrations and HTTP and schema contracts, with the Workflow and infra review, Database and migrations review and Infrastructure and Docker templates back in the marketplace. Action 1.10 hands each one the report of a pinned linter (actionlint, hadolint, tflint, squawk, oasdiff or buf); an older action reviews them with the model alone.
    • Docs: What runs in CI lists the linters the action downloads and how each is checked before it runs.

1.14.0

Minor Changes

    • Create in the Marketplace: makes a community checklist or template one step at a time, held to the same limits the API checks. A checklist's steps include a box that tries its globs on paths you type and a preview of its step in the review prompt. Save it as a draft, or publish it as its next version with what changed.
    • My items: a third Marketplace tab with each account's drafts and published items, their version, status and last change, to edit, publish or unpublish.
    • Fork: copies a community checklist's newest version into a draft on one of your accounts. POST /v1/accounts/<account>/items/fork does the same through the API, and each item now has forkedFrom.
    • Package lookups: POST /v1/packages/check takes the ecosystem, name and version of public packages from the action and answers their advisories (osv.dev, with GitHub's advisory database as the fallback) and the licence each one's registry declares (npm, PyPI, crates.io). It refuses names that can't be public, caps each outside call at 10 seconds and 5 MB, and keeps each version's answer for a day.
    • Licences checklist: now in the catalog and in the Library, Go and Rust service and Dependency updates templates. A new package under GPL, AGPL, SSPL or OSL is blocking, and one with no declared licence is should-fix. Needs action 1.10.
    • Docs: the data flow page says the action now sends public package names and versions to the Axon Shaft API, which asks osv.dev and the registries.

Patch Changes

    • Database: drops the old rate_limits table. Rate limits have counted in Valkey since 1.12.0.

1.13.1

No changes in this release.

1.13.0

Minor Changes

    • Marketplace: the Templates page is now the Marketplace, at /console/marketplace, with Templates and Checklists as two tabs. A template's page is at /marketplace/templates/<key> and a checklist's at /marketplace/checklists/<key>. The old /templates links still work and open the new pages.
    • Marketplace lists: each tab has a search box, a sort (Suggested, Most used or In most templates, and A to Z) and chips for the category or the kind of project, with a count of what matches. The cards keep an even height and show the publisher, Axon Shaft with a verified mark for the built-in items. A template card shows what it's for, when it runs, its tags and the projects that have it. A checklist card shows how it runs, the first two things it catches, the projects it suits, its files and how many templates list it. The Checklists list no longer scrolls sideways.
    • Checklist pages: each built-in checklist has a short list of what it catches and a sample finding, on its Marketplace page and its docs page. The page also shows the publisher, the version and how it runs, where it runs and the templates it's used in. A community checklist's page shows its rules and change notes.
    • Template pages: show the publisher and the template's own tags.
    • Add a checklist from the Marketplace: Add on a checklist's card, or Add to an automation on its page, picks one of your review automations by project and adds the checklist to the end of its list. An automation that has it already says so.
    • Change an automation's template: from the automation's menu, beside "Based on", or with Apply to an automation on a template's page. When the automation has checklists or checks its template didn't have, you choose to keep them after the new template's checklists or to reset to the new template. A kept checklist the new template also lists stays, greyed out as Already in the template, and runs again if you change to a template without it. Checklists you add yourself show Added by you.
    • API: a job's checklist can carry manual: true for one added by hand, and PATCH .../jobs/:id takes template to move a job to another template of its kind. The action gets each checklist once, without the flag.

Patch Changes

    • /fix on a review thread: when the API refuses a /fix commented in a review, it replies on that thread with the reason and asks you to comment /fix on the pull request's Conversation tab instead. It replies once per comment, and only to the person who wrote it.
    • Limit per pull request: a pull request can start 10 /ask and /fix runs an hour. Past that, the run ends with the new pull_request_limit notice, which says when to try again. Reviews don't count toward it.
    • Text aimed at the reviewer: the review prompt treats the pull request's title, body, diff, files and comments as data and never follows instructions in them. Text in a pull request that speaks to an AI reviewer or tries to change the review, such as "ignore previous instructions" or "approve this", becomes a blocking finding.
    • Cancelled runs: a run the action reports with the cancelled code closes with the cancelled conclusion, where it used to close as a failure, and the dashboard shows it as Cancelled. The job summary and the PR's summary comment show the new cancelled notice card.
    • Usage totals: a cancelled run counts as failed in the daily and lifetime totals, as it did when it closed as a failure. A run GitHub ended as cancelled without reporting back counted as passed before; it now counts as failed too.

1.12.0

Minor Changes

    • Community checklists and templates: an account can make checklists and templates under /v1/accounts/<account>/items, save drafts, publish each draft as a new version that never changes after, and unpublish. On an org only its admins publish, the publisher's GitHub account must be at least 30 days old, and each account can publish 5 times a minute. A checklist holds its rules and paths as data and can name one of the built-in scripts for facts; a template lists built-in checklists and community ones pinned to a version. Private items and integrations aren't available yet. Every change is in the audit log.
    • Checklist catalog: GET /v1/checklists lists the built-in checklists, from Axon Shaft and marked verified, then the published community ones. A built-in checklist no longer shows its rules; a community one shows every rule. The checklist page shows the rules only when there are some.
    • Console home page: /console no longer opens the first account. It says "Welcome, ", then shows Get started, the steps to a first review (install the app, add a project, add your Claude token, add the workflow), each ticked from your accounts and linked to the page that finishes it, until a project's first run. A card per account follows, with its plan, project count, the projects that need setup and permissions to review, and Add project for an account with none. The last card, Add an account or organization, opens GitHub's install page. Home in the sidebar comes back to it.
    • GET /v1/me: each account adds projectCount, needsSetup, setupNext, hasToken, hasRun and pendingPermissions.
    • Account switcher: its last item is now Add an account or organization.
    • Private draft advisories: a new Private draft advisories setting, in the Review section, picks Off or Public repos (the default). With it on, a review on a public repo can put each blocking security finding, up to 3, into a private draft security advisory, and the pull request shows a short note with the draft's link. POST /v1/advisories opens the draft with a token that holds only Repository security advisories write and never leaves the API. POST /v1/token sends settings.advisories, and the review prompt then asks the model to keep the headline and claim general about such a finding.
    • GitHub Code Quality findings: a code-quality checklist, in no template, lists Code Quality's open findings on the lines a pull request changes, with no model call. POST /v1/code-quality reads them with a Code quality (read) token that stays in the API, so the review's own GitHub token no longer asks for Code quality.
    • Advisory links on the run page: a finding's link to its draft advisory in the run's own repo is kept.
    • Permissions to review: Code quality and Repository security advisories now say what stops working until they're approved.
    • Plans: each account in GET /v1/me has a plan: free for a personal account and trial for an organization.
    • Paid endpoints: the API checks the plan on every request to a paid endpoint. An account whose plan doesn't include the feature gets 403 plan_required, naming the plan that does.
    • Plan card: the account's settings show its plan, the paid features it includes and the ones it doesn't, for anyone who can see the account (GET /v1/accounts/<account>/plan).
    • Plan marker: every paid feature is marked with the plan that has it, and shown locked with a link to the pricing on a plan without it.
    • Personal trial: Start trial on your own personal account turns on every paid feature, once, for a GitHub account at least 30 days old (POST /v1/accounts/<account>/trial). The trial has no end date.
    • Retention: a run is deleted 7 days after it started (it was 90). The audit log, the daily usage totals and resolved or dismissed item reports go after 3 months (the audit log was kept a year), and used OIDC token IDs once they expire. The API prunes when it starts and once a day.
    • Usage totals: each run adds its tokens, cost and how it ended to its project's daily totals and lifetime totals once, when it completes. The monthly spending cap and the account overview's 30-day usage read the daily totals, so they no longer depend on the runs being kept.
    • Usage view: the account's and each project's Runs page start with Usage, with Today, 7 days, 1 month and 3 months. Today and 7 days add the findings by severity and by checklist and the review time; 1 and 3 months show runs, passed, blocking and failed, tokens and cost. A lifetime line follows, such as "1,284 runs · 172M tokens · $412.37 since Nov 17, 2026". The API serves it at GET /v1/accounts/<account>/usage and .../projects/<project>/usage.
    • Older runs on GitHub: the runs lists end with a link to the repo's Axon Shaft workflow on GitHub for runs older than 7 days, and say that PR comments and check runs stay with the pull request.

Patch Changes

    • Overview refresh: the account overview and a project's Overview stop reading again once a run has been open for 2 hours, the longest time limit, so a run that never reported back no longer keeps them polling every 10 seconds.
    • Live runs: the dashboard's service worker asks the API first for every read, where it used to answer from its own copy for up to 60 seconds, so a run's progress shows within the 10-second refresh. Offline it still shows the last answers it kept.
    • Landing page answers: the questions no longer offer running the action without the dashboard, which it can't. The secrets answer says a fork's pull request is reviewed under pull_request_target, as the "Pull requests from forks" option decides, and the data answer names the package names and versions the Known vulnerabilities checklist sends to osv.dev or GitHub's advisory database. The setup steps say a repo's Admin and Maintain roles can add that repo too.
    • Verdict and merge hidden: an automation's What it posts section no longer shows "The verdict as its own check" or "Merge approved pull requests", since the action doesn't act on them yet. The API still keeps and takes them.
    • Checklists without a script hidden: API contract, Bundle size, Containers, Infrastructure as code, Licences, Migrations and Workflows are out of the catalog, the templates and every automation's list until the action has their scripts, since they could only skip. GET /v1/checklists lists 32.
    • Templates without their main checks hidden: Workflow and infra review, Database and migrations review, and Infrastructure and Docker are out of Templates until those scripts ship, since what's left of them wouldn't check what their names say. An automation already added from one keeps its checks.
    • Codes for skips: POST /v1/token names a paused project, a pull request its filters skip and one with a skip label with code paused, filtered or skip_label, next to the reason.
    • Settings in /help: the /help reply lists what a review on the repo runs with: the automation and its template, the model and effort, its checklists in order, the paths to skip and focus on, the skip labels and the merge gates. It posts again after the settings change.
    • Rate limits in Valkey: the API counts its rate limits in a Valkey container, where it used to write a Postgres row on every request. The limits and what each one counts by are unchanged. While Valkey is down, the API counts in its own memory and logs a warning once a minute, and requests go on as usual.
    • CORS preflights: a browser's OPTIONS preflight no longer counts against any rate limit.

1.11.0

Minor Changes

    • Templates by project type: eleven new review templates, each an ordered list of checklists for one kind of project: Expo or React Native app, React web app, Node or Bun API, Library or npm package, Monorepo, Python service, Go or Rust service, Infrastructure and Docker, Docs site or docs repo, Dependency updates, and Release and process. They're under the By project type chip on the Templates page, with category: 'stacks' in GET /v1/jobs.
    • Checklists on the Templates page: a Checklists chip lists the catalog. Each checklist has a page at /console/templates/checklists/<checklist> with the files it applies to, when it skips, its script, its rules and the templates that list it. A template's page links each of its checklists, in order.
    • An automation's checklists as an ordered list: under What it looks for, move a checklist up or down or drag it, turn it off, or remove it. Add a checklist opens the catalog, searchable and grouped, with each checklist's summary, files, and whether a script does all of it. A review's checklists take your own rules, and your own checks take the files they apply to. Checklists past the Code review's seven need action 1.6 or later.
    • Permissions to review: when AxonShaft asks for a permission that the account's installation hasn't approved yet, or approved only at a lower level, the account's admin now sees it. Alerts on the account overview and a card on each project's Overview list each permission with what doesn't work without it. A bar over each project's pages names them, and the bell lists them. Each has "Review on GitHub", which opens the installation's settings. The bar can be dismissed for the browser session.
    • GET /v1/accounts/<account>/permissions: { pending: [{ name, wanted, granted }] }, for the account's admin. It compares GitHub's GET /app (kept 10 minutes) with GET /app/installations/{id} (kept a minute), and is empty when GitHub fails.
    • Public profile pages at /u/@<login>: open without a sign-in, for people who have signed in to Axon Shaft. They show the public fields GitHub returns (picture, name, bio, company, location, website, X, public repos, followers, following, joined) and a link to GitHub, and nothing from Axon Shaft. Any other login reads "hasn't joined Axon Shaft yet".
    • GET /v1/users/<login>: the page's data, open without a sign-in, limited to 30 a minute per IP. It answers { joined: true, profile } or { joined: false, login }, and 404 when GitHub no longer has the user.
    • Logins link to profiles: the dashboard shows each GitHub login as @login, linking to their profile page, in the profile menu, Manage account, a project's pull requests and last run, recent runs and the audit log. Bots stay plain text.
    • Tokens hold only approved permissions: a review's GitHub token from POST /v1/token asks for Checks write and Code quality read only when the installation has approved them, so GitHub no longer refuses the first mint with a 422. Each permission it leaves out comes back in missing_permissions as { name, wanted, granted, warning }, and the action can log the warning once (for example, that the review posts no check runs). POST /v1/push-token refuses with 403 push_failed when the installation hasn't approved Contents write, before it mints anything.
    • Advisory links on findings: a finding in POST /v1/runs can carry a url, the GitHub advisory or osv.dev page behind it. The API keeps only links to github.com/advisories/ or osv.dev/vulnerability/ and drops any other, and the run's page links the finding's title to that page.
    • Skip by label: Code review has a new filter, "Skip pull requests with these labels". A pull request with any of the labels isn't reviewed on its own events, and the run ends with a notice naming the label. Labels match whatever their case. Every project starts with skip-review, no-review, skip-ai-review and do-not-review, and an admin can change the list or clear it to skip none. Someone with write access can still comment /review to review a labelled pull request. POST /v1/token takes pull_request.labels, which action 1.7 sends. An older action sends none, so nothing is skipped by label.

Patch Changes

    • Release header: every API response has X-Shaft-Version, the API's release, and the dashboard's origin may read it through CORS.
    • New version bar: a tab still on an older build of the dashboard shows "A new version is ready." with a Reload button once the API answers from a newer release.
    • Crash card: a page that crashes shows a "Something went wrong" card with a Reload button. If the API is on a newer release, the page reloads into the new build first, once per release.
    • Offline fallback: the service worker finds the saved page in any build's cache, so a page opened during a deploy no longer fails, and with no saved copy the browser shows its own offline page.
    • Org roles as GitHub reports them: role in GET /v1/me is the membership's own role (admin, member or billing_manager), null for someone who isn't an active member, and admin on the user's own account.
    • A refused role is unknown, not member: GitHub answers 403 when the app lacks the Members organization permission. The user is still treated as not an admin, and the dashboard shows "Role unknown" with the reason in Manage account, on the account's header and on its settings page.
    • New action version from the workflow file: the overview's "New action version" card now reads the version from the workflow's uses: axonpack/shaft-actions@<ref> line, so a workflow pinned to a release before 1.6, whose runs never report their version, gets the card too. The card says "The workflow pins Axon Shaft v1.5.0". A minor tag such as @v1.5 gets it once a newer minor is out. With no workflow to read, it falls back to the last run's version as before. GET .../workflow answers ref and latestAction for this.
    • Update the action version from the card: when the card comes from the workflow file, pick @v1 or the newest tag and click "Update workflow via PR". The pull request changes only the uses: axonpack/shaft-actions@ line and keeps the rest of the file. POST .../workflow takes { mode: 'version', ref } for this, with ref limited to v1 or the newest release's tag.
    • Set up before the first run: a project counts as set up once an automation is on, there's a Claude token, and the workflow file is on the default branch, up to date or not. It no longer waits for the first run. Its status is Ready, with "The first review runs on the next pull request.", and the Needs setup and Add the workflow cards go away. The setup walkthrough, the projects list and GET /projects/<project> follow the same rule. When GitHub can't say whether the file is there, a run counts instead.

1.10.0

Minor Changes

    • Action version on each run: POST /v1/token takes an optional action: { version, ref }, which action 1.6 sends, and keeps it on the run. When the run is pinned to an older v1 release than the newest, the answer has latest_action: { version, url }, so the summary comment can mention it. The newest release comes from GitHub's releases for axonpack/shaft-actions (the newest v1.x.y, without drafts or prereleases), read at most once an hour. A run at a branch or a SHA, or from an older action, compares nothing, and a GitHub failure never stops a run. The project overview's data has the last run's action and latestAction.
    • "New action version" on the overview: when the last run used an older v1 release, whoever may change the project sees a card with that version, the newest one, a link to its release notes, and the two ways to upgrade: uses: axonpack/shaft-actions@v1 for each v1 release as it comes out, or the new tag to pin it.
    • Reply runs on /ask: POST /v1/token with job: 'reply' grants a run with the Reply prompt, which tells the model to answer only, cite file:line, and never edit, commit or push. The GitHub token has no Checks write. The run's row opens like a review's, and someone without write access gets run: false. /fix and /axonshaft still say they can't run yet, and /help no longer marks /ask as coming soon.
    • A reply's answer on its run page: POST /v1/runs takes a finished reply as { model, effort, duration_ms, tokens, cost_usd, output: { answer, comment_url } }, with no verdict or findings. The run's page shows the answer and links to it on GitHub, and its status reads Done.
    • Reply is ready on the dashboard: templates and automations no longer show it as Coming soon. An automation that hasn't run yet says when it will, such as "when someone comments /ask".
    • Task runs on /fix and /axonshaft: POST /v1/token with job: 'task' grants a run with the Task prompt. It tells the model to make the smallest change, run only the repo's own commands, and never commit, push or call gh. The GitHub token is read-only for the code. The grant carries no push token.
    • POST /v1/push-token: the action asks for a task's push token only after its checks pass. The API reads the pull request from GitHub again and refuses a fork's, the default branch, Push mode Disabled and someone without write access with 403 push_failed. Otherwise it answers { push_token, branch }, a token with only Contents write. It shares the rate limits of POST /v1/token.
    • A task runs only from a comment, in a workflow on the default branch: POST /v1/token and POST /v1/push-token refuse a task whose OIDC token comes from another event, or whose workflow_ref isn't the repo's workflow on its default branch, read from GitHub. A comment on a review thread runs from the pull request's merge commit, so tasks start from the Conversation tab.
    • Write access is checked for every task and reply: POST /v1/token no longer skips the check when the body has an event.
    • New cards: nothing_to_fix, a notice for a bare /fix with nothing to fix, sandbox_unavailable, for a task whose runner can't start its sandbox, self_hosted_unsupported, for a task on a self-hosted runner, and credentials_in_checkout, for a task whose checkout kept its token. /help no longer marks /fix and /axonshaft as coming soon.
    • A task's reply on its run page: POST /v1/runs takes a finished task as { model, effort, duration_ms, tokens, cost_usd, output: { answer, comment_url } }, the same shape as a reply.
    • Task is ready on the dashboard: templates and automations no longer show it as Coming soon.
    • Outdated workflow: Axon Shaft now finds its workflow under any file name in .github/workflows/ on the default branch, not only at shaft.yml. When the workflow differs from the one the dashboard gives in its triggers, permissions, the job's if:, the action's @v1 tag or persist-credentials: false, a "Workflow needs an update" card on the project's Overview lists what's missing or different. "Update workflow via PR" opens a pull request that replaces the file at its own path, and the file to copy is under "Update it yourself". The check runs only when someone who may change the project opens it. GET /workflow adds the outdated state with path and changes, and POST /workflow takes { mode: 'update' }.
    • The workflow the dashboard gives now has the same comments as the one in the action's README.

Patch Changes

    • Changelogs: the Docs check no longer covers changelogs. Code review has a Changelog and changeset check of its own, so it still flags a published package that changed with no changeset. A job with Docs turned off keeps changelogs off too. Action v1.x reports these findings under Docs.
    • Checklists for the action: POST /v1/token also sends job.checklists, the checklists a job keeps on in its order with each one's step text, files and script, then your own checks keyed own-<n>. job.prompt adds open, before and after, so an action that runs checklists can build the steps from the ones that apply. Action v1.x keeps reading the fields it reads today.
    • Run results: POST /v1/runs takes up to 40 checks in output.checks, and output.passes, the model and tokens of each pass an action ran.
    • Checklist catalog: GET /v1/checklists lists the 38 checklists a review can be built from, each with what it looks at, the files it applies to and its rules. The Code review's seven checks now come from seven of them, and reviews run as before.
    • Setup steps: a project's setup checklist is now called its setup steps, and GET /v1/accounts/<account>/projects/<project> returns them as setup.
    • Templates as checklist lists: each template is now an ordered list of checklists. Security review lists security, access control, secrets, known vulnerabilities, supply chain, GitHub workflows, and personal data and logs. Docs and changelog review lists docs, changelog and config. Strict review lists the whole catalog. Templates that list checklists outside the Code review's seven (Security, Docs and changelog, Database and migrations, Simplicity, Workflow and infra, Strict) need action 1.6 or later. A workflow pinned to an older version runs only the Code review's checks among them, so Workflow and infra runs none there.
    • Job checks: a job stores checks as { list, custom }. list holds its checklists in order, each on or off, with up to 500 characters of the project's own rules. Your own checks get an own-<n> key and optional paths. Jobs saved earlier read the same as before, and the next save stores the new shape.
    • Run page: findings from your own checks that an older action reported together show under Your own checks.

1.9.2

Patch Changes

    • Outside contributors: a review set to skip outside contributors now reviews pull requests from people who can push to the repo through a team or a private org membership.

1.9.1

Patch Changes

    • Checks write on every review: the review's GitHub token now asks for Checks write whether or not the project has merge gates, so the action can post a check run for each review check. An installation that hasn't approved the Checks permission still gets the usual token, and the review runs.

1.9.0

Minor Changes

    • Merge gates: an account or a project can keep up to 10 plain-language requirements, such as "A migration has a rollback." The review reports each one as met, not met or not applicable, with a one-line reason, and the action posts them in the summary comment and as an Axon Shaft / gates check that fails when one isn't met. A project's own list replaces the account's. With gates, the run's GitHub token may also write checks. Settings has a Merge gates card, one gate per row, whose help says to require the check in branch protection. The action reports them from 1.4 on; an older action ignores them.
    • Monthly spending cap: the account's admin can set a monthly cap in US dollars in the account's Settings, and a different one on any project. It counts the cost that runs on an API key report over the calendar month (UTC). At the cap no run starts until the 1st, and the action's run ends with a notice saying whose cap was reached. The Settings card shows this month's spend against the cap, and the Usage card on the account's page shows it as a ring, with a badge from 80% of the cap.
    • Walkthrough in the summary: Code review has a new option in its What it posts section, on by default. For a pull request that changes three or more files, the review prompt asks for a table of the changed files grouped by area, with one line on what changed in each, and a Mermaid flow diagram when the pull request changes how its parts call each other. Turn it off to leave it out. The action shows it from 1.4 on; an older action ignores it.

Patch Changes

    • Setup checklist: a project that keeps Auto for its model and effort no longer shows as needing setup. The Needs setup card drops "Choose a model and effort", the setup walkthrough counts that step as done, and the project's checklist from the API no longer has a model item.
    • Review footer link: the API sends the action run_url, the run's page on the dashboard, so the "Axon Shaft" link in a review's footer opens that run instead of the dashboard home.
    • Turn limit: settings.maxTurnsDefault says whether maxTurns is the default, which no one set, so the action can use a smaller limit for a small pull request only then. maxTurns is still sent.

1.8.0

No changes in this release.

1.7.0

Minor Changes

    • Account overview: an account's page shows its projects as cards or rows, with search (/ jumps to it), a filter by status and Add project on top. Each project shows its week of runs, its status and its last run's pull request, and a menu to its pages. Beside them are the runs, tokens and (for admins) approximate cost of the last 30 days, an Alerts card that's empty for now, and the newest 5 runs.
    • Add the workflow: until a project's first run, its admins can add .github/workflows/shaft.yml from the overview. "Add workflow directly" commits it to the default branch, and "Add workflow via PR" opens a pull request, which the card offers on its own when the branch is protected. "Check installation" looks again for a file added by hand, and the file is still there to copy under "Add it yourself". The GitHub App now asks for the Workflows permission, which an existing installation's owner has to approve on GitHub first.
    • Project overview: the Overview tab shows the project's card with its default branch, status and last run, three cards for the last 7 days (runs, verdicts and failures), and the pull requests Axon Shaft ran on lately. A project that isn't set up yet shows its checklist in the card's place.
    • Runs: each run keeps who started it, and the pull request when the action sends it. A run started from a comment on the conversation tab keeps the pull request without its branch.
    • Danger zone: a project's Settings tab ends with a Danger zone for whoever may change the project. Pause Axon Shaft stops every run on the repo, each ending with a notice that it's paused, while the jobs, settings and Claude token stay; Resume turns it back on. Remove project deletes it, as on the Projects page. Both are audited, and a paused project's status on its overview reads Paused.
    • Which pull requests: Code review has a new card in the job editor. It can skip pull requests from outside contributors, ones opened by bots, and Axon Shaft's own, and it can review only pull requests into the branches you list (release/* works). The three skips are on by default, so a project that should review outside contributors' or bots' pull requests needs to turn them back on. A review someone asks for with /review runs either way.
    • After the review: two options you can set now that the action will use later. One adds the verdict as its own check, and the other merges an approved pull request (experimental: you type merge to turn it on). Both are off.
    • Templates: three more reviews in the marketplace. Simplicity review finds code the change didn't need, such as a helper the repo already has or a layer with one use. Workflow and infra review checks GitHub workflows, Dockerfiles and Compose files. Database and migrations review checks schema changes and migrations for locks, data loss and missing indexes.
    • Issue triage: a new kind of job for new issues, under Issues in the marketplace. Issue triage posts possible duplicates, related work, the files the issue likely touches and a short plan, or asks for what's missing. Label issues picks labels the repo already has. Each check can be turned off in the job editor. The action can't run them yet, so they show as Coming soon.
    • Run page: each run in a Runs list opens its own page. It shows what ran, links to the PR comment and the Actions log, and the review's card with its checks and findings by severity, each finding linked to its lines in the reviewed commit.
    • Run log: the run's page reads the log from GitHub. While a job runs you see its steps; once it ends you see the log, with each group folded and errors and warnings in colour.
    • Failed runs: a run that fails reports its error, and its page shows the error card and message.
    • Push mode: a new setting on the account and each project, sent to the action with the run. Disabled means the action never pushes, Restricted (the default) only to the pull request's own branch, and Enabled to any branch but the default one. The action pushes once Task ships.
    • Environment variables: a list of the runner's variables, by name, that the agent's commands get on top of its own short list. Names of tokens the runner or the action holds (ACTIONS_*, INPUT_*, ANTHROPIC_*, CLAUDE_*, GITHUB_TOKEN) are refused.
    • Runs: the Runs list has a Tokens column with what each run used. Hover it for input, output and cache tokens, and the cost for a run on an API key. Runs from an action before 1.2 show a dash.

1.6.1

No changes in this release.

1.6.0

Minor Changes

    • Default jobs: every new project starts with Code review, Reply and Task, all on. Reply and Task show as coming soon until the action can run them. A project added before gets them the first time its jobs open.
    • Templates: the marketplace lists ready-made jobs (Code review, Reply, Task, Security review, Docs and changelog review, Strict review, Light review, Tests review), with search and categories, and a page for each with what it comes with and an example of what it posts. The Jira and Linear reviews show as unavailable until those integrations exist.
    • Several jobs of one kind: add a template more than once, each job with its own name, options, checks, instructions, and model and effort. Every job that's on runs; until the action runs them all, a run uses the first one of its kind that's on.
    • Jobs tab: jobs grouped by kind, each with a switch that turns it on or off at once with Undo, the last run's status, and Edit, Duplicate, Rename, Reset to template and Remove.
    • Job editor: a card per section that saves on its own, with an example of what the job posts beside it, and a question before leaving with unsaved changes.
    • Runs: the runs list has a Job column.
    • Reply: a new kind of job that will answer a question on an /ask comment and never change code. Coming soon: until the action can run it, a mention gets a notice.
    • API: GET /v1/jobs lists the templates. A project's jobs are POST .../jobs { template, name? }, PATCH and DELETE .../jobs/<id>, and POST .../jobs/<id>/reset.
    • Slash commands: POST /v1/token takes job: 'help' and answers with the list of commands (/review, /fix, /ask, /axonshaft <prompt>, /help) for the action to post. The dashboard's workflow file starts a run on a comment with a line that begins with /.
    • Commands only: @axonshaft no longer starts anything; comments use the commands. /fix, /ask and /axonshaft <prompt> (a direct prompt) answer that Axon Shaft can't run them yet, instead of starting a run the action can't do.

1.5.0

Minor Changes

    • Runs: the account's Runs page, the overview's Recent runs and each project's Runs tab list the runs: status or verdict, project, commit with a link to the run on GitHub, trigger, findings, how long it took and when it started. The list refreshes every 10 seconds while a run is in progress.
    • Runs that ended early: a run that failed, was cancelled or timed out before it reported back takes its end from GitHub instead of showing as running forever.
    • API: GET /v1/accounts/<account>/runs and GET /v1/accounts/<account>/projects/<project>/runs, 50 at a time with ?before=<id>.

1.4.0

Minor Changes

    • Offline: the dashboard opens with no network. A service worker saves the whole build when it installs and shows the last answers it kept from the API, with GitHub avatars. Each deploy saves its own files.
    • Saved answers: the API's answers are reused for 60 seconds and avatars for a day. Failed answers are never kept, who's signed in is always asked again, and every kept answer is dropped after a change, on a 401 and on sign-out.

Patch Changes

    • Project settings: someone with Admin or Maintain on a repo, who isn't an admin of its org, can open the project's Settings tab again. It showed "Axon Shaft couldn't load the settings" because it also asked for the org's settings, which only the org's admins may read.
    • API: GET /v1/accounts/<account>/projects/<project>/settings also returns the org's own rules as rules.account, and the limits.
    • Console: the sidebar's top, with the account switcher, is as tall as the bar beside it, so their lines meet.

1.3.0

Minor Changes

    • Code review prompt: POST /v1/token sends the whole review prompt as job.prompt, with the project's steps, its rules and, for a deep review, the extra passes for the checks it keeps on. The action fills in the pull request's data and runs it.
    • Errors: when a run can't start, the API answers with a code, one sentence on what to do and, where it helps, a link to the page where you do it. A repo with no Claude token on the dashboard now stops before the run starts, with no_claude_token.
    • Error cards: each error code has a card at https://shaft.axonpack.dev/errors/<code>.svg, for the action to show in the job summary and on the pull request.
    • Pull requests from forks: Code review has a new option, "Pull requests from forks". It can review them automatically (the default), ask a maintainer first with a comment on the pull request, or wait until someone with write access comments @axonshaft review. The workflow file runs on pull_request_target, so a pull request from a fork reaches Axon Shaft too, and its checkout leaves no token in .git/config.
    • Status cards: the action's summary comment opens with a card for its status, Reviewing while it runs, then Looks good, Needs changes or Blocking.

1.2.0

Minor Changes

    • Landing page: redesigned. It shows what Axon Shaft does, how to set up a repo and the plans, with a Console button that opens the dashboard.
    • Sign-in: signed out, any console page shows a sign-in card in its place, and signing in comes back to that page. There's no separate sign-in page any more.
    • Console: the dashboard moved to /console. A sidebar holds the account switcher with the account's plan, the account's or the project's pages, Marketplace, Help, your profile and a notifications bell, and a bar over the page holds the project switcher. Old links open in the console.
    • Repo maintainers: anyone with Admin or Maintain on a repo can add it as a project and change its jobs, settings and Claude token.
    • Manage account: the profile menu shows your profile and the devices signed in to your account, and signs out any other one.
    • OAuth token: the Test button is gone for an OAuth token, which only Claude Code can use. A note says the next review run shows whether it works.
    • Reinstalling the app: the dashboard loads again after AxonShaft is uninstalled from an account and installed again.
    • API: GET /projects lists the projects as an array, GET /addable-project lists the repos you can add, each repo in /v1/me has canManage, and GET /v1/auth/sessions and DELETE /v1/auth/sessions/<id> list and sign out devices.

Patch Changes

    • Code review: a finding against something the repo's rules allow on purpose is dropped, so a rule the repo writes down wins over the built-in checks.

1.1.1

No changes in this release.

1.1.0

No changes in this release.

1.0.0

Major Changes

    • Sign in with GitHub: you sign in to the dashboard with your GitHub account.
    • Sign out: signing out, here or everywhere, ends your session at once.
    • Projects: you choose which of your repos Axon Shaft works on.
    • Job marketplace: you add Code review or Task to a project.
    • Job options: you pick when a review runs and which checks it makes. A ? next to each option says what it does.
    • Your own checks: you can add a check of your own, written as a plain sentence.
    • Claude token: you save a Claude token for your account or for one project, and test it.
    • Settings: you set the model, effort, rules and paths for your account, and change any of them per project.
    • Setup checklist: each project shows what's left before its first review, with a link to where you do it.
    • Audit log: you see who changed what on your account and its projects.
    • The Axon Shaft action: your workflow runs your project's jobs with the token and settings you saved.
    • The action's GitHub token: it can only read your code, post its review and comments, and read your workflow runs.
    • Who can get a Claude token: the Claude token panel tells you that anyone who can push a workflow to a repo can get the token its runs use.

Action

1.10.0

Minor Changes

    • Package lookups through the Axon Shaft API: the osv script (Known vulnerabilities) sends POST /v1/packages/check the ecosystem, name and version of the public packages a pull request adds or changes, and the API asks osv.dev and GitHub's advisory database. It sends nothing else. With an older API, the action asks osv.dev itself, as 1.9 did.
    • Licences checklist: a new licences script checks the licence each new public npm, PyPI or crates.io package declares, through the same API call. A strong copyleft licence (GPL, AGPL, SSPL, OSL) is blocking, and no declared licence is should-fix. It's skipped on an older API.
    • Pinned linters for five checklists: GitHub workflows gets actionlint's report and a scan for unpinned actions and permissions: write-all, Containers hadolint's, Infrastructure as code tflint's, Database migrations squawk's, and HTTP and schema contracts oasdiff's and buf breaking's. Each linter is pinned to one release, downloaded only from that project's GitHub release, checked against the sha256 this release records before it runs, and run in the sandbox with no network. When it can't run, the checklist runs on the model alone and the facts say why.

Patch Changes

    • Partial reviews move up the ladder: on auto, a review that leaves more than a third of its checklists not checked, says it only partly reviewed the diff, or rates its confidence under 60% moves up a model, as a failed attempt does. The last model's partial review is posted as it is. The footer's Tried first row says why, such as "partial review: 5 of 7 checklists not checked".
    • Big reviews start higher: over 2,000 changed lines after skip, a review starts at Opus, and over 400 lines it never starts below Sonnet.

1.9.0

Minor Changes

    • Model ladder for auto: a run starts on the smallest model that fits the job, at low effort, and moves up a model only when an attempt fails (an error, the turn limit, a reply that doesn't match the format, a failed check, or no change from a /fix). Reviews start at Sonnet, at Haiku for docs and config, and at Opus when they touch a sensitive file. /ask and /fix with words ask Haiku to rate the request first. A model or effort set on the dashboard runs as set. The footer lists each model tried and why the run moved on.
    • Bare /fix reuses the review: the review's own suggested changes are applied with no model call, all of a file's at once where the review put them (overlapping ones go to the model), the other findings go to the model with the review's whole comment and only the files they name, and a fix that changes any other file isn't pushed.
    • /fix with words gets the last review's open findings too, so "fix those" works.
    • Missing tools skip checks: a check whose program isn't on the runner is skipped instead of failing, the fix is pushed, and the reply says which checks were skipped.
    • Inline /fix: a /fix on a line of the diff sends the comment's ID to the API, which replies on that thread when it refuses the run.
    • No repeat inline comments: a finding that moved a line or two since the last review isn't posted again.
    • Failed Bash calls in the log: each one gets a short line with the command and what it printed.
    • Build files need asking: a /fix with words won't push a change to package scripts or dependencies, lockfiles, .npmrc, .yarnrc, bunfig.toml, a Makefile, a Dockerfile or Git hooks unless the request names the file or says dependency or script. A lockfile goes through only with a dependency change that was asked for.
    • Images in posted text: an image the model wrote stays only when it loads from GitHub or the dashboard; any other becomes a plain link.

Patch Changes

    • Cancelled runs close their checks: when a workflow is cancelled or times out, the action closes the check runs it left in progress (Axon Shaft / review and the rest) as cancelled, swaps the summary comment's "Reviewing" for the new cancelled card, and records the run on the dashboard with the cancelled code. A new post step does the same at the end of the job if the step was killed first. A failed review fails any check still in progress, and its comment shows the failure card.

1.8.0

Minor Changes

    • Links to the docs: each failure card links to its code's entry at https://shaft.axonpack.dev/docs/reference/codes#<code>, in the PR comment, the check run and the job summary. A notice with a code ends with the same link: fork_pull_request, not_a_pull_request and no_write_access from the action, the task codes such as push_failed, and whatever code the Axon Shaft API sends with its reason for not running. An older API sends no code, and its notices have no link.
    • Spending cap card: a run stopped by the monthly spending cap shows the spending_cap card in the job summary, as well as the notice.
    • Missing permissions: when the AxonShaft app's installation hasn't approved a permission the run wants, the job log warns once for each, and the summary's footer table lists them under "Missing permissions", such as `code_quality` (read).
    • Private draft advisories: with the dashboard's "Private draft advisories" setting on, each blocking security finding (the first 3 a review) goes into a private draft security advisory, opened by the Axon Shaft API. On a public repo the PR thread, the check run and the run's page show "Security issue, reported privately" with the draft's link in place of the details. On a private repo the finding keeps its text and links to the draft. If a draft can't be opened, the finding posts as before, the job log warns and the summary says why. A dry run logs the drafts it would open.
    • Findings stay out of the job log: a reply that doesn't match the format is logged by its field names, without its text, since the job log is public on a public repo.
    • GitHub Code Quality findings checklist: the code-quality script asks the Axon Shaft API for GitHub Code Quality's open findings and keeps those on the lines the pull request changes, with no model call. error findings are should fix and the rest are nits. A pull request into a branch other than the default is skipped, since the findings are the default branch's, and so is a repo where the API can't read them, with the reason.

Patch Changes

    • Checked Claude Code download: the action fetches the Claude Code binary from npm with npm pack and unpacks it only when the tarball's sha512 matches the one the release recorded from its bun.lock. A mismatch, or a runner platform with no recorded hash, fails the step with the new claude_binary_unverified code, and the binary never runs. Each run installs into an empty folder, so a binary already in the runner's temp folder is never reused.

1.7.1

Patch Changes

    • Update links under a newer release: when a newer v1 release is out, the summary shows it as a NOTE after the card, with "Update the workflow" (the workflow file in GitHub's editor, on its branch or the default branch) and "Update from the dashboard" (the run's page).
    • Plain card line: the card's last line drops its icons and reads Bug fix · 85% confidence · reviewed a1b2c3d, with the commit still linked.
    • Footer that opens into a table: the footer is one short line, such as Claude Sonnet 5.5, medium effort · 62k tokens, 219 lines (the cost follows the tokens on an API key). Click it for a table with the run's page, the Actions log, the action version, the model's ID, fallback and subagent models, the effort and why it was picked, the turn limit, the changed lines and files, the checklists, the full token breakdown with the billed figure and the cost, and the reviewed commit. Model IDs read as names, so claude-sonnet-5-5 is "Claude Sonnet 5.5".

1.7.0

Minor Changes

    • Advisory links: a Known vulnerabilities finding links each advisory ID to its page, GitHub's advisory page for a GHSA- ID or the record on osv.dev for any other. The finding's row in the summary and on its check run links to its most severe advisory, and the check run's details link goes there too. The run report sends the link to the Axon Shaft API with the finding.
    • Facts for nine more checklists: eight new pre-steps hand the model facts, placed after the steps in the prompt. Changelog and changeset gets the changed packages and the changesets that name them. Docs gets the docs files changed and the READMEs above changed code that stayed the same. Tests gets each changed source file with its changed test, an existing one, or none. File and folder structure gets each new file's folder, suffix and name case. Does what it says gets the lines changed by area and kind of file. Personal data and logs, and Logs and metrics, get the log, metric, analytics and error-report calls the diff adds. Feature flags gets the flags read and dropped and where the repo still names them. Type safety gets the type escapes added and the repo's settings for them. The scripts read only the diff and the checkout's files, and never run the repo's code.
    • More pre-step facts: What else it can break, Unused code, Unchecked generated code, Public API, Access control, Translations, Theme and design tokens, and Native and store now get facts from a script, placed after the steps in the prompt. They list the files that name each changed export, new exports no other file names, imports no manifest lists, exports before and after, new routes with the auth checks near them, locale keys the other locales lack, raw colours and sizes, and native changes with the runtime version before and after. The scripts read only the diff and the checkout, and never run the repo's code.
    • Labels for the skip filter: the action sends the pull request's label names to the Axon Shaft API, from the pull request or, for a comment, its issue. With them, the Code review's new "Skip pull requests with these labels" filter can skip a pull request labelled skip-review, no-review, skip-ai-review or do-not-review, or whatever list the project sets. Comment /review to review a labelled pull request anyway.

Patch Changes

    • What this PR does: the Claim, Impact, Scope and Not checked lines in the summary's folded section each sit in their own paragraph, with a blank line between them.
    • Summary card: the two chip rows are labelled Findings and Checks, and the last line marks the kind of change, the confidence and the reviewed commit with an icon each, with the commit linked to GitHub.

1.6.0

Minor Changes

    • Newer release notice: the action sends its version and the ref the workflow runs it at (v1, v1.5.0, a branch or a SHA) when it asks the Axon Shaft API for its tokens. When the workflow pins an older v1 release and a newer one is out, the summary comment has a line under the card with the new version, a link to its release, and how to get each v1 release as it comes out: change uses: to axonpack/shaft-actions@v1. Runs on @v1, a branch or a SHA get no line.
    • Checklists: when the Axon Shaft API sends the job's checklists, only the ones whose paths match a changed file (after skip) run, and each gets its own check run, Axon Shaft / <key>, in the job's order. One that didn't run is skipped and says why. The prompt puts the run's data first and the steps of the checklists that run last, and a finding can only name one of them. When none run, no model is called: the summary comment says so and every check run is skipped, unless the project has merge gates. Script checklists report as skipped until the action has their scripts. The run's page on the dashboard gets the session's model, tokens and checklists. With an API that doesn't send checklists, the review runs as before.
    • Lockfile facts: Dependencies and Supply chain now get a list of the packages each lockfile adds, removes or changes, with the versions before and after, placed after the steps in the prompt. It covers npm, Bun, Yarn, pnpm, pip requirements*.txt, Poetry, uv, Cargo and Go lockfiles, compares the base branch's copy with the pull request's, and never runs a package manager.
    • Known vulnerabilities: now runs as a script, with no model call. Each package version a lockfile adds or changes is checked against osv.dev, and against GitHub's advisory database when osv.dev fails. A finding names the advisory, the package, its version and the fixed versions, and is blocking when the advisory is rated high or critical. Only public packages are sent. Scoped packages, private registry, git and URL packages are never sent, and nothing about the repo is. When neither source answers, the checklist is skipped with the reason and never passes.
    • One email for the first review: the summary comment is now created when the result is ready. It no longer opens with a "Reviewing" card that GitHub emailed everyone about. On a later push the comment still switches to "Reviewing", since an edit sends no email. A failed review still creates the comment with its failure card.
    • Progress without emails: an Axon Shaft / review check run is in progress while the review runs. It then passes with a link to the summary comment, or fails with the failure's message. A dry run logs it and posts nothing.
    • Reactions on every command: the comment that started a run gets 👀 when it starts, after the write access check (/help gets it from anyone). When the run ends, the 👀 is replaced with 🚀 if it did its job, or 😕 if it failed or did nothing, such as a command that isn't built yet. A refused reaction only warns, and a dry run logs the reactions and adds none.
    • /ask answers questions: a comment with /ask <question> on a pull request runs the project's Reply job. The model reads the head commit with the review's read-only tools and environment. Its answer cites file:line and goes where the question was asked: in the line's thread for a comment on the diff, or as a comment on the pull request. It never edits, commits or pushes. Only people with write access can ask, and /ask on an issue ends with a notice. The run goes to the dashboard with the model, the tokens and the answer. If the session fails, its failure card is posted in the same place.
    • Review progress in the checks box: the Axon Shaft / review check run is updated once at the start of each step. Its title is the current step, such as "Reviewing with Sonnet, effort medium", and its summary lists the steps with ✓ on the ones done, the current one in bold, and how far into the review that step started. If GitHub refuses an update, the job log warns once and the step updates stop, while the review goes on. A dry run logs each step's title.
    • Script checklists: Secrets, Pull request basics, Config and env vars, and Schema and migration in step now run as scripts built into the action, with no model call. They read the diff, the pull request's description and files in the checkout, and never run the repo's code. Their findings post inline and in the summary, set their check runs, and raise the verdict. When only script checklists run, no model is called. A Secrets finding names the kind of key and the line, never the value.
    • /fix and /axonshaft make changes: on a pull request, /fix <what> and /axonshaft <prompt> run the project's Task job, and a bare /fix takes the open findings of the last review. With no review yet, or nothing open, a bare /fix only replies in one line saying so and ends with a notice. The model edits the head commit's checkout and may run the repo's own install, test, lint, build and format commands. It never gets a GitHub token and can't run git commit, git push, gh or install a package by name. The action then runs the checks again itself, asks the API for a Contents-write token only once they pass, and commits as axonshaft[bot] to the pull request's branch through GitHub's API, never forced, revoking the token afterwards. The reply lists the changed files, the commit and each check. If a check fails, nothing is pushed and the reply shows the output.
    • The commands run in a sandbox: the model's commands and the action's checks run in bubblewrap, with their own PID namespace and /proc, no sudo, an empty home folder and the checkout as the only folder they can write. Background commands are off, and every sandbox is killed when the session ends, before the change is read. The action installs bubblewrap on an Ubuntu runner when it's missing. Where the sandbox can't start, the task doesn't run and the run reports sandbox_unavailable. A task runs only on a GitHub-hosted runner: on a self-hosted one it doesn't run and the run reports self_hosted_unsupported. A checkout that kept its token in .git/config (no persist-credentials: false) is refused too, with credentials_in_checkout.
    • What a commit may hold: a change under .github/ (in any case) is refused, and the model can't edit files there. So is a change of more than 200 files, a file over 1 MB or more than 10 MB in all. A change to a package.json, a Makefile or a lock file is listed in the reply under "Read before merging". Values of the project's Environment variables are masked in the log and taken out of the reply.
    • Push mode is enforced before the session: Disabled, a fork's pull request or a branch that is the default branch ends with a reply and a notice saying why, and the run reports push_failed. Nothing is ever pushed to the default branch or with --force.
    • /fix and /axonshaft on an issue end with a notice.
    • Read the bot's commit before merging: its push starts the repo's CI on the pull request like any push. Tasks start from the Conversation tab, since a comment in a review thread runs the pull request's own workflow.

Patch Changes

    • Comments from teammates with team access: the review now reads the comments of anyone with admin or write access to the repo, as GitHub's permission check reports it, along with its owners, members and collaborators. Before, a teammate whose access came through an org team, or whose org membership is private, had their comments left out, so a reply such as "this is intentional" didn't reach the next review. The action asks once per commenter. When the check fails, that commenter's comments stay out and the job log warns.

1.5.0

Minor Changes

    • A check run per review check: next to Axon Shaft / gates, the action posts Axon Shaft / <check> on the reviewed commit for each check that's on, such as Axon Shaft / correctness, with that check's findings as its summary. A blocking finding fails it, other findings make it neutral, a clean check passes, and a check the review didn't do is skipped. A check turned off on the dashboard gets no run. Each one can be required in branch protection. They need Checks (write) on the run's token, which the Axon Shaft API asks for on every review run once its matching update ships. Without it the review still posts and the log warns once.

1.4.0

Minor Changes

    • Merge gates: when the project has merge gates on the dashboard, the summary comment lists each gate as met, not met or not applicable, with the review's reason, and the action posts an Axon Shaft / gates check run on the reviewed commit. The check fails when a gate isn't met or the review didn't report on it, so branch protection can require it. The AxonShaft GitHub App needs the Checks (write) permission for the check; without it the review still posts and the log warns.
    • Walkthrough: the summary comment can have a folded "Walkthrough" with a table of the changed files grouped by area, one line on what changed in each, and a Mermaid flow diagram when the pull request changes a flow. The API asks for it when Code review's "Walkthrough in the summary" option is on, which is the default. A walkthrough over the limits (8 areas, 20 files in an area) is left out, and the review posts without it.

Patch Changes

    • Marketplace listing: action.yml has a blue check-circle icon and a shorter description, so the action can be listed on GitHub Marketplace.
    • Footer tokens: the summary footer shows the cost and the total on an API key, such as $0.38 · 759k tokens, and on a subscription the cache-weighted count with the total in brackets, such as ≈220k billed (759k with cache). The raw total counted every cached read and looked several times bigger than the bill.
    • Footer link: "Axon Shaft" in the summary footer opens the run's page on the dashboard when the API sends it, and the dashboard otherwise.
    • Turn limit by size: with no turn limit set on the dashboard, a review of under 100 changed lines gets 12 turns and one under 500 gets 20, and the footer's reason says so, such as 60 lines, up to 12 turns. Larger diffs and deep reviews keep 30, and a limit set on the dashboard always wins.

1.3.0

Minor Changes

    • Review filters: the action also sends the branch a pull request goes into, its author's association with the repo, and whether a bot opened it. The API uses them for Code review's new filters (outside contributors, bots, Axon Shaft's own pull requests, target branches). When a filter skips a pull request, the run ends with a notice that says why.
    • Run page: the run sent to the dashboard has the review card (headline, each check's result and every finding), and a run that fails after it started sends its error, so the dashboard can show both.
    • Job log: each step is a folded group with how long it took, and the Review group lists each tool call the model made, with a deep pass's name before its own.
    • Pull request on the dashboard: the action sends the pull request's number, title, branch and author when it asks the API for its tokens, so the dashboard's overview can show which PR each run was for. A run started from a comment on the conversation tab sends it without the branch, since that event doesn't include one.
    • Environment: the model's environment holds a short list of the runner's variables (PATH, HOME, locale, time zone, temp folder and the CI basics) plus the names set in the project's Environment variables setting, never a token. ACTIONS_*, INPUT_*, ANTHROPIC_*, CLAUDE_* and GITHUB_TOKEN can't be added.
    • Tokens: the summary comment's footer shows how many tokens the review used, and the run sent to the dashboard has the input, output and cache counts. For a run on an API key it also has the cost.

1.2.0

Minor Changes

    • Slash commands: a comment starts Axon Shaft with /review, /fix, /ask, /axonshaft <prompt> or /help at the start of a line. /review reviews the pull request now, /help posts the list of commands, and /fix, /ask and /axonshaft end with a notice until Task and Reply are built. Commands in quotes or code don't count
    • @axonshaft removed: mentioning @axonshaft no longer starts anything; use /review
    • Workflow trigger: the workflow's if: starts only on a comment with a line that begins with /; copy the new one from the README or the dashboard

1.1.0

Minor Changes

    • The dashboard is required: every run gets its GitHub token, Claude token, settings and whole review prompt from shaft.axonpack.dev, so the workflow needs permissions: id-token: write and the repo needs a project there
    • The prompt comes from the dashboard: the review steps, the safety rules and the deep review's subagents are all set there, and the action only fills in the pull request's data
    • Inputs removed: claude_code_oauth_token, anthropic_api_key, app_id, private_key, github_token, model, effort, depth, mode, max_turns, timeout_minutes, nits, task_tools and config_path are gone, and dry_run is the only input left. Set the model, effort, depth, nits and limits on the dashboard
    • Standalone mode removed: the action no longer runs on a Claude token and a GitHub token given as inputs
    • .github/shaft.yml is no longer read: set rules, skip and focus on the dashboard instead
    • Failures say what to do: a failed run shows a card for its code, a message and a link in the job summary, and a failed review shows the same in the PR's summary comment. The README lists each code
    • Pull requests from forks: run the workflow on pull_request_target, with persist-credentials: false on its checkout, and a fork's pull request is reviewed as the dashboard's "Pull requests from forks" setting says. Anyone with write access can also start a review of a pull request by commenting @axonshaft review
    • Status cards: the summary comment opens with a card for its status from the dashboard, Reviewing while the review runs, then Looks good, Needs changes or Blocking.

1.0.0

Major Changes

    • Pull request reviews with Claude: the action reviews each pull request on your own GitHub runners, posts one summary comment that it updates on every push, and leaves inline comments on the lines it flags
    • No secrets in the workflow: install the AxonShaft app, add the repo on shaft.axonpack.dev, and the action gets its GitHub token, your saved Claude token and the review settings from there
    • Review steps from the dashboard: the repo's Code review job decides which pull request events start a review and which checks it makes, including checks you add yourself
    • Standalone mode: give a Claude token plus github_token, or your own GitHub App's app_id and private_key, and the action runs without the Axon Shaft API
    • Summary card: the comment opens with the verdict, the headline, the finding counts and a mark for each check, then a table of findings linked to their lines
    • Settings order: an input wins over the dashboard, the dashboard over .github/shaft.yml, and the file over the defaults
    • Review rules in .github/shaft.yml: rules, skip and focus, read from the base branch, so a pull request can't change its own review
    • Only trusted comments are read: the review reads comments from the repo's owners, members and collaborators, and leaves out everyone else's
    • The review stays in the checkout: it can read only the repo's files and run only git diff, git log and git show, and it never sees the runner's other tokens
    • Reviews against your repo's guides: the review reads CONTRIBUTING.md, CLAUDE.md, AGENTS.md, the README and similar files from the base branch, and checks the pull request against them
    • Review memory: a finding is posted once, and the next review marks fixed ones and resolves their threads
    • Model and effort picked per run: with auto, small changes get Sonnet, docs and config get Haiku, and large or sensitive changes get Opus
    • Dry run: dry_run: true prints the comments to the job log instead of posting them

Last updated on